The Routledge Handbook of Mechanisms and Mechanical Philosophy 1nbsped 1138841692 978-1-138 84169 7 978 1 315 73154 4 Compress

THE ROUTLEDGE HANDBOOK OF
MECHANISMS AND MECHANICAL

PHILOSOPHY
Scientists studying the burning of stars, the evolution of species, DNA, the brain, the economy,
and social change, all frequently describe their work as searching for mechanisms. Despite this
fact, for much of the twentieth century philosophical discussions of the nature of mechanisms
remained outside philosophy of science.
The Routledge Handbook of Mechanisms and Mechanical Philosophy is an outstanding reference
source to the key topics, problems, and debates in this exciting subject and is the first collec-
tion of its kind. Comprising over thirty chapters by a team of international contributors, the
Handbook is divided into four Parts:
• Historical perspectives on mechanisms

• The nature of mechanisms
• Mechanisms and the philosophy of science
• Disciplinary perspectives on mechanisms.
Within these Parts central topics and problems are examined, including the rise of mechanical
philosophy in the seventeenth century; what mechanisms are made of and how they are organ-
ized; mechanisms and laws and regularities; how mechanisms are discovered and explained;
dynamical systems theory; and disciplinary perspectives from physics, chemistry, biology, bio-
medicine, ecology, neuroscience, and the social sciences.
Essential reading for students and researchers in philosophy of science, the Handbook will
also be of interest to those in related fields, such as metaphysics, philosophy of psychology, and
history of science.
Stuart Glennan is the Harry T. Ice Professor of Philosophy and Associate Dean of the College
of Liberal Arts and Sciences at Butler University, USA.
Phyllis Illari is Senior Lecturer in Philosophy of Science in the Science and Technology Studies
Department at University College London, UK.
ROUTLEDGE HANDBOOKS IN PHILOSOPHY
Routledge Handbooks in Philosophy are state-of-the-art surveys of emerging, newly refreshed, and
important fields in philosophy, providing accessible yet thorough assessments of key problems,
themes, thinkers, and recent developments in research.
All chapters for each volume are specially commissioned, and written by leading scholars in
the field. Carefully edited and organized, Routledge Handbooks in Philosophy provide indispensa-
ble reference tools for students and researchers seeking a comprehensive overview of new and
exciting topics in philosophy. They are also valuable teaching resources as accompaniments to
textbooks, anthologies, and research-orientated publications.
Recently published:
The Routledge Handbook of Epistemic Injustice
Edited by Ian James Kidd, José Medina and Gaile Pohlhaus
The Routledge Handbook of Philosophy of Pain

Edited by Jennifer Corns
The Routledge Handbook of Brentano and the Brentano School

Edited by Uriah Kriegel
The Routledge Handbook of Metaethics

Edited by Tristram McPherson and David Plunkett
The Routledge Handbook of Philosophy of Memory

Edited by Sven Bernecker and Kourken Michaelian
The Routledge Handbook of Evolution and Philosophy

Edited by Richard Joyce
The Routledge Handbook of Mechanisms and Mechanical Philosophy

Edited by Stuart Glennan and Phyllis Illari
THE ROUTLEDGE HANDBOOK
OF MECHANISMS AND
MECHANICAL PHILOSOPHY
Edited by
Stuart Glennan and Phyllis Illari
First published 2018
by Routledge
2 Park Square, Milton Park, Abingdon, Oxon OX14 4RN
and by Routledge
711 Third Avenue, New York, NY 10017
Routledge is an imprint of the Taylor & Francis Group, an informa business
 2018 selection and editorial matter, Stuart Glennan and Phyllis Illari;
individual chapters, the contributors.
The right of Stuart Glennan and Phyllis Illari to be identified as the authors
of the editorial material, and of the authors for their individual chapters,
has been asserted in accordance with sections 77 and 78 of the Copyright,
Designs and Patents Act 1988.
All rights reserved. No part of this book may be reprinted or reproduced or
utilised in any form or by any electronic, mechanical, or other means, now
known or hereafter invented, including photocopying and recording, or in
any information storage or retrieval system, without permission in writing
from the publishers.
Trademark notice: Product or corporate names may be trademarks or
registered trademarks, and are used only for identification and explanation
without intent to infringe.
British Library Cataloguing-in-Publication Data
A catalogue record for this book is available from the British Library
Library of Congress Cataloging-in-Publication Data
Names: Glennan, Stuart, editor. | Illari, Phyllis McKay, editor.
Title: The Routledge handbook of mechanisms and mechanical
philosophy / edited by Stuart Glennan and Phyllis Illari. Other titles:
Handbook of mechanisms and mechanical philosophy
Description: 1 [edition]. | New York : Routledge, 2017. |
Series: Routledge handbooks in philosophy | Includes bibliographical
references and index.
Identifiers: LCCN 2017001422| ISBN 9781138841697 (hardback : alk.
paper) | ISBN 9781315731544 (e-book)
Subjects: LCSH: Mechanical movements—History. | Mechanical
engineering—History.
Classification: LCC TJ15 .R68 2017 | DDC 621—dc23
LC record available at https://lccn.loc.gov/2017001422
ISBN: 978-1-138-84169-7 (hbk)

ISBN: 978-1-315-73154-4 (ebk)
Typeset in Bembo
by Swales & Willis Ltd, Exeter, Devon, UK
CONTENTS
Contributors ix
Foreword xiv
William C. Wimsatt
Acknowledgments xvii
1 Introduction: mechanisms and mechanical philosophies 1

PART I
Historical perspectives on mechanisms 11
2 Mechanisms: ancient sources 13

Tiberiu Popa
3 From the mechanical philosophy to early modern mechanisms 26

Sophie Roux
4 The origins of the reaction mechanism 46

William Goodwin
5 Mechanism, organicism, and vitalism 59

Garland E. Allen
6 Mechanisms and the mental 74

Marcin Miłkowski
v
Contents
PART II
The nature of mechanisms 89
7 Varieties of mechanisms 91
8 Mechanisms, phenomena, and functions 104

Justin Garson
9 The components and boundaries of mechanisms 116

Marie I. Kaiser
10 Mechanisms and the metaphysics of causation 131

Lucas J. Matthews and James Tabery
11 Mechanisms, counterfactuals, and laws 144

Stavros Ioannidis and Stathis Psillos
12 What would Hume say? Regularities, laws, and mechanisms 157

Holly Andersen
13 Probability and chance in mechanisms 169

Marshall Abrams
14 Mechanistic levels, reduction, and emergence 185

Mark Povich and Carl F. Craver
15 Mechanisms and natural kinds 198

Emma Tobin
PART III
Mechanisms and the philosophy of science 211
16 Mechanistic explanation and its limits 213

Marta Halina
17 Models of mechanisms 225

John Matthewson
18 Explaining visually using mechanism diagrams 238

Adele Abrahamsen, Benjamin Sheredos, and William Bechtel
vi
Contents
19 Strategies for discovering mechanisms 255

Lindley Darden
20 Mechanisms and dynamical systems 267

David Michael Kaplan
PART IV
Disciplinary perspectives on mechanisms 281
21 Mechanisms in physics 283

Meinard Kuhlmann
22 Mechanisms in evolutionary biology 296

Lane DesAutels
23 Mechanisms in molecular biology 308

Tudor M. Baetu
24 Mechanisms and biomedicine 319

Brendan Clarke and Federica Russo
25 Developmental mechanisms 332

Alan C. Love
26 Mechanisms in ecology 348

Viorel Pâslaru
27 Systems biology and mechanistic explanation 362

Ingo Brigandt, Sara Green, and Maureen A. O’Malley
28 Mechanistic explanation in neuroscience 375

Catherine Stinson and Jacqueline Sullivan
29 Mechanisms in cognitive science 389

Carlos Zednik
30 Social mechanisms 401

Petri Ylikoski
31 Disaggregating historical explanation: the move to social mechanisms 413

Daniel Little
vii
Contents
32 Mechanisms in economics 423

Caterina Marchionni
33 Computational mechanisms 435

Gualtiero Piccinini
34 Mechanisms and engineering science 447

Dingmar van Eck
Index 462
viii
CONTRIBUTORS
Adele Abrahamsen is Project Scientist in the Center for Research in Language at University
of California, San Diego. In addition to work in philosophy of science on mechanistic expla-
nation and interdisciplinary relations, she has done cognitive science research on mental
representations of meaning, syntactic processing, language–cognition relations, and the onset
of the symbolic function.
Marshall Abrams studies probability and modeling in biological and social sciences. After a
Ph.D. in philosophy from the University of Chicago, he worked at Colgate, Duke, George
Washington University, and the University of Alabama at Birmingham, where he is currently
an Associate Professor.
Garland E. Allen is Professor Emeritus of Biology at Washington University in St. Louis. He

received his Ph.D. in history of science from Harvard University. His research focuses on the
history of biology, particularly genetics and its relationships to evolutionary theory and embry-
ology in the twentieth century. He is author of, among other works, Life Science in the Twentieth
Century (Cambridge University Press, 1978), Thomas Hunt Morgan: The Man and His Science
(Princeton University Press, 1978), and Scientific Process and Social Issues in Biology Education
(Springer, 2016).
Holly Andersen is Associate Professor in the Philosophy Department at Simon Fraser

University, in Burnaby, British Columbia. She works in philosophy of science and in meta-
physics and epistemology broadly construed. Much of her work relates to causation: causal
explanation, application of causal methodology to case studies in philosophy of science,
problems related to mental causation, and the metaphysics of causation.
Tudor M. Baetu is Lecturer in Philosophy of Science at the University of Bristol. His research
interests are in philosophical issues concerning explanation, modeling, and experiment in
biology and medicine.
ix
Contributors
William Bechtel is Distinguished Professor of Philosophy and a member of the Center for
Circadian Biology and the Interdisciplinary Program in Cognitive Science at University of
California, San Diego. His research has focused on mechanistic explanation, modeling, and
discovery in cell and molecular biology; cognitive science; and most recently, systems biology.
Ingo Brigandt is Canada Research Chair in Philosophy of Biology at the University of Alberta,
Canada. His research on evolutionary developmental biology, molecular biology, and systems
biology concerns the topics of explanation, reduction and integration, concept change, and the
role of values.
Brendan Clarke is Lecturer in History and Philosophy of Medicine at the Department

of Science and Technology Studies, University College London. Originally qualified in
medicine, his research now concentrates on the intersection between epistemology and
clinical medicine.
Carl F. Craver is Professor of Philosophy in the Philosophy-Neuroscience-Psychology

Program at Washington University in St. Louis. He is the author of Explaining the Brain (Oxford
University Press, 2007) and coauthor (with Lindley Darden) of In Search of Mechanisms: Discoveries
Across the Life Sciences (University of Chicago Press, 2013).
Lindley Darden (Ph.D., conceptual foundations of science, University of Chicago) is Professor

of Philosophy, Affiliate in History and Biological Sciences, and Distinguished Scholar/Teacher
at the University of Maryland, College Park. She was President of the International Society for
History, Philosophy, and Social Studies of Biology (ISHPSSB) from 2001–3, and was elected
Fellow of the American Association for the Advancement of Science in 1995. Her most recent
book, with Carl F. Craver, is In Search of Mechanisms: Discoveries Across the Life Sciences (University
of Chicago Press, 2013). Her current work is on the discovery of disease mechanisms.
Lane DesAutels is Assistant Professor in the Department of Philosophy and Religion at

Missouri Western State University in St. Joseph, Missouri. His primary research interests are
history and philosophy of science (especially biology), metaphysics, and value theory.
Justin Garson is Associate Professor of Philosophy at Hunter College of the City University
of New York. His interests are in the history and philosophy of science and, in particular,
the intersection of biology and mind. He is the author of The Biological Mind: A Philosophical
Introduction (Routledge, 2015) and A Critical Overview of Biological Functions (Springer, 2016).
He is also a co-editor of The Routledge Handbook of Philosophy of Biodiversity (Routledge, 2016).
Stuart Glennan is the Harry T. Ice Professor of Philosophy and Associate Dean of the College
of Liberal Arts and Sciences at Butler University (Indianapolis, USA). His research focuses
on mechanisms, causality, modeling, and scientific explanation. He is the author of The New
Mechanical Philosophy, forthcoming from Oxford University Press.
William Goodwin is Associate Professor in the philosophy department at the University of

South Florida. He is a philosopher of science focused on applied and constructive sciences, such
as chemistry and climatology. Much of his work investigates how standard philosophical topics
such as the nature of explanation, mechanism, and modeling look different in the context of
these philosophically under-investigated scientific fields.
x
Contributors
Sara Green is a postdoctoral fellow at the Department of Science Education, University of

Copenhagen, Denmark. She works on the epistemic and social implications of research strate-
gies in systems biology and systems medicine.
Marta Halina is University Lecturer in the Department of History and Philosophy of Science
at the University of Cambridge. She works in the areas of philosophy of biology, philosophy of
cognitive science, and general philosophy of science.
Phyllis Illari is Senior Lecturer in Philosophy of Science in the Science and Technology Studies
Department at University College London, and, from July 2017, joint Editor-in-Chief (with Federica
Russo) of the European Journal for Philosophy of Science. Her current interests are mechanisms, causality,
and information, and how they impact on evidence assessment in biomedical sciences.
Stavros Ioannidis is a philosopher of biology and currently a postdoctoral researcher at the

University of Athens, Greece. His research focuses on causation and explanation in biology,
particularly on mechanistic explanation and developmental explanations of evolution, as well as
on more general issues in the metaphysics of science.
Marie I. Kaiser is Assistant Professor at the University of Bielefeld, Germany. Her main
research interests are the philosophy of biology, the general philosophy of science, and the
metaphysics of biological practice. In particular, her work focuses on the concept of reductive
explanation, mechanisms, part–whole relations, causal modeling, complex systems, biological
individuality, and the methodology of philosophy of science.
David Michael Kaplan is Senior Lecturer in the Department of Cognitive Science, an Associate
Investigator in the Australian Research Council (ARC) Centre of Excellence in Cognition and
its Disorders, and an Associate Investigator in the Perception and Action Research Centre at
Macquarie University, Australia.
Meinard Kuhlmann, Professor of Philosophy at Mainz University in Germany, received dual

degrees in physics and philosophy and has worked at the universities of Oxford, Chicago, and
Pittsburgh. His main fields of research are the philosophical aspects of quantum field theory and
complex systems.
Daniel Little is Chancellor and Professor of Philosophy at the University of Michigan-Dearborn

and Professor of Sociology at the University of Michigan, Ann Arbor. He is the author of many
books on philosophy, social science, economic development, and China. His most recent book
is New Directions in the Philosophy of Social Science (Rowman & Littlefield, 2016). He also main-
tains an academic blog at UnderstandingSociety.blogspot.com.
Alan C. Love is Associate Professor of Philosophy at the University of Minnesota and Director
of the Minnesota Center for Philosophy of Science. His research concentrates on concepts,
methods, and reasoning in developmental and evolutionary biology with a special focus on
interdisciplinary explanation.
Caterina Marchionni is Academy Research Fellow at the Academy of Finland Centre of

Excellence in the Philosophy of the Social Sciences (TINT), University of Helsinki. Her
research mainly concerns scientific modeling, explanation, and interdisciplinarity in economics
and the social sciences.
xi
Contributors
Lucas J. Matthews is a philosopher of biology who received his Ph.D. in 2016 from the
University of Utah. He is currently a postdoctoral research associate in a behavioral genetics lab
in the Department of Psychology at the University of Virginia.
John Matthewson received his Ph.D. at the Australian National University in 2012, and
teaches philosophy at Massey University in Auckland, New Zealand. His research focuses on
scientific modeling, mechanistic explanation, and the medical sciences.
Marcin Miłkowski is Associate Professor in the Institute of Philosophy and Sociology, Polish
Academy of Sciences (Warsaw, Poland). For his Explaining the Computational Mind (MIT Press,
2013), he received the Tadeusz Kotarbiński Prize from the Polish Academy of Sciences and
the National Science Center Award. He received the Herbert A. Simon Award for significant
contributions to the foundations of computational neuroscience.
Maureen A. O’Malley is a philosopher of biology at the University of Bordeaux, France. She

specializes in philosophy of microbiology, including microbial systems biology.
Viorel Pâslaru is Associate Professor of Philosophy at the University of Dayton. He has pub-
lished on topics in philosophy of science and of ecology, such as mechanistic and causal explana-
tion, causation, laws, and nomological explanation in ecology.
Gualtiero Piccinini (Ph.D., University of Pittsburgh, history and philosophy of science, 2003)
is Professor of Philosophy and Associate Director of the Center for Neurodynamics at the
University of Missouri-St. Louis. His book, Physical Computation: A Mechanistic Account, was
published in 2015 by Oxford University Press.
Tiberiu Popa is Associate Professor of Philosophy at Butler University (Indianapolis, USA).

His recent publications deal, among other things, with Aristotle’s science and philosophy of
science and with topics of philosophical interest in the Hippocratic Corpus.
Mark Povich is completing his Ph.D. in the Philosophy-Neuroscience-Psychology Program at

Washington University in St. Louis. He works on the philosophy of scientific explanation, and
has publications in Philosophy of Science and The British Journal for the Philosophy of Science.
Stathis Psillos is Professor of Philosophy of Science and Metaphysics at the University of Athens,
Greece and a member of the Rotman Institute of Philosophy at the University of Western
Ontario (where he held the Rotman Canada Research Chair in Philosophy of Science). He is the
author or editor of seven books and of more than 120 papers and reviews in learned journals and
edited books, mainly on scientific realism, causation, explanation, and the history of philosophy
of science. He is a member of Academia Europaea.
Sophie Roux is Professor at the École Normale Supérieure, Paris. Her main research is in the
history of early modern thought (Descartes, Galileo, mechanics, mechanical philosophy). She
has also published in philosophy of science (thought experiments, mathematization) and history
of philosophy of science (Pierre Duhem, Louis Couturat).
Federica Russo is Assistant Professor at the philosophy department, the University of

Amsterdam. Her research covers causality, evidence, and modeling in the social, biomedical,
and policy sciences, as well as the relation between science and technology.
xii
Contributors
Benjamin Sheredos received a joint Ph.D. in philosophy and cognitive science from University
of California, San Diego in 2016. In addition to philosophy of science, he works in the history
of philosophy and has developed a novel understanding of Brentano and Husserl’s accounts
of the origins of intentionality in mental acts. As a postdoc in UCSD’s Center for Circadian
Biology, he also applies his research in a team advancing web-based science education.
Catherine Stinson is Postdoctoral Fellow in Philosophy of Neuroscience at the Rotman

Institute of Philosophy. Her research covers mechanistic (and non-mechanistic) explanation,
computational models, animal models, the metaphysics of abstraction, classification in psychia-
try, and body perception.
Jacqueline Sullivan is Associate Professor of Philosophy at the University of Western Ontario.

She has published numerous articles on topics in philosophy of neuroscience and philosophy
of psychiatry, and co-edited Classifying Psychopathology: Mental Kinds and Natural Kinds (MIT
Press, 2014).
James Tabery is Associate Professor in the Department of Philosophy and a member of the
Department of Pediatrics and the Department of Internal Medicine at the University of Utah.
Emma Tobin is Senior Lecturer at University College London. She completed her Ph.D. at
Trinity College Dublin in 2006. She worked at Bristol University with the Arts and Humanities
Research Council Metaphysics of Science project. She has authored a number of articles on the
topic of natural kinds.
Dingmar van Eck is Postdoctoral Researcher in Philosophy of Science at the Centre for Logic
and Philosophy of Science, Ghent University. Most of his current research is on issues related
to scientific explanation in the life sciences and engineering sciences.
Petri Ylikoski is Professor of Science and Technology Studies at the Department of Social
Research in the University of Helsinki and Visiting Professor at the Institute for Analytical
Sociology in Linköping University. His research interests include theories of explanation and
evidence, science studies, and social theory. His current research focuses on the integration of
findings from biological sciences (neurosciences, genetics, and evolutionary biology) into the
social sciences and the foundations of mechanism-based social science. His papers have been
published in journals such as Annual Review of Sociology, Philosophical Studies, Erkenntnis, and
Synthese.
Carlos Zednik is Assistant Professor at the Otto von Guericke University in Magdeburg,
Germany. His work concerns the explanatory norms and practices in cognitive science and
neuroscience, as well as the nature of embodied and situated cognition.
xiii
FOREWORD
My aim in this Foreword is to give some background to the New Mechanism, which has arisen
with sources in biology and philosophy, primarily at Chicago, where I started teaching in 1969.
About 15 years ago, Carl Craver referred to us as “The Chicago Mechanists”—comprising
Peter Machamer (who left before I arrived), Lindley Darden, Bill Bechtel, Bob Richardson,
Stuart Glennan, and me. And to some extent, the seeds for these developments emerged in my
classroom.
My education in the early 1960s was framed by two things in tension: three years in engi-
neering physics and a year in industry as a designer of mechanical devices, and by contrast, the
formal approaches of Rudolph Carnap and Carl Hempel of the Deductive-Nomological model of
explanation and the law-based account of reduction of Ernest Nagel. My father, a classical biolo-
gist, talked about tissue structures and mechanisms, but I supposed (as did most philosophers who
thought about it) that the action of mechanisms could be captured by the intersection (often tem-
porally distributed) of multiple causal laws and boundary processes, leaving laws as fundamental.
My first step toward mechanism came in 1972 when, in analyzing function, I classified as
functional new modifications that caused an increase in probability of survival and reproduc-
tion of a system under selection. But this new functional trait wasn’t a law—it was one in an
evolutionary series of causal factors. In a postscript, I saw resonance with Wesley Salmon’s new
account of explanation in terms of statistically relevant factors, which he contrasted with law-
based accounts.
In this period, two sources moved me toward a mechanistic account of explanation. The first
was Herbert Simon’s classic 1962 paper, “The Architecture of Complexity,” which talked about
the evolution of complex systems through the aggregation of stable sub-assemblies, and also
introduced the idea of near-decomposability as a way of dynamically decomposing a complex
system into its parts in terms of relative strength of interaction.
The second and even more direct source for seeing biological explanations as explaining
phenomena in terms of mechanisms was Stuart Kauffman’s 1971 paper, “Articulation of Parts
Explanations in Biology and the Rational Search for Them.” Kauffman argued that we must
first find a way of decomposing a system into parts (there might be multiple such ways), then
constructing a “cybernetic model” using one of these decompositions for how the parts articu-
lated to cause the behavior, and then discovering how these elements were instantiated in the
biological system. Kauffman roughed out that paper in our faculty-student philosophy of biology
xiv
Foreword
discussion group at Chicago in 1970, and presented his paper at PSA-1970. That group included
Lindley Darden and Nancy Maull as students, and evolutionary biologists Richard Lewontin
and Richard Levins, Ken Schaffner, molecular biologist Arnold Ravin, and me among others
as faculty. I count Kauffman’s paper as the first analytical account of mechanistic explanation
in biology. In my PSA-1972 paper, I combined Kauffman’s ideas of multiple decompositions
of a system into parts with Simon’s idea of near-decomposability to ask what happened when
strongly interacting variables or parts from different partial perspectives of the system were
required for a given explanation.
In 1974, mechanisms came to the fore, in a symposium at the PSA meetings on reduction in
biology, with Ken Schaffner, David Hull, and Michael Ruse. The focus of the other papers (pro
or con) was the Nagel-Schaffner model of reduction, but I puzzled that:
At least in biology, most scientists see their work as explaining types of phenomena
by discovering mechanisms, rather than explaining theories by deriving them from or
reducing them to other theories, and this is seen by them as reduction, or as integrally
tied to it.
(Wimsatt 1976, p. 671)
I went on to argue for a mechanistic account, arguing that discovering a mechanism as a rela-
tively stable and manipulable articulation of causal factors better fit the activity of biologists than
a search for laws. This notion of mechanism was developed correlatively with a realist account
of the nature of levels of organization, and argument against a “nothing-but-ist” eliminative
reductionism.
I taught the Simon and Kauffman papers as well as my PSA papers regularly in classes,
and Bill Bechtel and Bob Richardson were influenced, I think, by all of them, as well as by
a 1970 paper on complexity by Richard Levins, where he distinguished between aggregate,
engineered, and evolved systems. Levins’ account of engineered systems fit the description
of paradigmatic mechanistic explanations of the behavior of a system in terms of interac-
tions between its parts. Evolved systems had parts boundaries that were less clear because
they had evolved together, and required the multiple perspectives and parts boundaries of my
1972 paper for their analysis. Simon’s work on near-decomposability probably came through
most strongly in Bechtel and Richardson’s groundbreaking 1993 book, Discovering Complexity,
which involved decomposition, localization, and re-synthesis (of parts as articulated mecha-
nisms) as explanatory strategies. I continued to interact with Bob and Bill, and to encourage
them with their complex and long-maturing book.
My own work turned to other matters until the late 1980s when Stuart Glennan turned up.
I was then deeply involved in simulation modeling with Jeffrey Schank. He and Glennan were
both very talented programmers, with interests intersecting mechanistic explanation. They had
independently discovered the framework and inspiration that object-oriented programming pro-
vided for mechanistic thinking and modeling, and I encouraged them to write about it. Glennan
chose in his dissertation to develop an account of the nature of mechanisms, and put it to use in
developing a novel account of causation. While he had initial troubles getting it accepted, the
last chapter of his dissertation became his now well-known 1996 paper, “Mechanisms and the
Nature of Causation.” The mechanisms literature really took off when Machamer, Darden, and
Craver responded to Bechtel, Richardson, and Glennan in their groundbreaking 2000 paper,
“Thinking about Mechanisms.”
I feel fortunate to have been present as an increasing awareness of the role of mechanisms
as elements of scientific explanation emerged, and to have contributed to that development.
xv
Foreword
Early ideas like Simon’s and Kauffman’s on interlevel articulatory explanations, and of my own
on levels, robustness, aggregativity, multiple complementary decompositions, and emergent
complexities, helped to fuel more detailed and focused articulations by Bechtel, Richardson,
Machamer, Darden, Craver, and Glennan. Our work also involved two important turns with
philosophical import. The first is to take the work and claims of scientists seriously, and to look
at what they can bring to philosophy rather than to suppose that the primary mission of philoso-
phers is to bring edification to scientists. The second is a commitment to realism, which tempers
the investigations of working scientists and should illuminate ours.
The new mechanistic approach has now become a dominant view across a range of debates
in philosophy of science. The chapters in this Handbook analyze the historical, metaphysical,
and epistemological dimensions of mechanistic science, and document the spread of philosophi-
cal work on mechanism through multiple disciplines. These analyses range from chemistry and
(ironically!) physics on up through the biological and social domains where ontological geno-
cide was once promised in the service of eliminative reductionism. Collectively, they show how
mechanistic approaches better fit the kinds of investigations actually pursued in these diverse sci-
ences, while explaining why the search for laws in these areas has for the most part been fruitless.
I welcome the multi-dimensional expansion of mechanistic philosophy of science, and the
varieties of practice in the compositional sciences, from chemistry to the life and social sciences,
that it illuminates. This Handbook is a rich, systematic, and illuminating encyclopedia of what
mechanism has wrought. I am proud to be able to introduce it.
William C. Wimsatt
Peter B. Ritzma Professor of Philosophy, University of Chicago, and
Winton Chair of Liberal Arts, University of Minnesota, USA
Reference
Wimsatt, W. C. (1976) Reductive explanation: a functional account, in A. C. Michalos, C. A. Hooker,
G. Pearce, and R. S. Cohen, eds., PSA-1974 (Boston Studies in the Philosophy of Science, volume 30).
Dordrecht: Reidel, pp. 671–710, reprinted in my (2007) Re-Engineering Philosophy for Limited Beings:
Piecewise Approximations to Reality. Cambridge, MA: Harvard University Press.
xvi
ACKNOWLEDGMENTS
A Handbook like this does not come to be without the work and support of many people,
and we would like to thank them for their help. We are very grateful to Tony Bruce, sen-
ior publisher at Routledge, who approached Stuart with the idea for a Handbook and who
enthusiastically supported his request to bring Phyllis on as co-editor. We are also indebted to
Adam Johnson, who was our primary contact at Routledge and did a great deal to advise and
support us as we brought this project to completion. Additionally, we would like to thank those
who gave us feedback on early plans for the Handbook, including three anonymous reviewers.
We also received valuable advice from Bill Bechtel, Carl Craver, and Lindley Darden.
Our biggest debt, of course, is to our contributors. We are pleased to have assembled such
an accomplished group of scholars, forty-one men and women from all over the world and at
all different stages of their careers. They were attentive to our schedules and responsive to our
feedback, and many of them communicated with each other to maximize the coordination and
impact of their chapters. They have been, without exception, a joy to work with, and if this
book is a success, it is because of them. Also, we thank Bill Wimsatt, who has been a mentor to
many in this book, for writing a Foreword to the volume.
We could not have completed this book at all without the support of those close to home.
Stuart is grateful to his wife Lesley and son Elliot for their encouragement, and for their under-
standing during the nights and weekends taken up with editorial work. He also appreciates the
support of his Butler colleagues, particularly those in the Dean’s office who covered for him
while he slipped into cyberspace for meetings with Phyllis. Phyllis is grateful to her husband
David, parents Sylvia and Ian, and her colleagues in Team Philosophy for their endless support,
and for enduring many mutterings about “the book.”
Finally, the editors would like publicly to acknowledge each other. We collaborated closely
on every stage of this project, and the result is far better than either of us could have achieved
on our own, and we had a lot of fun doing it. Philosophy is always best done together.
xvii
1
INTRODUCTION
Mechanisms and mechanical philosophies
Mechanical philosophy is of ancient origin. For philosophers and scientists in many epochs,
thinking about mechanisms has proven to be a fruitful way to understand nature. Although
mechanical philosophy receded for much of the twentieth century, it is again resurgent; that is
the occasion of this Handbook.
The re-emergence of mechanical philosophy has come chiefly from two directions. The
first, what has come to be called the New Mechanism, has its origins in the work of a number
of philosophers of science (Bechtel & Richardson 1993; Glennan 1996; Machamer et al. 2000;
Wimsatt 1976; Thagard 1999) working in the life sciences – including biology, medicine, and
cognitive and neuroscience. The second is a growing body of literature on mechanisms and
mechanistic explanation in the social sciences, including sociology, political science, economics,
and history (Elster 1989; Hedström & Swedberg 1996; Little 1998). For lack of a better term, we
will call this approach social scientific Mechanism. Although they arose largely independently,
New Mechanism and social scientific Mechanism were both motivated by dissatisfaction with
the philosophical image of logical empiricism. In both the life and social sciences, it seemed
more plausible to construe scientific inquiry as a search for mechanisms than a search for laws
of nature, and to see scientific explanation as causal and mechanistic rather than as a matter of
subsuming phenomena under general laws.
The two strands of mechanist thought also share much in common in their conception of
what mechanisms are. Mechanisms in both cases are conceived to be complex systems or pro-
cesses that are “real and local” (Illari & Williamson 2011). They have parts whose activities and
interactions are responsible for the phenomena that scientists study. Mechanisms are conceived
to have both a “vertical” (constitutive) and “horizontal” (causal) dimension: mechanisms as
wholes do what they do because of the activities of their parts. In the social sciences, these
parts might be individual people, families, or political parties. In the life sciences, they might be
proteins, cells, or organisms. Mechanistic accounts are reductionist in the sense that they take it
that the properties and activities of parts underlie the properties and activities of wholes, but it
is not nothing-but reductionism. Mechanists take seriously the reality of complex things. They
recognize that distinctive scientific domains have distinctive kinds of entities, which engage
in distinctive kinds of activities. For instance, social interactions are not the same as chemical
interactions. Also, in explaining how mechanisms work, one must look to the organization of a
1
mechanism’s parts, and the context in which a mechanism is embedded. As Bechtel has put it,
mechanistic explanation must look “down, around, and up.”
This Handbook is concerned with both mechanisms and mechanical philosophies; that is to
say, it is concerned both with what mechanisms are as things in the world, and with philosophi-
cal or scientific theories or approaches in which mechanisms or mechanistic methods figure
prominently. Since the word “mechanism” is used to refer both to mechanisms and mechanical
philosophies, we will signal this distinction by capitalizing the term when referring to a philo-
sophical or scientific approach, and using lower case when referring to mechanisms themselves.
So, for instance, we will speak of New Mechanism or Cartesian Mechanism, but of economic
mechanisms or the mechanism of protein synthesis.
What, then, is a mechanism? Within both the New Mechanism and social scientific
Mechanism literature there have been quite a number of attempts to offer a working definition
or characterization, but we believe that a common core of assumptions can be succinctly cap-
tured in a formulation we call minimal mechanism:
A mechanism for a phenomenon consists of entities (or parts) whose activities and
interactions are organized so as to be responsible for the phenomenon.
This formulation, which comes from (Glennan forthcoming) and which closely resembles
a proposal from (Illari & Williamson 2012), is minimal in two different senses. First, it is a
generic and permissive definition – one by which a great many things will count as mechanisms.
Second, it represents the common denominator in a set of proposals from new mechanists and
social scientific mechanists.1 We offer minimal mechanism as a characterization not to sug-
gest that questions about the nature of mechanisms are settled – they are not – but rather to
provide a framework for organizing discussions and disagreements about what mechanisms are
and what role they play in the scientific enterprise. Also, we should emphasize that within this
broad category of minimal mechanism there are many kinds of mechanisms, and an informa-
tive philosophical analysis of mechanisms must attend to the important differences between the
many species of mechanisms.
Just as there are many kinds of mechanisms, there are many kinds of mechanical philosophies.
Most broadly we can distinguish them by whether they are concerned primarily with ontologi-
cal and metaphysical questions, or with epistemic and methodological questions. Metaphysical
mechanists are concerned with questions about the constituents and organization of the natural
world. They see mechanisms as part of nature and seek to understand questions about what they
are and how they are organized. They likely see mechanisms as central to understanding some
traditional metaphysical questions, e.g. about the nature of objects and properties, parts and
wholes, causal relations and laws of nature. Metaphysical mechanists operate within a tradition
of natural philosophy that dates to Democritus. Methodological mechanists, on the other hand,
are primarily concerned with questions in the philosophy of science. They understand much
of the scientific enterprise as being concerned with the search for mechanisms, so they seek to
give accounts of how scientists discover mechanisms, and how they represent, refine, and justify
their claims about them, as well as how mechanistic knowledge helps us explain and control
natural phenomena.
A second dimension with respect to which we can classify mechanical philosophies is their
scope. Some, who argue for a narrow conception of what mechanisms are, see mechanistic
science as a particular kind of science that is powerful, but applicable only in limited domains.
Others, who argue for a broader conception, see mechanistic methods as applicable to a far
2
Introduction
wider range of scientific domains. We should emphasize that both of these classifications are
idealizations. Metaphysical and methodological questions are of course not that easy to separate,
and there is no simple ordering of the scope of mechanical philosophies. Nonetheless, it can be
quite helpful in exploring contemporary and historical debates to attend to where philosophers’
views fit within this framework.2
This Handbook is divided into four parts. The first part delves into the history of scientific
and philosophical thinking about mechanisms. The second is focused on the nature of mecha-
nisms themselves; these are areas in which mechanical philosophy addresses basic ontological
and metaphysical questions. The third part turns to methodological and epistemological ques-
tions raised by the mechanistic approach to natural and social sciences. The final and longest part
offers a view of mechanisms and mechanistic approaches in a variety of scientific fields. In the
remainder of this chapter, we will survey the contents of these parts and suggest some important
connections between them.
Part I: Historical perspectives on mechanisms

As we noted above, mechanical philosophy has a long history. The chapters in Part I of this
book recount some important episodes in this history. Understanding these episodes is impor-
tant in part because they are all major debates and developments within the history of science.
But, equally importantly, they allow us to understand both continuities and discontinuities
between contemporary and historical mechanistic approaches to nature and science.
In “Mechanisms: ancient sources,” Tiberiu Popa begins with a discussion of the ancient
sources of mechanistic thought. These commence with Democritus and the ancient atomists
and are taken up by the Epicureans. Perhaps most interestingly, Popa shows how Aristotle,
despite his criticisms of the atomists, nonetheless adopted – within certain domains – forms of
explanation that we would recognize as mechanistic. Sophie Roux, in “From the mechani-
cal philosophy to early modern mechanisms,” examines the origins and meanings of the term
“mechanical philosophy.” She draws on recent historical scholarship that shows that the var-
ious metaphysical and scientific projects carried out in the name of mechanical philosophy
were quite diverse. While she agrees with critics that many mechanistic explanations from the
period do not stand up to scrutiny, she argues that others, especially in the biological realm, are
more empirically credible, and bear noticeable resemblances to contemporary-style mechanistic
explanations. To illustrate this, Roux examines some of Descartes’ uses of mechanism diagrams,
and her points show remarkable parallels between early modern and contemporary techniques
for visually representing mechanisms – techniques addressed in the contemporary mechanisms
literature (see Chapters 18 and 19).
William Goodwin’s “The origins of the reaction mechanism” pays welcome attention to a
discipline not often covered by new mechanists: chemistry. Goodwin examines the develop-
ment of the chemical reaction mechanism, including Lapworth’s work decomposing reactions
into elementary steps. He shows how these decompositions were used to explain kinetic data.
In a complementary chapter, “Mechanism, organicism, and vitalism,” Garland Allen traces the
history of debates about Mechanism, organicism, and vitalism in biology, focusing on devel-
opments in the nineteenth and twentieth centuries. Allen situates this debate within broader
metaphysical debates about materialism and its limits – showing the interplay between experi-
mental and theoretical work and broader cultural and political developments. Finally, Marcin
Miłkowski’s “Mechanisms and the mental” covers the long history of mechanistic models of
the mental, beginning once again with Descartes but now sweeping all the way to cybernetics
3
and contemporary artificial intelligence. Miłkowski explores the interplay between attempts to
build mechanical minds and robots with attempts in cognitive science to understand human and
animal cognition mechanistically.
While these chapters are far from sufficient to paint a full portrait of the history of mechanis-
tic philosophy and science, they do show that mechanistic approaches go a long way back, and
that from the first they were applied to explain a wide variety of natural phenomena, not just
those within the realm of physics. Moreover, these historical episodes show a continued and
lively interplay between metaphysical debates about the nature of the material and living world
and more concrete empirical explorations that developed and used mechanistic approaches.
Part II: The nature of mechanisms

Part II in this book focuses on some core questions and debates about the nature of mechanisms
that have arisen over the last twenty-five years. These chapters flesh out the features that new
mechanists point to in their accounts of mechanisms. The part begins with the editors’ contribu-
tion to the Handbook, “Varieties of mechanisms.” Given that mechanisms in the minimal sense
form a very broad genus, our aim in this chapter is to provide some taxonomic principles for
identifying different species of mechanisms. Investigation into specific varieties of mechanisms
allows us to understand methodological differences in different areas of mechanistic science, as
well as patterns of similarity that cut across very different scientific domains.
One point of almost universal agreement among contemporary advocates of Mechanism
is that mechanisms are identified by their phenomena; that is, they are identified – at least
partially – by what they do or produce or how they act. Justin Garson’s chapter, “Mechanisms,
phenomena, and functions,” explores just what these phenomena are, and how they are used
to individuate mechanisms. As Garson observes, a mechanism’s phenomenon can often be
construed as a function, and mechanisms, construed functionally, have important additional
characteristics – like a distinction between proper function and malfunction. Garson explores
some practical implications of this fact, like the ways in which diseases can be understood as
breakdowns of normally functioning mechanisms.
As the definition of minimal mechanism makes clear, mechanisms have two fundamental
ingredients – the entities (or parts, or components) of which they are composed, and the activi-
ties and interactions that these entities engage in. But while contemporary Mechanists agree
about this basic point, there are many questions about just what these constituents are and how
they are related. These questions are taken up in Marie Kaiser’s chapter, “The components
and boundaries of mechanisms.” A very basic question about these constituents is whether
we should take them to be real things, as opposed to explanatory constructs. And while many
mechanists favor a realist interpretation, that realism must be cognizant of the fact that how one
identifies the constituents of mechanisms depends upon the phenomenon one seeks to under-
stand. There is, for instance, no single way to carve a multicellular organism into parts. Kaiser
also discusses the relative priority of entities and activities. She asks: can one exist without the
other, or is one category ontologically primary?
Several chapters in Part II explore the relationship between mechanisms, causation, and
laws of nature. Lucas Matthews and Jim Tabery, in “Mechanisms and the metaphysics of cau-
sation,” situate new mechanist views about the metaphysics of causation within the broader
background of recent responses to Humean puzzles about causation. They argue that four dif-
ferent approaches to causation that have been connected with the mechanisms literature – from
Anscombe, Lewis, Salmon, and Glennan – all respond to Hume, but with very different desid-
erata. They suggest that understanding these different desiderata can help untangle a number of
4
Introduction
debates about the relation between mechanisms and causes. In “Mechanisms, counterfactuals,
and laws,” Stavros Ioannidis and Stathis Psillos take up the specific question of the relation-
ship between mechanistic and counterfactual conceptions of causation, two approaches that
have been quite popular over the last fifteen years. They argue that all ways of understanding
mechanisms depend in some way on a prior notion of law or counterfactual. Holly Andersen,
in “What would Hume say? Regularities, laws, and mechanisms,” takes up the relationship
between mechanisms, laws, and regularities, critically discussing the debate, and arguing that
there are at least two roles for laws that mechanisms cannot subsume. Together these chapters
show how thinking about mechanisms has both informed and been informed by recent debates
about causation and laws of nature.
In his “Probability and chance in mechanisms,” Marshall Abrams maps out the various
ways that the mechanisms studied in the sciences involve activities and outcomes that can
be characterized by probabilities. As Abrams notes, it is widely acknowledged that many
mechanisms studied in biology and other disciplines are in some sense stochastic, but little
has been done to spell out the various kinds of probabilities that are appropriate to describing
them. Using the example of the stochastic processes involved in bacterial chemotaxis, Abrams
identifies the kinds of probabilities that can be used to characterize the activities of a mecha-
nism’s parts. He also discusses how probabilities can be measured and put to use in models
and simulations of mechanisms.
In “Mechanistic levels, reduction, and emergence,” Mark Povich and Carl Craver turn from
questions of causation to constitution. They examine the concept of mechanistic levels, explain-
ing how they are related to other kinds of levels, and how they are different from realization
relations. They explore the implications of mechanistic levels for our understanding of reduc-
tion and emergence, and mechanistic explanation.
Finally, Emma Tobin, in “Mechanisms and natural kinds,” shows how the popular view that
natural kinds are property clusters sustained by homeostatic mechanisms requires a clearer con-
ception of what counts as a homeostatic mechanism, and at the same time that there is a possible
circularity, because individuation of mechanisms may require appeal to kinds. Both mechanisms
and natural kinds prove hard to define and individuate because of boundary issues, raising ques-
tions about whether these boundaries are real or conventional.
Collectively, the chapters in this part of the book show how the philosophical debate moved
toward an increasingly serious engagement with the metaphysical implications of Mechanism.
They do so both by exploring more carefully just what mechanisms are, and by showing how
a better understanding of the nature of mechanisms can give us purchase on many traditional
questions about the structure of the natural world.
Part III: Mechanisms and the philosophy of science

Part III of the book turns from primarily ontological questions about the nature of mechanisms
to epistemological and methodological questions that arise in the activities of mechanism-
centered science. The part begins with Marta Halina’s chapter on “Mechanistic explanation
and its limits.” Halina sets out the core features of recent models of mechanistic explanation,
and considers the advantages of mechanistic explanations over earlier approaches. She then
explores recent challenges to the new mechanist approach, discussing arguments that suggest
that knowledge of mechanisms is not necessary for genuine explanations, as well as claims
that mechanistic approaches to explanation do not do justice to the centrality of abstrac-
tion and idealization in explanation, and do not account for the generality of some kinds of
scientific explanations.
5
John Matthewson’s “Models of mechanisms” explores the question of what it is to be a

model or representation of a mechanism, situating his account within the broader context of
recent work in the philosophy of science on the nature of modeling. Matthewson shows how
this can both illuminate our conception of mechanistic explanation and add to the literature
on modeling. Abrahamsen, Sheredos, and Bechtel’s “Explaining visually: mechanism diagrams”
focuses on a specific kind of mechanism representation. Drawing primarily on examples from
recent research on circadian rhythms, they explore the many techniques with which scientists
use diagrams to efficiently represent the parts and operations in mechanisms, as well as the spa-
tial, temporal, and causal organization of mechanistic processes. They also discuss how diagrams
complement other forms of mechanism representation.
Lindley Darden’s contribution to this volume is “Strategies for discovering mechanisms.”
Darden shows how mechanism discovery fits within Norwood Russell Hanson’s account of
models of discovery. In her account, she emphasizes the iterative character of the discovery pro-
cess as scientists move back and forth between the phenomena to be explained and the entities
and activities that may be responsible for those phenomena.
Part III concludes with David Kaplan’s contribution, “Mechanisms and dynamical sys-
tems.” Kaplan addresses many questions that have been raised about the relations between
mechanisms and mechanistic explanation and dynamical models. While some have claimed
that dynamical systems theory (DST) provides an alternative framework that can explain
phenomena in a number of domains (including systems biology and neuroscience), Kaplan
argues that DST, while an important tool for describing the behavior of certain systems,
cannot be said to explain them unless the models provide information about the mechanisms
responsible for that behavior.
A prominent feature in the chapters in Part III is that their analyses are based upon substantial
engagement with contemporary scientific practice and literature, especially in the biological
sciences. This, no doubt, is one reason why this research has provided such a promising new lens
on traditional issues in the philosophy of science.
Part IV: Disciplinary perspectives on mechanisms

Our final part is by far the longest, with fourteen chapters. Here, we aim to give space for
authors to show how debates about mechanisms, however they arise, take place in different parts
of the natural, life, social, and engineering sciences. Our selections are tilted toward the biologi-
cal sciences, reflecting the concentration of research in this area, but we have aimed to balance
this with discussions of mechanisms in the social sciences, and in other areas less explored.
The first chapter in this part is Meinard Kuhlmann’s “Mechanisms in physics.” As Kuhlmann
notes, recent philosophical discussions of mechanisms have largely avoided physics. The new
mechanistic approach was born in discussions of the special sciences, and many who have advo-
cated for it see it as a repudiation of earlier philosophical approaches to understanding scientific
theory and practice that are rooted in physics as a paradigm science. In light of this, Kuhlmann
focuses on two questions: first, to what extent do the entities, structures, and processes studied
by physics involve mechanisms? Second, how, if at all, are certain features of the fundamental
physical structure of the world compatible with the ontological presuppositions of mechanistic
approaches in the special sciences?
Lane DesAutels, in “Mechanisms in evolutionary biology,” looks at whether the mechanistic
framework that has been successfully applied in other areas of biology can be helpful in under-
standing biological evolution. DesAutels sets out what he takes to be core metaphysical features
of mechanisms, and examines whether, in this light, natural selection, drift, and mutation count
6
Introduction
as mechanisms. He suggests that there may be strategic benefits to thinking about evolution
mechanistically, regardless of whether there are any mechanisms of evolution.
In “Mechanisms in molecular biology,” Tudor Baetu shows how molecular biology helped
shape New Mechanism, and considers how recent developments in the field are leading to
reconsideration of the nature of molecular mechanisms. Earlier idealizations that characterized
mechanisms, like those responsible for genome replication and expression, as quasi-deterministic
molecular machines have given way to a more nuanced picture in which there are stochastic
elements and noise. Additionally, increased understanding of the organization of cells and their
molecular constituents has shown that assumptions about the modularity of these systems can
be false, and also that the intracellular environment has considerable structure which determines
the rates of molecular interactions. These features imply that understanding these mechanisms
requires new tools, like those provided by systems biology.
In their contribution, “Mechanisms and biomedicine,” Brendan Clarke and Federica Russo
explore the many ways in which “mechanisms are sought, formulated, and used in medicine.”
They do so by reflection on six episodes from past and contemporary biomedicine, ranging
from the evolving understanding of the various effects and applications of aspirin to the mech-
anisms connecting asbestos to lung cancer, and to Semmelweis’ work on puerperal fever in
nineteenth-century Vienna. Their analysis of these episodes shows the epistemic role of evidence
of mechanisms, and suggests the importance of a pluralistic approach to evidence in medicine.
Alan Love studies developmental mechanisms, particularly focusing on the relationship
between lower-level molecular-genetic mechanisms and higher-level cellular-physical mecha-
nisms. Love uses the distinction between these two classes of mechanisms to elucidate the
multi-level processes involved in development. Among other issues, he explores the trade-off
between the explanatory generality gained by investigating molecular-genetic mechanisms that
operate across taxa and the explanatory completeness gained by integrating them with cellular-
physical mechanisms to understand more fully morphological outcomes in specific species.
Viorel Pâslaru’s “Mechanisms in ecology” extends the new mechanistic approach from
its more established areas toward mechanisms responsible for the distribution and abundance
of organisms in their environment. Pâslaru uses research on the shrub Lonicera maackii (Amur
honeysuckle), applying the minimal account of mechanisms to elucidate individual-level mech-
anisms that are generally accepted and used in ecology. He also speculates on the challenges
associated with the analysis of population-level ecological mechanisms.
Rounding off the chapters on biology widely construed is Ingo Brigandt, Sara Green, and
Maureen A. O’Malley’s “Systems biology and mechanistic explanation.” The authors give a
succinct account of the emerging field of systems biology, with an emphasis on its novel tech-
niques for representing and explaining complex biological processes. They show that these
explanatory techniques can be used both to extend the range of mechanistic explanation and to
construct complementary explanations that are non-mechanistic.
The next two chapters are broadly concerned with mental mechanisms. Catherine Stinson
and Jacqueline Sullivan’s “Mechanistic explanation in neuroscience” explores the nature of the
mechanisms studied in neuroscience by tracing the development of theories of learning and
memory. They show that the mechanisms scrutinized by neuroscientists are multi-level and
that understanding them requires the integration of models and theories from multiple scientific
fields. Carlos Zednik, in “Mechanisms in cognitive science,” interprets the best-known model
of explanation in cognitive science, David Marr’s three-levels account, as an early articulation of
a mechanistic approach to explanation. Zednik argues that interpreting Marr’s levels mechanisti-
cally resolves ambiguities in the original account and allows it to be fruitfully extended to cover
research programs in cognitive science that have emerged since Marr’s time.
7
Moving to the social realm, Petri Ylikoski’s “Social mechanisms” draws attention to the
recent work on mechanism-based explanation in the philosophy of the social sciences. Ylikoski
shows how mechanistic explanation has proved to be a useful tool for criticizing existing
research practices and metatheoretical views on the nature of the social scientific enterprise.
Daniel Little, in “Disaggregating historical explanation: the move to social mechanisms,”
extends the social mechanisms approach to historical explanation. Little argues that the mecha-
nistic approach provides a far superior model for causal explanations in history than Hempel’s
covering law model. He suggests that the mechanistic approach fits naturally with current
thinking about the character of causal narratives, and can account for the inevitable difficulties
of making large historical predictions.
In “Mechanisms in economics,” Caterina Marchionni characterizes how mechanisms are
conceived in economics in comparison with the other social sciences, and examines how mech-
anisms are appealed to in philosophical debates about methodological individualism, causal
inference, and the extrapolation of causal claims.
In “Computational mechanisms,” Gualtiero Piccinini charts the changing understanding of
the relationship between mechanisms and computation over the last century. Influenced by
the power of the mathematical theory of Turing machines, many philosophers saw Turing
machines as providing a precise model of mechanistic processes. But in the past twenty years,
the situation has reversed, as philosophers have used an account of mechanisms associated with
New Mechanism to identify the particular features of mechanisms that can perform physical
computations. As Piccinini puts it, it has been a shift from computation explicating mechanism
to mechanism explicating computation.
In the last chapter in our collection, “Mechanisms and engineering science,” Dingmar van
Eck applies the ontological and explanatory framework of New Mechanism to conceptual prob-
lems in engineering science. Van Eck argues that to properly understand mechanistic explanation
in engineering science, one must attend to the different kinds of function with which engineers
are concerned. He then uses this analysis to explore the interplay between the explanatory work
of reverse engineering and the developmental work of engineering design.
The chapters that comprise Part IV of our volume are far from comprehensive, but they give
some inkling of the diversity of domains in which the methods of mechanistic science are fruit-
ful, as well as the challenges to and limits of those methods. These investigations, both of the
distinctive features of mechanisms and mechanistic explanation within a given domain, and of
the surprising points of contact across domains, will, we hope, demonstrate the value of thinking
about mechanisms across the sciences.
This Handbook is the first reference work on contemporary philosophical research on mechanisms
and mechanical philosophy, and we hope that it will serve as a resource to students, phi-
losophers, and scientists. The chapters in this volume provide an entry into a large and diverse
literature, but also break some new ground and point to unresolved questions and avenues for
further research.
As Sophie Roux (Chapter 3) reminds us, when Robert Boyle first popularized the term
“mechanical philosophy” 350 years ago, he lumped together under this moniker a number of
distinct and sometimes conflicting philosophical positions and scientific projects. The same no
doubt can be said of the mechanistic projects of scientists and philosophers today. But now, as
then, we think it safe to say that those projects are worthy of pursuit.
8
Introduction
Notes
1 See Illari & Williamson (2012) for a detailed discussion of the prominent new mechanist formulations,
and an argument for a generic account like minimal mechanism. Hedström & Ylikoski (2010) provide
a survey of proposed definitions from both the new mechanist and social mechanist literatures, and
similarly argues for a great deal of commonality across the projects.
2 There have been a number of attempts to identify varieties of mechanisms and mechanical philosophies,
among them Andersen (2014a, 2014b); Kuorikoski (2010); Levy (2013). See our further discussion in
Chapter 7.
References
Andersen, H. (2014a) “A Field Guide to Mechanisms: Part I.” Philosophy Compass, 4: 274–83.
——. (2014b) “A Field Guide to Mechanisms: Part II.” Philosophy Compass, 9(4): 284–93.
Bechtel, W., & Richardson, R. C. (1993) Discovering Complexity: Decomposition and Localization as Strategies
in Scientific Research. Princeton, NJ: Princeton University Press.
Elster, J. (1989) Nuts and Bolts for the Social Sciences. Cambridge: Cambridge University Press.
Glennan, S. S. (1996) “Mechanisms and the Nature of Causation.” Erkenntnis, 44(1): 49–71.
——. (forthcoming) The New Mechanical Philosophy. Oxford: Oxford University Press.
Hedström, P., & Swedberg, R. (1996) “Social Mechanisms.” Acta Sociologica, 39: 281–308.
Hedström, P., & Ylikoski, P. (2010) “Causal Mechanisms in the Social Sciences.” Annual Review of
Sociology, 36(1): 49–67. DOI: 10.1146/annurev.soc.012809.102632.
Illari, P. M., & Williamson, J. (2011) “Mechanisms Are Real and Local.” In P. McKay Illari, F. Russo, &
J. Williamson (eds) Causality in the Sciences, pp. 818–44. New York: Oxford University Press.
——. (2012) “What Is a Mechanism? Thinking about Mechanisms across the Sciences.” European Journal
for Philosophy of Science, 2(1): 119. DOI: 10.1007/s13194-011-0038-2.
Kuorikoski, J. (2010) “Two Concepts of Mechanism: Componential Causal System and Abstract Form of
Interaction.” International Studies in the Philosophy of Science, 23(2): 1–19.
Levy, A. (2013) “Three Kinds of New Mechanism.” Biology & Philosophy, 28: 99–114.
Little, D. (1998) Microfoundations, Method and Causation: On the Philosophy of the Social Sciences. New
Brunswick, NJ: Transaction Publishers.
Machamer, P., Darden, L., & Craver, C. F. (2000) “Thinking about Mechanisms.” Philosophy of Science,
67(1): 1–25.
Thagard, P. (1999) How Scientists Explain Disease. Princeton, NJ: Princeton University Press.
Wimsatt, W. C. (1976) Reductive Explanation: A Functional Account. In A. C. Michalos, C. A. Hooker,
G. Pearce, & R. S. Cohen (eds) PSA-1974 (Boston Studies in the Philosophy of Science, volume 30),
pp. 671–710. Dordrecht: Reidel.
9
PART I
Historical perspectives on
mechanisms
2
MECHANISMS
Ancient sources
Tiberiu Popa
1. Introduction
A search for possible precursors of our concept of mechanism in ancient texts is potentially
rewarding—that is, if we remain mindful of the original methods and aspirations that informed
those works and we handle our own terminology with caution. At the beginning of an important
article on “Ancient Automata and Mechanical Explanation” (2003), Sylvia Berryman addresses
the use of “mechanical” and “mechanistic” in connection with purely material explanations
based on contact action. The use of these words, she points out, can be sometimes more opaque
than illuminating. One of Berryman’s goals—in that article, in her 2009 book on the “mechani-
cal hypothesis,” and elsewhere—is to reveal the usefulness of such terminology with specific
reference to “a method of investigating the natural world through terms and principles drawn
from the discipline called ‘mechanics’” (2003: 344). I agree with her assessment that a promis-
ing direction of research was largely ignored, while teleology and materialism were regarded
by many and for too long as the only positions in ancient natural philosophy worth studying.
At the same time, I believe that, as long as we mark the semantic scope of these terms
with reasonable clarity, we should be able to apply them profitably and legitimately to ancient
thought from several angles. And that includes the search for early notions—emerging from
questions or concerns which were often significantly different from ours—that might cor-
respond in some measure to contemporary understandings of mechanism. The stipulative
and composite definition (see also Glennan 2013; Craver and Tabery 2015; Chapter 1 of this
volume) I am going to rely on is that a mechanism is a system of entities whose interactions,
organization, and specific activities are responsible for its overall stability or for producing
regular changes. Philosophically interesting treatments of systems—thus understood—were
often instrumental in ancient efforts to articulate topics as diverse as the relations between
chance, necessity, and teleology, self-organization in biological contexts, and the sort of cau-
sality governing mental processes. The relevance of mechanisms to these topics remained
central to early and later modern debates.
As I attempt to sketch a prehistory of the idea of mechanism, I need to be quite parsimo-
nious in my selection. Late antiquity will have to be mostly omitted, given the nature of the
sources for that period, which are less amenable to a brief historical survey. I shall focus on the
claims and arguments of several philosophers and philosophically minded scientists rather than
13
Tiberiu Popa
on more purely scientific studies on mechanics, which might yield less insight into general,
theoretical approaches to causation and to what we would call mechanisms. Much of my discus-
sion is devoted to comprehensive views of nature that mobilize causal explanations in terms of
interactions between component parts and their operations either universally (Democritus and
the Epicureans) or within the confines of some fairly pervasive types of phenomena (Aristotle’s
inquiry into the properties and operations of the two exhalations; pseudo-Aristotle’s study of
the lever principle in Mechanics). I also evaluate the significance of analogies and technological
models in a handful of examples (Aristotle’s biology, Lucretius’ philosophical poem, and the
anonymous On the Cosmos).
2. Early atomism
In studies on ancient philosophy, “mechanistic” most frequently describes the worldviews of
Democritus and of the Epicureans. Are there features, beyond the synonymy of “mechanistic”
and “materialist” in such studies, that warrant a comparison with our concept(s) of mechanism?
Let us start with a look at early atomism. For Democritus of Abdera1 (c. 460–c. 370 bce), each
atom is in important ways much like Parmenides’ one being: ungenerated, imperishable, homo-
geneous. Unlike in Parmenides’ metaphysical doctrine, however, plurality and change are not
self-contradictory notions. The atoms remain intrinsically the same, but can change their spatial
positions with respect to each other. Indeed, all changes are reducible to (re)combinations of
these material principles (67A9, 68A58),2 which are always in motion (67A16, 67A18). Besides,
they are perfectly solid or “full,” in sharp contrast to the void or “the empty” (67A1, 67A7).
The indefinitely many uncuttables differ from each other with regard to shape and size (and
possibly weight, 68B168). Their shapes allow some of them to form clusters or entanglements
more easily, following their collisions—if, for example, they have hook-like appendages or
some are concave and others convex (68A37). Different atomic shapes can also underlie percep-
tible qualities (heat is caused by especially small and sharp atoms, 67A14; round ones are liable
to be perceived as “sweet,” and atoms with many edges can produce the impression of rough-
ness, 68A129), but it is mainly the structures of the aggregates they form that ultimately explain
observable properties (e.g. compressibility, 67A19).
We may find the overall spirit of Democritus’ explanations quite appealing, but he was often
a target for Aristotle, who insisted on the importance of goal-directedness in nature, especially
in the realm of life. He takes Democritus to task for failing to furnish a convincing explanatory
apparatus for both the complexity of organisms and for the regularity that can be found within
species, and for confusedly enlisting spontaneity (to automaton) and chance (tuchē) in his cosmo-
gonic and cosmological accounts. In Physics II.4 (196a24–35; cf. Parts of Animals 641b15–23),
Aristotle objects that “some”—quite clearly the atomists, chiefly Democritus—suppose incon-
sistently that our world, much like other worlds in this totality or universe, has come about
through the agency of a vortex which was generated spontaneously, whereas they do not take
animals and plants to be produced by chance or spontaneously. The surviving fragments and tes-
timonia (e.g. 68A66, 68A68) do not convey a clear picture of the relationship between chance
and necessity in Democritus, but the two concepts are probably complementary. As Glennan
(2010: 621; cf. Johnson 2009 and Gregory 2013: 464–68) puts it,
Democritus . . . insists that every event happens from necessity, in the sense of having
sufficient antecedent conditions to bring about the event; but that event can still be,
from another perspective, a chance event, in that it does not happen for a purpose.
14
Mechanisms
Democritus seems to have tried to mitigate the potential discrepancy between our experience
(we can easily notice the repeatability and predictability of many kinds of phenomena) and
his insistence on chance encounters between atoms, by claiming that a sort of self-selection of
atoms with comparable or corresponding shapes and sizes is possible. Just as irrational (alogōn)
animals tend to seek the company of animals of the same kind, inanimate (apsuchōn) things
are sorted by size and shape, as when we use a sieve or a winnowing basket to sort grains or
as when the waves gather long pebbles and round ones in different places. Fragment 68B164
does not spell out the details of this “mechanism” for selection, but the apparent mutual
“attraction”3 of the atoms is likely due to, say, hooked atoms coming together by chance and
clinging to each other or convex and concave atoms happening to collide with one another
(68A37, cited above), their congregations producing an enormous number of phenomena,
including the operations of our minds.
If our interpretive standard is adherence to the principles of mechanics as a scientific
domain, Democritus’ physical theory is hardly mechanistic, as Berryman rightly argues
(2009: 34–9); indeed, mechanics proper was yet to come into existence. Still, his under-
standing of causation—based on assumptions about the nature of solid particles, their
collisions (plēgai) and vibration (palmos), and their entanglements—looks decidedly mecha-
nistic in a more general sense of the word. It bears mentioning that his explanations apply
at various scales, from the formation and disintegration of entire worlds (kosmoi) to physi-
ological and psychological processes. Some causal mechanisms appear thus to be nestled
within larger ones, according to hierarchies that are transparently hinted at in the extant
fragments or testimonia, although they are not explicitly analyzed there.
Among the phenomena caused by interactions between atoms and compounds of atoms,
psychological processes are of special interest. In his survey of earlier theories about the soul
as a motive force (On the Soul I.2, 405a5–13 = 68A101), Aristotle notes that, for Democritus,
the mind (nous) or the soul (psuchē) is an assemblage of round atoms, their fine grain and their
roundness ensuring extraordinary mobility. It is not implausible that such a conglomeration
of round atoms might be the result of a process of self-selection, like the one described in
the passage that mentions the sieve and the waves. In the next chapter of On the Soul (I.3,
406b15–22), Aristotle brings up Democritus again and attributes to him the view that what
causally explains the movement of the soul, namely the continuous motion of its round atoms,
also accounts for the movements of the body (which is set in motion by the soul). To stress
the weakness of this argument, Aristotle writes that Democritus’ explanation is reminiscent of
the way in which—according to Philippus, an author of comedies—Daedalus made a wooden
statue of Aphrodite move as if by itself by pouring quicksilver into it. The problem with
Democritus’ approach is that it is logically vulnerable (how can rest depend on those moving
atoms?) and also that we seem able to move sometimes deliberately and as a result of reflection,
not merely because of chance happenings. The fact that Aristotle ridicules the atomists’ view
of self-motion is clear enough and important in itself, but the very comparison between the
workings of the soul in Democritus’ theory and an automaton, such as an automaton fashioned
by a mythical figure, suggests, I think, that even in classical times Democritean explanations
were deemed “mechanistic” in the sense that an organized assemblage of parts (of atoms and
of structured collections of atoms) and its activities could produce an observable phenomenon.
Democritus’ direct and indirect influence was to be profound and lasting, from Epicurus
(341–270 bce) to Gassendi and beyond. The avatars of his doctrine include the famous
Epicurean qualification to the Democritean absolute necessity: the uncaused swerve, one of
the most distinctive features of Epicurus’ doctrine.4
15
Tiberiu Popa
3. The Epicureans
We have seen that the first atomists explained psychological processes and states by appeal
to atomic motions and the shapes of atoms, in keeping with their dominant metaphysi-
cal dogmas. The Epicureans, while embracing the spirit of such accounts, made a number
of original contributions, most notably with respect to their treatment of will. Indeed,
Lucretius’ (who wrote around the middle of the first century bce) memorable section on
the swerve5 makes it clear that Epicurus’ innovation is meant partly to salvage the concept
of (free) will. It is also presumably crucial to understanding how aggregates of atoms can
ever form, since the atoms tend to “fall” along parallel trajectories and do so, importantly,
at equal speed: there is no resistance in void (see Epicurus, Letter to Herodotus, 61), as there
would be for a body falling through air or water. The speed of bodies falling in air or water
varies in proportion to their weights and to the resistance of the medium in which they
move (II.225–39). Hence, no atom could possibly catch up with other atoms in a void.
The fact that there are compounds of atoms, however, compels Epicurus—and his follower,
Lucretius—to posit the swerve, a slight and uncaused deviation of the atoms from their
parallel trajectories that allows them to collide with each other.
When such swerves take place in the (material) soul, they can cause chain reactions which
propagate through the body and instigate our actions. Significant though this is for Epicurean
physics in general—the formation of conglomerates of atoms—and for elucidating general traits
of animal behavior, this explanatory device was, again, probably meant mainly to justify any
reference to intentional action and moral responsibility.6 What is noteworthy for our purpose
is that, in Lucretius’ celebrated passage in Book II, psychological aspects like will (voluntas) and
desire (voluptas) and their manifestations (e.g. our going against the movement of a crowd which
surrounds us, II.277–83) are tightly connected with notions that were pivotal to theoretical and
applied mechanics—weight, collision, air- or water-resistance, etc. Accordingly, speaking of
mechanistic explanations in the case of this evolved version of atomism is legitimate for the rea-
son just stated as well as for its general compatibility with our conceptions of mechanisms—even
when no explicit technological models come into play.
There is one analogy with a “machine” in Lucretius, by the way, that is rather well known
and worth a glance. In Book V he reminds us that the enormous mass, moles, and machinery
of the world, machina mundi (96), with its threefold constitution (sea, earth, sky), is not perma-
nent, despite appearances and false expectations. The formulation seems intended to emphasize
at the same time vastness and order—and maybe complexity.7 Berryman (2009: 38–9) finds
this formulation unimpressive: “The meaning of machina here could be little more than a
vague sense of an arrangement or system, perhaps in contrast to mere undifferentiated mass;
there is no suggestion of technology here.” I think it is helpful, however, to take into account
some of the preceding lines (73–90 in Book V; cf. Epicurus’ Letter to Herodotus 81, which that
passage may reflect), where Lucretius raises the possibility that people’s feeble understanding
of the laws (leges, foedera) of nature might lead them to embrace superstition and to believe
that the heavenly bodies, with their wondrously regular revolutions, are divinities (or at least
that they are steered by divinities). He appears to mark there a distinction between animate,
divine celestial bodies, and a view of the universe as a largely orderly and dynamic mechanism.
Incidentally, Lucretius’ insistence on the order and regularity of nature is one of the most
prevalent motifs in this philosophical poem. A contrast between (a special sort of) living beings
moving spontaneously and a functioning mechanical device is thus probably implied here.
Finally, and more to the point of the present chapter, this analogy is very much in keeping with
the whole Lucretian poem and Epicurean outlook: all phenomena, from the revolutions of the
16
Mechanisms
stars to our most intimate thoughts, are explainable through the causal relations between and
within the vast realms of nature, and are reducible to interactions between indivisibles and to
the “laws” that govern their behavior.
4. Aristotle on the two exhalations

The atomists argued that the world is an infinite swarm of discrete minimal particles moving in
“the empty.” Aristotle (384–22 bce) begged to differ. For him, the finite universe is a plenum
consisting of five simple stuffs, one of them, the aithēr, moving in circular fashion; the other four,
which make up the sublunary world, having naturally a rectilinear motion, toward or away from
the center of the universe. Perceptible bodies are continua, as are movement and time. He also
decried the sheer absence of anything like his natural teleology in early atomism, an error that,
he thought, inevitably led to grave contradictions. In light of his demolition of Democritus’
physics,8 among other things, one can see why the term “mechanistic” is not generally applied
to Aristotle’s worldview. Yet, we can ask, are at least some aspects of Aristotle’s scientific works
or some of his models mechanistic in a sense that accords in part with the notion of mechanism
explored in this Handbook? I start with his “meteorology,” a domain where natural teleology
is marginal or absent and which may rather aptly be considered to accommodate systematic
appeals to and analyses of mechanisms, to use a deliberate anachronism, and then I tackle a few
controversial passages from his biological writings.
Aristotle’s interest in mechanics proper is no doubt limited (whatever mechanics, mēchanikē,
might have meant for him),9 but he does dwell at length on what we might call mechanisms in
a more comprehensive sense. He devotes three books, Meteorology I–III, to a quasi-classification,
description, and, especially, causal explanation of “meteorological” processes. They include
what we would recognize as meteorological phenomena, but their scope extends far beyond
rain, frost, or wind to include earthquakes, comets, and the Milky Way, all of them occurring, as
he thought, in the sublunary sphere. The unifying principle of this inquiry is the theory accord-
ing to which rainbows and mock suns, rain, and the formation of minerals etc. are caused by the
inherent nature of and also by the interaction between two exhalations (anathumiaseis) present in
the sublunary world: one is dry and smoke-like and emerges when the soil is sufficiently heated
by the sun; the other is moist and vaporous and is due chiefly to the evaporation of water. The
former rises to become a layer in the proximity of the first celestial sphere and can be easily
ignited by it. The latter is situated generally below the dry exhalation and is cooler. The two are
normally combined, but one can predominate to various degrees in different regions.
Their behavior is described notably in mechanical terms,10 often by reference to
objects which are ejected under intense pressure. Hot air can be expelled from a certain region
when the surrounding cooled air contracts, as one hurls a projectile (4.341b37–342a3, 8); it
moves much as a fruit stone is ejected when squeezed between fingertips (4.342a10–11). To
take another example from ch. 4 in the first book, to bolster his demonstration of what causes
a shooting star, Aristotle wonders whether that process might not be comparable to an exhala-
tion coming from below two lamps, one placed on top of the other. The exhalation causes “the
lower lamp to be lit from the flame of the upper; the speed at which this happens is astonishing
and is more similar to a projection than to fires being lit in succession” (342a2–7). If this was an
experiment carried out by Aristotle, which is not unlikely but is impossible to ascertain, the two
lamps and the interaction between them and their properties can be regarded as a rudimentary
mechanism (and a sort of testing model for a natural phenomenon) which produces and explains
the second flame, in the lower lamp. If so—or even if this is just a thought experiment, used as
a model for what we might consider a mechanism—the natural process itself which is supposed
17
Tiberiu Popa
to be thus illustrated and clarified, the production of shooting stars, is a kind of mechanism,
with identifiable parts and specific activities. Aristotle did not use a technical term for such
systems, but his analysis in Meteorology I–III of phenomena (the explananda) and of the objects
and processes responsible for their generation and defining attributes seems to share some gen-
eral concerns with modern treatments of mechanisms. His conceptual framework was decidedly
different from ours; yet, his effort in that treatise to establish the common and fundamental
causal principles of natural systems which could produce extraordinarily different phenomena is
an exploit worthy of our admiration, even if many of his conclusions are spectacularly wrong.
Indeed, throughout the first three books of the Meteorology we find explanations based on
the two exhalations and resorting consistently to thermal and mechanical processes. Cooling,
condensation, increased pressure, combustion, and rapid ejection are the main elements of
Aristotle’s accounts, for instance, of comets and shooting stars (take, e.g., Meteorology I.7, where
the movement and appearance of shooting stars are compared to what happens when one
throws a lit torch into a large amount of chaff, 344a25–8). The quantity of dry or moist exhala-
tion present in a certain region, and its shape and dynamics, are crucial to the production of
phenomena which involve the display of particular colors, movements, and shapes rather than
others. The Milky Way, therefore, does not look and behave like a comet, although the causes
of the two phenomena are related, as we read in Meteorology I.8. Generally, modified condi-
tions will be responsible for correspondingly altered phenomena. He makes this point plainly
at I.4.341b23 ff. (see also Meteor II.9.370a29–33) where he notes that, under the most favorable
conditions, a particular kind of smoky exhalation he calls “fire,” pur, is ignited by the move-
ment of the innermost celestial sphere. The exact outcome differs (diapherei), depending on
the position (thesin) and on the quantity (plēthos) of the flammable mixture. If that amount of
“fire” extends both broadwise and lengthwise, what we often see is a flame similar to the one
produced when the stubble on a farmland is set on fire. If, on the other hand, the combustible
material only extends lengthwise, the result will be “torches,” “goats,” and “shooting stars” (he
then proceeds to distinguish the respective characteristics of these phenomena).
The behavior of the two sorts of exhalation constitutes a mechanism not just because
of the nature of his analogies with various artifacts, but also and principally because of his
explicit and consistent reliance on what will later become core elements of mechanics,
including ballistics and pneumatics: trajectories of projectiles, relation between forces (or
powers), the effects of increased pressure, etc. Besides, I would venture to suggest that his
causal explanations can also be said to involve mechanisms roughly in the sense explored in
this volume, even if their components (masses of exhalations, positioned, shaped, and inter-
acting in specific ways and productive of corresponding phenomena) do not form very stable
and conspicuously organized systems.
5. Aristotle’s biology
Given the pride of place claimed by qualitative changes and also by formal and final causes in
many11 of his works on the science of animals, the rarity of analogies between the workings of
living organisms and mechanical devices is no surprise. Such analogies do exist, however, and
are relevant here. In ch. 7 of On the Movement of Animals, Aristotle invokes a number of dunameis
or capacities of the soul, including sense-perception, desire, and phantasia (very roughly, capac-
ity for mental representation), to explain how movement is initiated in animals. As a result of
the operations of those dunameis, certain alterations (alloiōseis, 701b18) or qualitative changes (not
reducible to rearrangements of particles, as the atomists would have held) take place and lead to
movement. The original impetus for movement is an almost imperceptible change in the region
18
Mechanisms
of the heart, which can become warmer or cooler as a result of our desires or of our perception
of imminent danger, etc. At 7.701b2–10, he compares the movement of animals to little car-
riages12 ridden by children and to automatic mechanical devices (in Greek automata) which start
moving as a result of a slight change but can subsequently sustain their movement in virtue of
their internal organization. This is the case, we read, when some strings are released and wooden
pegs in it strike one another. The analogy continues with a more detailed comparison between
organic parts and components of automata (and perhaps of the toy carriages): the sinews and
the bones correspond to the wooden pegs and the pieces of iron, while the sinews, which can
be relaxed to allow movement, are analogous to strings, which can be loosened by a puppeteer.
This analogy, then, works with respect to the interaction between (bodily or mechanical) parts,
the initiation of movements by a slight original change responsible for a series of consecutive
changes, and the similarity between genuine self-motion and the seemingly autonomous move-
ment of automata. One should not lose sight, however, of the fact that Aristotle urges us to also
keep in mind the striking difference between the two comparanda: mechanical gears do not
need to undergo qualitative changes to perform their respective functions.
Another notable analogy is found in the second book of Aristotle’s Generation of Animals, in
chapters 1 and 5. Some of the difficulties contemplated and solved in the first chapter have to do
with whether the semen itself contains any actual parts of the future embryo (the answer is cat-
egorically negative, since the matter for the embryo is contributed by the mother) and whether
it is the father or rather the seed produced by the father that is responsible for informing the
matter. (For current views on mechanisms and developmental biology, see Chapter 25.) Just as
the father “moves” the sperm, the latter moves, or informs, the matter. To clarify this, Aristotle
musters again the imagery of mechanical devices and notes that:
Their parts, even while at rest, have in them somehow or other a capacity (dunamin),
and when something from outside moves the first part, then immediately the next
part comes to be in actuality. Hence, as with the automata, in one way that [external
mover] moves it, not by being in contact with it anywhere now, but by having at
one time been in contact with it; so too that from which the semen originally came,
or that which made the semen [moves it], namely not by being in contact with it
still, but by having once been in contact with it at some point. In another way, it
is the movement within [which moves it], just as the activity of building causes the
house to get built.
(Generation of Animals II.1.734b11–18, trans. Peck, modified)
The automatic puppets (ta automata tōn thaumatōn; thaumata means literally “wonders” or
“marvels”) seem to be regarded as suggestive analogs, then, in so far as they can sustain their
movement, following an initial impetus, and can, moreover, move in ways that differ signifi-
cantly from the nature of that impetus. According to Henry’s reading, which I find plausible,
what is explained here is not the development of the embryo, but the moving capacity of the
sperm. The analogy with an automaton is intended “to show how the sperm can continue to
move once it is no longer in contact with the father and how the father can still be said to fashion
the matter without being in contact with it” (2005: 31).
In ch. 5 (741b7–15) Aristotle turns his attention to the development of the fledgling organ-
ism. His main point is that the parts that will become distinguishable in the embryo are present
in matter potentially and they begin to emerge when a principle of motion (archē kinēseōs), present
in the semen, comes into play. When the matter has been “moved” in this manner, one stage
follows another in a continuous process comparable with the movements of automatic wonders
19
Tiberiu Popa
or marvels (automatois thaumasi). The coming about of the actual parts of the embryo, beginning
with the heart, involves successive qualitative changes (741b12). The mention of changes with
respect to differentiae such as softness, hardness, and color alerts us to the fact that the preceding
analogy (signaling spatial, not qualitative, changes) should not be pushed too far, lest it become
more misleading than elucidating. I take this implicit concern to reflect Aristotle’s caveat in On
the Movement of Animals 7, although there is also a potentially significant difference between the
two texts with regard to his handling of analogies. Henry argues (2005: 38) that, unlike in On the
Movement of Animals, the internal motion of the sort of automata brought up here, in Generation
of Animals, “is the actualisation of a single potential rather than a causal sequence passing through
a series of mechanical gears.” A (hypothetical)13 self-moving automaton would thus be a more
appropriate analog for a developing embryo than a mechanical puppet whose activity is gener-
ated by an external principle and consists of a series of distinct events (causes and effects), not of
different stages in a continuous change.
Aristotle’s use of this analogy was to prove inspiring, if controversial, and was often discussed
in antiquity and later on, but his attitude toward analogies with mechanical devices is quite
ambivalent. They can call attention to the relation between an original impetus and self-motion
and to the functions of and interactions between the parts of an organism, but they can also be
misleading in so far as models based on mechanical processes only convey part of the story, and
arguably not the most important one.14
6. Mechanica
The Mechanics (or Mechanical Problems, in Greek: Mēchanika; Mechanica in its Latin version) has
been traditionally included in the Aristotelian corpus. While its authorship is dubious—if not
spurious, it is not incompatible with the Stagirite’s natural philosophy (as Heath suggested as
early as 1921, vol. 1: 344–6). It seems safe to ascribe it to the Peripatetic school in the late
fourth or early third century bce, possibly to Strato of Lampsacus. What makes the Mechanics
especially interesting is its application of geometrical explanations to physical phenomena and
to their consequences for us (e.g. he points out that larger balances indicate weight with greater
accuracy, 849b23, which is explainable in the final analysis by appeal to abstract models, such
as concentric circles). Mechanics frames a host of problems—related to the lever (and, in smaller
measure, to pulleys and windlasses, which turn out to illustrate the same principle)—in a way
that emphasizes their mutual similarities and the author’s belief that they can be expressed in
a simple, unifying mathematical language. He stresses the dependence of mechanics both on
mathematics (with respect to the “how,” the method) and on “physics” or rather natural phi-
losophy (with regard to the object of study, “that about which,” 847a24–9). The manifest and
sustained interest in tracking down what is common to a variety of physical phenomena and in
revealing the philosophically significant points of this investigation makes this treatise a land-
mark in the history of mechanics, which has only some timid, unsystematic forerunners, judging
by the extant evidence.
What distinguishes this text from some of the later and, in some cases, more impressive
contributions to advances in mechanics is that, owing to the author’s avowed interest in natural
philosophy, this constitutes an early self-conscious attempt to articulate a theoretical treatment
of what a mechanism (of a particular but very common kind, the lever) is essentially. The main
topic of this work and some of the main questions it answers are summarized at the outset of
ch. 3. The central paradox which needs to be elucidated in the Mechanics is that great weights
can be moved by using a comparatively small force.15 Here is the author’s concise description
of how a lever (mochlos) works:
20
Mechanisms
Since under the same weight the greater radius from the center moves more rapidly,
and there are three elements in the lever—the fulcrum, that is the cord16 or center, and
the two weights, the one which causes the movement, and the one that is moved—the
ratio of the weight moved to the weight moving it is, then, the inverse ratio of the dis-
tance from the center. Now the greater the distance from the fulcrum, the more easily
it will move. The reason has been given before that the point further from the center
describes the greater circle, so that by the use of the same force, when the mover is
farther from the lever, it will cause a greater movement.
(850a37–b7, translation by Hett (1936), with slight modifications)
As a mechanical principle, a lever can be instantiated, according to this work, in the arm of
a balance, in the movement of an oar, in an aggregate of wheels in contact with each other,
and, more abstractly, in the radius of a circle, or the radii of a set of concentric circles. Note,
however, that this passage deliberately ignores the nature of the lever as an instrument which
is meant to transmit some force; instead, the reader’s attention is focused on the three aspects
which, irrespective of the specific type of lever used and the specific conditions under which the
phenomenon is produced, can reduce the apparent paradox stated earlier to ratios that are easy
to grasp and to formulate in generic fashion. This principle governs an immense range of seem-
ingly incongruent types of processes, from the functioning of scales in markets (chs. 1–3) and
the force imparted by oars (chs. 4–5) to the ratio between the height of the yard-arm on a ship
and the ship’s speed (assuming a wind of constant strength, ch. 6), and from the movement of
the potter’s wheel (ch. 8) to the distance covered by a missile launched with a sling. Some of the
illustrations offered there are about as delightfully or cringingly colorful as mechanics can ever
aspire to get (ch. 21: Why is it easier for dentists to remove teeth with the help of a forceps than
with their bare hands?). Even certain natural phenomena—for instance, pebbles being shaped by
the waves—conform to the same basic explanatory scheme, which can be formulated in discar-
nate mathematical terms (the relation between circles, etc.). In short, this is a superbly clear and
very influential17 account of the fundamental (interacting) components of a type of mechanism
that can be found under the guise of many categories of artificial and natural processes.
7. On the Cosmos
The two most pervasive analogies involving ancient references to technology are those pertain-
ing (A) to the nature of the entire cosmos (see also my comments on Lucretius), and (B) to the
structures and functions of living organisms, to their parts and generation (we have seen some
early examples in Aristotle’s biological works).18 Such analogies became more frequent with the
acceleration of progress in technology in Hellenistic and Roman times.19
Let me mention one final example—belonging to the first category and quite typical of
post-classical thought in its systematic blend of tenets that can be traced back to different, not
always compatible, doctrines. The short exhortative treatise entitled On the Cosmos was assigned
to Aristotle, but was probably written at some point during the first century ce. It deals suc-
cinctly with the structure of the heavens, various regions of the earth, and especially with the
relation between the world and its divine keeper (as well as his attributes). Its substance seems
to be based on earlier scientific compendia and betrays an eclectic philosophical background,
borrowing especially from Aristotle’s cosmology and “meteorology.”20 The universe as a whole
is imperishable (4.397b8), but the region around the earth is eminently a domain of change. In
ch. 5, the author claims that the overall harmony of the world depends on the mutual balancing
of opposites, an idea he explicitly attributes to Heraclitus. The interactions between the powers
21
Tiberiu Popa
of the elements as well as between the exhalations are responsible for the, admittedly imperfect
and fluid, order of the sublunary world. This order, however, is ultimately preserved by a god
who acts remotely—from the outermost celestial sphere—through successive powers, dunameis,
which relay his plans with diminishing degrees of accuracy; the earth is, accordingly, a far less
glorious example of harmony than the heavens.
We can find in this theology and teleology echoes, distorted though they might be, from
Aristotle, among others. To make the divine portrait more vivid and to justify this god’s aloof-
ness, while also arguing for his causal efficacy, the author uses several analogies. After comparing
the divinity with the Persian Great King, he embarks on a double analogy of the universe with
a piece of machinery (Furley 1955: 390–1, note a, argues it might be a catapult or a ballista;
cf. Thom 2014: 117) and with a puppet (an analogy rather reminiscent of several passages in
Aristotle’s biological corpus—see above). The divinity is able
to produce all kinds of forms with ease and a simple movement, just as indeed the
engineers (megalotechnoi) do, producing by means of the single release mechanism of an
engine of war many varied activities. In the same way puppeteers, by pulling a single
string, make the neck and hand and shoulder and eye and sometimes all the parts of
the creature move with a rhythmical movement.
(Trans. Thom 2014: 45)
The image of a mechanical device is also evoked by the more dynamic notion of the transmis-
sion of change21 along an assemblage of components and conveying powers, which seem to be
semi-autonomous once the ruler and begetter of all things (399a31) provides his initial impulse.
What is vigorously emphasized in On the Cosmos, therefore, is that the divine keeper does not
have to tend to every single component of the cosmic system and to direct every stage in its
operation. We have here the picture of a world that, while imbued with divine rationality, is
capable of producing its own phenomena, even if the long and complex causal chains behind
them lead ultimately beyond nature itself.
8. Conclusion
Recent debates about the essential features of mechanisms and about the relation between
mechanisms and laws of nature cannot find, of course, any close antecedents in the texts and
theories just surveyed in this chapter. We can track down, nonetheless, a few attempts to caus-
ally and comprehensively explain natural events in terms of the interplay among the component
parts of some natural system and the activities inherent in it. For the authors discussed here,
the discovery of causes and their handling of explanations follow sometimes radically different
methods. They investigate systems—mechanisms avant la letter—of varying degrees of complex-
ity (instantiations of the lever principle, for instance, are simpler than the formation, structure,
and operations of organisms).
Some treatments of explanatory mechanisms are quite self-conscious and crisply articulated;
others are vaguer and fairly tentative. Attitudes concerning the appeal to technological models
and analogies meant to represent what we would call mechanisms were determined, sometimes
in complicated ways, by various philosophical doctrines (natural or providential teleology; dis-
tinct approaches to more particular problems, such as the nature of processes characteristic of
embryogenesis, etc.). A host of other, more elusive, factors too must have contributed to the
shaping of those attitudes: evolving concepts, shifting cultural contexts, polemical stance, and
even literary tastes.
22
Mechanisms
There is no robust unity, then, that we can hope to find among these possible and distant
precursors of our conception(s) of mechanisms. A few general common points, however, can
be identified. The texts examined here tend to mark the phenomena to be explained22 before
distinguishing entities (e.g., uncuttables, exhalations, simple and complex parts of organisms) and
activities (collision, uncaused swerve, friction, compression, expulsion, slight “movements” in
the region of the heart, relaxation or tightening of sinews, etc.) which, when organized in the
right ways, are conducive to changes, often to regular changes. New phenomena can be gener-
ated if the conditions—i.e. the organization of the entities making up a mechanism and their
activities—change. The atomists’ natural philosophy and Aristotle’s “meteorology,” for example,
richly illustrate this aspect. Most of those accounts tend to elucidate: (1) the relation between
some original impulse and genuinely or apparently autonomous activity either in the case of indi-
vidual organisms or of automata or even of the entire cosmos (whose order is ensured remotely
by a transcendent divinity); (2) implicitly, the transmission and persistence of the force impressed
by that impetus with a higher or lower degree of accuracy; and (3) the interaction between parts:
minimal material particles and aggregates of particles, continuous masses of exhalations, celestial
spheres and the material constituents of the sublunary world—all this under conditions that hinge
on mechanical processes (foreshadowing or following central tenets in ballistics and pneumatics).
Beyond this, we would be hard pressed to find a consensus on what a mechanism is in general.
Such dissonance or rather polyphony, however, should be no cause for disappointment. After all,
the search for full clarity and consensus is still underway, as this whole volume bears witness.23
Notes
1 Democritus probably followed Leucippus quite faithfully with regard to his physical doctrine, which
he further developed; Aristotle and several late ancient Aristotelian commentators, among others, often
mention Leucippus and Democritus in the same breath.
2 My references to fragments from and testimonia about Democritus follow the Diels-Kranz numbering
system.
3 Taylor (1999: 194) believes that there is some evidence that Democritus invoked several types of—
attractive and repulsive—forces. Gregory (2013: 449–54) raises a number of serious objections to this
interpretation.
4 Epicurus also departed from Democritean physics by rejecting the notion that the atoms are of infinitely
many sizes and shapes.
5 Clinamen, translating the Greek paregklisis; II. 216–93 in Lucretius’ De rerum natura.
6 Epicurus rejects the notion that everything is ruled by fate in his Letter to Menoeceus, 133–4.
7 Bailey (1966: 1335–6): “[A]gain a careful expression; the mundus is not only a mass of matter, moles, but
a complex construction, machina.”
8 Aristotle also virulently criticizes Plato’s “geometric atomism,” marred by absurdities like assigning
weight to bodies analyzable into planes (the two types of triangles posited in Timaeus 53c ff.). Plato’s
image of the cosmos as a whole is that of a living being (a reflection of its transcendent model, the sum
total of ideas), rather than of a mechanical device.
9 There is a mention in Posterior Analytics (I.13.78b37) of mechanics as a scientific domain subordinate
to stereometry, but he does not elaborate on it. Modern references to Aristotelian “mechanics” gener-
ally point to his discussion of pushing, pulling, and the bodies’ resistance to such actions and forces
(especially the ship-hauling demonstration and the rare quantitative formulas conveying the ratio
between velocity and force in Physics VII.5), and to his comments on impetus, the movement of a
projectile after its release and the presumed role of the medium in sustaining that movement.
10 For a detailed analysis of the quasi-technical terminology in Meteorology I–III, see Wilson’s section on
“Aristotle’s lexicon of mechanics” (2013: 65–70).
11 Many, not all: his natural teleology is virtually absent in treatises like History of Animals, where he studies
morphological and other differentiae (“the fact”) to prepare the terrain for inquiries into the causes of
those differentiae (“the reason why,” the cardinal feature of other works, including Parts of Animals and
Generation of Animals).
23
Tiberiu Popa
12 The toy carriages tend to move in a circle because the wheels on one side are smaller than on the other
side. Nussbaum (1978: 347) notes that “The difference between the two examples seems to be primar-
ily one of emphasis: the puppet example underlines the generation of a whole series of motions from a
single initial motion, the cart example the change in character or a motion because of the nature of the
functioning mechanism.”
13 Henry (2005: 38–40) believes that the two passages in Generation of Animals present a thought experi-
ment and could not refer to an actual device.
14 Recently, there have been some very interesting attempts, however, to demonstrate that Aristotle’s
biology relies to a more significant extent on the application of mechanical principles: De Groot
(2014 (especially 107–62)) and Johnson (forthcoming).
15 Or: power; I am not implying any unduly modern connotations here.
16 I.e. the cord on which, for instance, the beam of a balance is suspended.
17 Its impact is evident not just in later ancient works on mechanics, but also at the dawn of modern sci-
ence, when this work was widely read.
18 On these two categories of analogies, see, among others, De Solla Price 1964 and Berryman 2009. A few
remarkable examples of analogies between biological processes and the functioning of automata can be
found e.g. in Galen’s On Seed I.5 and in his On the Development of the Fetus 6.3–4; for useful clarifications
of these passages and of their connections with Aristotle’s own analogies (discussed in this chapter), see
Preus 1977 and Berryman 2002.
19 We have both reports of geared planetaria, complex water-clocks, automata meant to simulate not
just the appearance, but also aspects of the behavior of human beings and of non-human animals, and
tantalizing archeological evidence (the Tower of Winds in Athens, the Antikythera mechanism, etc.).
Nonetheless, the fragmentary nature of the evidence hampers any attempt to outline a coherent his-
tory of ancient mechanisms, to properly reconstruct many of them and to attribute their invention or
production with reasonable confidence. Berryman argues that, unlike mythical or imaginary examples
(e.g. the statues of Daedalus), “real devices offer evidence of the kinds of results that can be achieved
by mechanical craftsmanship. Because it seemed that technology could replicate some of the kinds of
features taken to be distinctive of living things, the argument that there is a distinction in kind between
living and nonliving things, and in the explanantia required for the former, faced a formidable challenge”
(2003: 365–6).
20 See his discussion of the two exhalations (anathumiaseis 4.394a9 ff.). In fact most of ch. 4 is a précis of
long tracts of Aristotle’s Meteorology I–III.
21 The author points out that the same impulse can lead to different outcomes, depending on the constitu-
tion of the thing being moved: “It is as if one would throw a sphere, a cube, a cone and a cylinder from
a vessel at the same time—for each of them will move according to its own shape—or if one would
have a water animal, a land animal and a bird in the folds of one’s cloak and throw them out at the same
time.” Trans. Thom (2014: 47).
22 A clear statement of the necessity of proceeding from “what” to “the reason why” can be found e.g. in
Aristotle’s Meteor. III.2.371b18–22.
23 I would like to express my gratitude to Phyllis Illari, Stuart Glennan, Andrew Gregory, and Monte
Johnson for their helpful comments on an earlier version of this chapter.
References
Bailey, C. (1966, reprinted) Titi Lucreti Cari De rerum natura libri sex, vols. I–III, Oxford: Clarendon Press.
Berryman, S. (2002) “Galen and the Mechanical Philosophy,” in Apeiron: A Journal for Ancient Philosophy
and Science, 35, 235–53.
—— (2003) “Ancient Automata and Mechanical Explanation,” Phronesis, XLVIII, 4, 344–69.
—— (2009) The Mechanical Hypothesis in Ancient Greek Natural Philosophy, Cambridge: Cambridge
University Press.
Craver, C. and Tabery, J. (2015) “Mechanisms in Science,” in the Stanford Encyclopedia of Philosophy: plato.
stanford.edu/entries/science-mechanisms.
De Groot, J. (2014) Aristotle’s Empiricism: Experience and Mechanics in the 4th Century BC, Las Vegas:
Parmenides Publishing.
De Solla Price, D.J. (1964) “Automata and the Origins of Mechanism and Mechanistic Philosophy,”
Technology and Culture, V, 1, 9–23.
24
Mechanisms
Furley, D.J. (1955, reprinted) Aristotle, On the Cosmos, Cambridge, MA: Harvard University Press.
Graham, D.W. (2010) The Texts of Early Greek Philosophy, Part I, Cambridge: Cambridge University Press.
Glennan, S. (2013) “Mechanisms,” in Martin Curd and Stathis Psillos eds., The Routledge Companion to
Philosophy of Science, London: Routledge, 420–8.
Gregory, A. (2013) “Leucippus and Democritus on Like to Like and ou mallon,” Aperion, 46, 4, 446–68.
Heath, T. (1981 [1921]) A History of Greek Mathematics, vol. 1 and 2, New York: Dover Publications.
Henry, D. (2005) “Embryological Models in Ancient Philosophy,” Phronesis, L, 1, 1–42.
Hett, W.S. (1936) Aristotle: Minor Works, London: Harvard University Press.
Johnson, M.R. (2009) “Spontaneity, Democritean Causality and Freedom,” Elenchos, XXX, 1, 5–52.
Johnson, M.R. (forthcoming) “Aristotelian Mechanistic Explanation,” in J. Rocca ed., Teleology in the
Ancient World: Philosophical and Medical Approaches, Cambridge: Cambridge University Press.
Nussbaum, M. (1978) Aristotle’s De Motu Animalium, Princeton, NJ: Princeton University Press.
Preus, A. (1977) “Galen’s Criticism of Aristotle’s Conception Theory,” Journal of the History of Biology,
X, 1, 65–85.
Sambursky, S. (1962) The Physical World of Late Antiquity, Princeton, NJ: Princeton University Press.
Taylor, C.C.W. (1999) The Atomists: Leucippus and Democritus, Fragments, Toronto: University of Toronto
Press.
Thom, J.C. (2014) Cosmic Order and Divine Power. Pseudo-Aristotle, On the Cosmos, Tübingen: Mohr
Siebeck.
Wilson, M. (2013) Structure and Method in Aristotle’s Meteorologica: A More Disorderly Nature, Cambridge:
Cambridge University Press.
25
3
FROM THE MECHANICAL
PHILOSOPHY TO EARLY
MODERN MECHANISMS
Sophie Roux
Early modern natural philosophers put forward the ontological program that was called
“mechanical philosophy” and they gave mechanical explanations for all kinds of phenomena,
such as gravity, magnetism, the colors of the rainbow, the circulation of the blood, the motion
of the heart, and the development of animals. For a generation of historians, the mechanical
philosophy was regarded as the main alternative to Aristotelian orthodoxy during the so-called
Scientific Revolution and mechanical explanations were presented as paving the way for the use
of experiments and mathematics in the understanding of natural phenomena.
However, the historical category of mechanical philosophy was later criticized as being too
broad, while early modern mechanical explanations were condemned by more epistemologi-
cally oriented minds for being incompatible with, or at least not necessarily connected to, the
use of experiments and mathematics. In the last ten years, just as the new mechanistic literature
emerged in philosophy of science, there has been a reevaluation of early modern mechani-
cal explanations in a domain that had been until then considered peripheral to the so-called
Scientific Revolution, namely the domain of biology, anatomy, physiology, and medicine.
Although they were neither confirmed nor predictive, some early modern explanations in these
domains appear to have a cognitive value similar to the value of contemporary mechanistic
explanations.
1. Establishing mechanical philosophy

Boyle is often said to have coined the term “mechanical philosophy.” It would be more exact to
say that he was the first to use this term to advertise the general program of explaining all natural
phenomena by matter and motion alone. There were indeed some earlier uses of the term. After
he read Meteors and Dioptrics, Libert Froidmont complained in a letter from 1637 that Descartes
fell too often into the “coarse and somewhat bloated (ruda & pinguiscula)” Epicurean physics; he
noted in particular that the reduction of the Aristotelian elements to small parts of various figures
was “too gross and mechanical (nimis crassa & mechanica).” In his answer to Froidmont, Descartes
allowed his philosophy to be qualified as “mechanical,” but he reversed the negative connota-
tions associated with this adjective and insisted that his physics manifested a certitude similar to
the certitude of mathematics only in as much as it was mechanical. Froidmont had only to see the
numerous problems that Descartes was able to solve to understand that there was absolutely no
26
Mechanical philosophy to early modern mechanisms
reason to condemn his “bloated and mechanical philosophy (pinguiscula & mechanica philosophia)”
(Descartes, 1964–74, vol. I, pp. 402, 406, 420–1, discussed in Roux, 1996, pp. 15–17, Gabbey,
2004, pp. 18–20, and Roux, 2004, pp. 32–4).
Just a few years after Descartes’ death, Henry More dubbed him “the great Master of this
Mechanical Hypothesis” (1653a, p. 44). Later on, More recommended “that admirable Master
of Mechaniks Des-Cartes” to be taught in universities to make the students see the limits of
mechanical explanations and to enable them to beat hollow the “pretender to Mechanick
Philosophy” (1659, n.p., discussed in Gabbey, 1982, pp. 220–2). These occurrences of the term
“mechanical philosophy” are interesting, but they should not be equated with the establishment
of mechanical philosophy. Descartes had a positive program for reducing all natural phenomena
to matter and motion alone (Descartes, 1964–74, vol. VIII, pp. 314–23, vol. XI, p. 47, passim),
but he was not ready to include in this program those who had other conceptions of matter and
motion than his own. As for More, he did not consider mechanical philosophy as a program
that would have delineated how to frame physical explanations in the years to come; for him,
Descartes’ explanations were essentially incomplete, and their gaps gave evidence that some
kind of immaterial substance should be introduced, whether it be God or the Spirit of Nature
(1653b, pp. 42–7 and 1659, pp. 193–204, discussed in Gabbey 1990b, pp. 22–5 and 30–1).
The story is different in Some specimens of an attempt to make chymical experiments useful to
illustrate the notions of the corpuscular philosophy that Robert Boyle wrote in the late 1650s and
published in his Certain Physiological Essays (1661). In the Preface, Boyle insisted that, not-
withstanding the differences between their philosophies, Descartes and Gassendi agreed upon
two things. First, contrary to Aristotelians and to chemists, they intended to “deduc[e] all the
Phenomena of Nature from Matter and Local Motion.” Second, they proposed to stand up
for the defense of Christian religion. At this point, Boyle needed a term to baptize the pro-
gram that both Descartes and Gassendi would have shared. He actually proposed several ones:
it could be called “Corpuscular,” because it explains natural phenomena through corpuscles,
“Phoenician,” because it its inventor is supposed to be the Phoenician Moschus, or, still,
“Mechanical,” because it gives an account of phenomena by motion, which is “obvious and
very powerfull in Mechanical Engines” (1999–2000, vol. II, p. 87, discussed in Roux, 1996,
pp. 18–20). This ontological program was an important program, even if it was only one of
Boyle’s two programs for natural philosophy, the other one being “experimental philoso-
phy”; that is, the epistemological program to acquire and secure knowledge of nature through
observations and experiments (1999–2000, vol. II, p. 14, vol. III, p. 12, passim; see Sargent,
1994, 1995; Chalmers, 1993, 2012; Gaukroger, 2006, pp. 352–99; Anstey and Vanzo, 2012,
2016; Anstey, 2014). Boyle was not sparing his words; to designate this ontological program,
he spoke also of the “New,” the “Real,” or the “Atomical” philosophy (1999–2000, vol. XI,
p. 292). But his most common terms are “the corpuscular (or corpuscularian) philosophy (or
hypothesis)” and “the mechanical philosophy (or hypothesis),” the first emphasizing the kind
of entities that appear in explanations (corpuscles), the second the kind of activities that these
entities are engaged in (motions).
In The Origin of Forms and Qualities according to the Corpuscular Philosophy (1666), publicized
by Henry Oldenburg in the Philosophical Transactions as “a kind of Introduction to the Principles
of the Mechanical Philosophy” (1666, p. 191), Boyle identified positively the eight tenets that are
constitutive of this ontological program:
1) There is only one universal matter, which is extended, divisible and impenetrable.
2) The diversity between bodies comes from various affections of matter, the main of these
affections being local motion.
27
Sophie Roux
3) Motion divides matter into parts of various sizes and figures.

4) Between several parts of matter, there are relations of order and situation, which determine
the texture of a body.
5) Parts of matter in motion make some impressions on the senses of animals.
6) The sensible qualities that are associated to these impressions are nothing but the effects of
matter in motion on the senses.
7) Substantial forms are nothing but the names given to aggregates of sensible qualities.
8) Generation, alteration and corruption are nothing but the names given to transformations
of matter in motion. (1999–2000, vol. V, pp. 305 sqq.)
Referring to those that Boyle himself designated as the two founding fathers of mechanical
philosophy, to wit Descartes and Gassendi, will make clear that the mechanical philosophy was
large enough to include various interpretations of most of these tenets.
Regarding 1, contrary to Aristotelian matter, which is a pure power, matter is a true
substance (Descartes, 1964–74, vol. XI, p. 33; Gassendi, 1658, vol. III, p. 636b). Being
“universal,” matter is everywhere one and the same: it is a “homogeneous” or “uniform”
substance (Descartes, 1964–74, vol. III, pp. 211–12, vol. VIII, p. 52, vol. XI, p. 17). Its
only properties are extension, divisibility, and impenetrability. That bodies are extended was
granted by all natural philosophers; the only question was if extension was the only essential
property of matter. By mentioning “divisibility,” Boyle seems to adhere to Descartes’ view,
according to which there are no atoms because any part of matter can be divided further away
(1964–74, vol. III, pp. 191, 213–14, 477, vol. VI, pp. 238–9, vol. VIII–1, pp. 51–2, 58–9).
Still, Descartes’ view was that matter is infinitely divisible, while Boyle claimed to remain
neutral with respect to this question (1999–2000, vol. VIII, pp. 103–4). In fact, since Boyle
admitted that there are minima or prima naturalia that are not naturally divided, though they
could be divided in thought or by God (1999–2000, vol. V, pp. 325–6), except for the words
that he uses, he subscribes to Gassendi’s view, according to which there are atoms that have
parts but that cannot be divided by any natural force (1658, vol. I, p. 256b). Similarly, put-
ting on a par extension and impenetrability (that is, the property that a body has of excluding
any other body from its location) is truer to Gassendi than to Descartes: Gassendi presented,
in addition to gravity or weight, extension and impenetrability as two equally fundamental
properties (1658, vol. I, pp. 55a, 267a), while Descartes asserted that impenetrability is to be
derived from extension (1964–74, vol. V, pp. 269, 341–2, vol. XI, p. 33).
Regarding 2 and 4, because of its homogeneity and uniformity, matter is not sufficient in
itself to account for the variety we see in natural phenomena. This variety comes from the
different shapes, sizes, and motions of parts of matter and from their relationships (Descartes,
1964–74, vol. VIII–1, pp. 52–3, vol. XI, p. 34; Gassendi, 1658, vol. I, pp. 366a–71b).
Regarding 3, here, Boyle adheres to Descartes’ view, according to which motion is prior
to shape and size because it is responsible for slicing matter into corpuscles of various sizes and
shapes (1964–74, vol. XI, p. 34). A related point is that, for Descartes, motion does not belong
to matter in itself, while, for Gassendi, matter is not inert but active, weight being a principle of
motion internal to atoms as such (1658, vol. I, pp. 276b, 335b). In that respect, Boyle goes along
with Descartes: against Gassendi and the Ancient atomists, he opposes the opinion that motion
is essential to matter, because it would pave the way to atheism (1999–2000, vol. V, p. 306).
Regarding 5 and 6, since there are only corpuscles in motion in the world, the various quali-
ties that animals perceive cannot but be caused by these corpuscles in motion. While Descartes
did not try to give specific names to combinations of corpuscles, Gassendi and Boyle sug-
gested that there are intermediate levels between the lowest atomic level and the highest level,
28
which are accessible through observations and experiments; they called these intermediate levels
respectively semina rerum or molecula (Gassendi, 1658, vol. I, pp. 282b, 472a, 335b) and textures
or primary concretions (Boyle, 1999–2000, vol. V, pp. 333–4).
Regarding 7 and 8, mechanical philosophy is an eliminative reductionism that aimed at
replacing the Aristotelian ontology. Considering the variety of ways in which the previous
tenets were interpreted, the rejection of Aristotelian ontology was probably the only common
denominator between early modern mechanical philosophers.
As he himself acknowledged, Boyle did not create this ontological program out of nothing:
in the first half of the seventeenth century, there had been attempts to replace the dominant
Aristotelian natural philosophy with a natural philosophy formulated in terms of matter and
motion alone. But, possibly because of his own “reconciling Disposition,” or because of a larger
Latitudinarian context, Boyle was one of the first authors to propose to put aside the differences
that were important to the members of the former generation and to offer to the members of
the generations to come to create a common program in which to inscribe their works (Garber,
2013). Thus, the resemblances between Descartes and Gassendi were brought to the forefront
and their differences were pushed into the background. These were metaphysical differences
that concerned God and the soul, most of which were violently expressed in the controversy
that went through Descartes’ Meditationes, Gassendi’s Fifth Objections, Descartes’ Fifth Replies,
Gassendi’s Disquisitio metaphysica, and finally Descartes’ letter to Clerselier of 12 January 1646
(Lennon, 1993; Lolordo, 2007; Osler, 1994; Roux, 2008). There were also differences concern-
ing the very principles of natural philosophy, as we just noted (Lennon, 1993; Roux, 2000).
More important for us here, there were what we perceive today as epistemological differ-
ences between types of explanations (Roux, 2009, 2012). Consider, for example, bodies that
fall on the Earth. This had to be explained, because gravity or attraction were considered as
mere words, action at a distance being proscribed in favor of action by contact. According to
Gassendi, the fall of heavy bodies is caused by the conjunction of two external forces, on one
hand the pushing force of the air from above, on the other the force of some magnetic corpus-
cles emitted by the Earth that pull bodies down to the Earth as small insensible ropes or hooks
would do (1658, vol. III, pp. 489–96). According to Descartes, the fall of heavy bodies follows
from the impossibility of the void and from his third law of nature, according to which every
body, when moving circularly, tends to recede from the center of its motion, a tendency which
is greater for swifter and smaller bodies. Void being impossible, the smaller and swifter bodies
would not recede from the center if the bigger and slower bodies did not approach the center,
which means that these bigger and slower bodies go down (1964–74, vol. VIII, pp. 213–17; the
explanation is slightly different in Le Monde (vol. XI, pp. 72–80), but its kingpin is the same law
of nature and the impossibility of the void).
Let us call Gassendi’s explanation “corpuscular,” in so far as the burden of explanation lies
on the properties of insensible magnetic corpuscles, and Descartes’ explanation “nomological,”
because the burden of his explanation lies on a law of nature. The distinction between cor-
puscular and nomological explanations does not exhaust the different types of mechanical
explanations, though. There were also explanations that consisted in exhibiting the causal
interactions of parts of more or less complex structures. This is the type of explanation that was
put in place when animals were compared with machines that can be decomposed in parts that
act on each other. For example, Perrault suggests that the influx of animal spirits into a mus-
cle causes its release, which itself causes the contraction of the antagonistic muscle, just as the
release of the guy supporting a mast causes the slackening of the opposed guy (Perrault, 1680,
pp. 75–7, discussed in Des Chene, 2005; Roux, 2012). Considering the importance that com-
parison with machines had for this third type of explanation, it could be called “machinical.”
29
Sophie Roux
To avoid neologisms, but also to comply with the distinction proposed by Ernan McMullin
(1978, discussed below), let us call these “structural” explanations.
Partly because of Boyle’s position at the Royal Society and partly because the ontological
program of explaining natural phenomena in terms of matter and motion alone was at this time
already well admitted by a number of natural philosophers, the terms “corpuscular philoso-
phy” and “mechanical philosophy” were successful, but not always with the same connotations.
Among the members of the newly founded Royal Society, the adjectives “mechanical” and
“experimental” were paired together, as if each of them contributed to make the meaning of
the other more precise. Henry Power presented himself as an “experimental and mechanical
philosopher” (1664, p. 193); Robert Hooke contrasted “the real, the mechanical, the experimental
Philosophy” and “the Philosophy of discourse and disputations” (1665, Sig.a2r); Samuel Parker
explained that he preferred “the Mechanical and Experimental Philosophy” to the Aristotelian
philosophy (1666, p. 45). In contrast, on the Continent, even those who thought that every-
thing occurs mechanically did not often use the term “mechanical philosophy”; when they did,
it referred to the demand that physical explanations are formulated in intelligible terms and con-
nected to the mathematical sciences.
It was only by reference to Boyle and the Royal Society that in this period the young Leibniz
happened to speak of “corpuscularians” and of “corpuscular philosophy” (1923, vol. VI–1,
pp. 489–90, VI–2, pp. 325, 327), a term that he used at least once as equivalent with “mechani-
cal philosophy” (1923, VI–2, p. 325). Still, he already called “mechanical” what does not rely on
the supposition of fictitious entities, but on clear and simple terms (1923, vol. II–1, pp. 266, 284,
287, 372, 379, 393). From the end of the 1670s onward, Leibniz continued to occasionally use
“corpuscular philosophy” (1923, vol. II–2, pp. 396, 845, VI–4, p. 477), but the more frequent
terms from his pen were “mechanical philosophy” and related words like “mechanism,” that
he used in association with laws and mathematics (1923, vol. II–2, p. 172, VI–4, pp. 485, 1559,
1566, 2009, 2118, 2342). The association of the mechanical with what is clear and intelligible,
if not with mathematics, is also pregnant in the eulogies that Bernard Le Bovier de Fontenelle
wrote as secretary of the Académie des sciences. Speaking of a chemist, he rejoiced that “the
sound philosophy . . . undertook to reduce this mysterious chemistry to the simple corpuscular
mechanics” and he defined the corpuscular philosophy as “the one where only clear ideas are
admitted, that is figures and motions” (1740, vol. II, pp. 444, 250, 322).
Thus, because of the variety of these connotations, “mechanical philosophy” changed from
a term for an already large ontological program into a kind of a broad umbrella term covering
several programs associated with the erosion of Aristotelian natural philosophy and with the
development of the new sciences (Roux, 1996). The first historians who promoted mechanical
philosophy as a key component of the Scientific Revolution took over this broad umbrella. In
her influential paper from 1952, Marie Boas identified, among many others, Descartes, Boyle,
and Newton as the three main early modern mechanical philosophers who were able to “put
atom to work” (1952, p. 540). She suggested a progression from each to the next one: Descartes
would have offered a theory of matter accounting for many unexplained physical properties,
Boyle would have expanded it thanks to his brilliant experiments, Newton would have put
the emphasis on the mathematized law of attraction and, if he “inclined towards the non-
mechanical one for want of experimental evidence,” “he would have preferred a mechanical
explanation” (1952, p. 519). For her, the early modern mechanical philosophy in its full form
consisted in adapting the ancient atomism to the new science of mechanics, characterized by the
use of experiments and mathematics (1952, pp. 414, 426, 520–1).
Another example of these first promoters of the historiographical category of the mechani-
cal philosophy is Eduard Jan Dijksterhuis. According to the opening words of his monumental
30
Mechanization of the world picture, first published in Dutch in 1950 and then translated into
several European languages, the emergence of “the conception of the world usually called
mechanical or mechanistic” had more profound and far-reaching effects than any other con-
ception of the world (1961, p. 3). In a word, these historians described the ontology of matter
and motion as necessarily implying a certain access to nature (that is, through observations
and experiments) as well as a certain development of knowledge (thanks to mathematics).
For them, the mechanical philosophy was thus essentially linked to the two most eminent
characteristics of the Scientific Revolution and, beyond that, of modern science; that is,
experimentation and mathematization.
2. Challenging mechanical philosophy

Still, questions about the delineation of mechanical philosophy emerged and the cluster of
associated programs began to come apart. Of course, some of these programs were sometimes, in
some places, and in some respects associated, but a natural philosopher engaged in one of them
did not necessarily engage in all the others. Already in The Mechanization of the world picture,
Dijksterhuis made a distinction between mechanics as the science of motion and mechanics as
the theory of machines; in his conclusion, he pointed out that what led to the mechanization
of the world picture was not the theory of machines, but rather a science of motion that should
be characterized as a “mathematism” rather than as a “mechanism” (1961, pp. 4, 498). In his
short textbook, The Construction of Modern Science, Mechanisms and Mechanics, which has achieved
an enduring success since its first publication in 1971, Richard Westfall dissociated two tradi-
tions that had been conflated by Boas: the mechanical philosophy, which “conceived of nature
as a huge machine and sought to explain the hidden mechanisms behind phenomena” and the
Platonic-Pythagorean tradition, which developed in the mathematized science of mechanics.
According to him, the tension between these two conflicting traditions was resolved only with
Newton, which is to grant some significance to mechanical philosophy (1977, pp. 1, 36, 42,
120). But when Westfall comes to the point of explaining what exactly the mechanical phi-
losophy brought to sciences such as optics, chemistry, and biology, he presents it as a language,
an idiom, a facade, a robe, or even a puppet regime (1977, pp. 41–2, 50, 56, 73, 77, 94, 104).
This amounts to saying that mechanical philosophy was something external to the sciences.
Introducing Gassendi, Westfall went as far as speaking of “the occupational vice of mechanical
philosophers, the imaginary construction of invisible mechanisms to account for phenomena”
(1977, p. 41, see also pp. 1, 56).
It was only, however, in the 1990s that the category of mechanical philosophy was systemati-
cally debunked from a historical point of view. It was pointed out that to regard the mechanical
philosophy as the principal alternative to Aristotelian orthodoxy was a significant oversimplifi-
cation of the historical situation. Learned studies emphasized that the great authors owed more
to the Scholastics than they had been willing to admit, and for the lesser authors, that they had
at times taken such complex positions that the division between the mechanical and the non-
mechanical philosophies became impossible to draw in practice, all the more so given that the
Aristotelian doctrine of minima naturalia helped the development of some corpuscularianisms
(Leijenhorst, 2002; Lüthy, 1997, 2001a, 2001b; Murdoch, 2001). It had been known for twenty
years that various magical and hermetic philosophies were available at the time (Rhigini Bonelli
and Shea, 1975; Westman and McGuire, 1976), but only in the 1990s were the alchemical roots
and active principles in some early modern corpuscularianisms recognized (Clericuzio, 1990,
2000; Henry, 1986, 1989; Newman, 1994, 2001, 2006; Principe, 1998). In addition, it was
pointed out that the first generation of the so-called mechanical philosophers did not identify
31
Sophie Roux
themselves as belonging to a common program, but rather as competitors in the search for new
philosophies intended to replace the philosophy of the schools (Gabbey, 2004; Garber, 2013).
In short, the category of mechanical philosophy came to be thought of as lacking any real
historical significance, except when applied to authors like Boyle who explicitly defined it. It
was particularly feared that it represented a bad partition of seventeenth-century natural phi-
losophies, ill-suited for drawing out their inexhaustible richness and their fast transformation.
While I concur with the historical criticisms that have been addressed to the obviously too
broad category of mechanical philosophy, in this chapter I want to focus on the epistemological
grounds that have been used to scrutinize mechanical explanations. For the sake of clarity, I will
distinguish a de jure evaluation that bears on mechanical explanations in general from a de facto
evaluation that affects some mechanical explanations only.
The de jure evaluation was based on a normative notion of science, according to which some-
thing deserves to be positively evaluated if and only if it complies with certain well-established
standards; that is, in the case of science, the use of mathematics and the recourse to experimental
proofs. For example, Gaston Bachelard famously said that Descartes’ physics “should be left
in its historical solitude,” because it would be a physics “where objects are not measured, a
physics without equations, a geometrical representation without scale, without mathematics”
(1951, p. 35). Similarly, with regards to the recourse to experimental proofs, Alan Chalmers
argued that Boyle’s successes were “achieved in spite, rather than because, of his allegiance to
mechanical philosophy” (1993, p. 541; and again, 2012). Contrary to the first historians who
promoted the category of mechanical philosophy, Bachelard and Chalmers described the link
between mechanical explanations on the one hand and mathematization or experimentation on
the other as a kind of union against nature: as such, mechanical explanations would stay outside
of proper science—they would even constitute an obstacle to the achievement of scientific
norms. If mechanical philosophy was saved, it was not for its explanatory successes, but only
because it functioned as a unifying discourse that legitimized scientific practices (Gaukroger,
2006, pp. 253–5, 260, 397–406).
There were also de facto criticisms addressed to mechanical explanations. In this case, the
objector did not say that they were flawed in principle and that they were doomed to be
false, but in a more descriptive way noted that some of them happened to be vacuous. Ernan
McMullin characterized structural explanations as obtaining when the properties or behavior
of a complex entity are explained by the structure of that entity; that is, by “a set of constitu-
ent entities or processes and to the relationships between them.” Moreover, he noted that,
by contrast with nomological explanations, such explanations are causal, the explanans being a
structure that causes the explanandum (McMullin, 1978, pp. 139, 145–7). Relying on the notion
of structural explanation, Alan Gabbey proposed to consider mechanical explanations as a kind
of structural explanations. (This is obviously a restriction; if mechanical explanations are defined
as those that are formulated in terms of matter and motion alone, they also include corpuscular
and nomological explanations, as was noted in the first part of this chapter.) This led him to
formulate a nuanced judgment on mechanical explanations.
Some of them were indeed successful when the phenomenon at stake was represented by
a working model which instantiated a physical law or property or still a more familiar reality
(the distinction between the two cases is presented as one of degree in Gabbey, 2001, p. 454).
It may happen that the working model at stake leads to a falsification of the explanation—
for example, Newton falsified the explanation that Descartes proposed of the motion of the
planets around the Sun by comparison with the motion of small bodies in a basin of swirling
water—but this, being only the usual game of science, does not mean that the explana-
tion was flawed in principle (Gabbey, 1985, pp. 11–12; 1990a, pp. 274–86; 2001, p. 453).
32
But, continues Gabbey, not all mechanical explanations were successful in this sense: most
notably, the explanations that pretended to account for sensations in terms of “adequate” or
“appropriate” corpuscles and motions failed because they turned out to be tautological or
circular. For example, Walter Charleton writes that the “Odours . . . can never be explained,
but by assuming a certain Commensuration, or Correspondency, betwixt the Particles amassing
the Odour, and the Contexture of the Olfactory Nerves” (quoted and discussed in Gabbey,
1985, pp. 12–14, 1990a, pp. 278–82, 2001, pp. 461–2; on the “incommensurability” between
a sensation and its mechanical explanation, see Meyerson, 1951, pp. 334–7).
When they proposed such explanations, mechanical philosophers did not progress one step
beyond Molière’s doctor. When this doctor said that opium causes sleep because it has a
dormitive virtue, he referred to an empirical property, since opium has indeed the property
to make one’s sleep, but redoubles the explanandum—this empirical property—by a vacu-
ous explanans—a dormitive virtue; asserting, as Cartesians did, that opium has a corpuscular
structure that is appropriate to make one sleep is no better, because the explanans—the “appro-
priate” corpuscular structure of opium—only redoubles the explanandum—the fact that one
sleeps when we swallow opium (Gabbey, 2001, p. 462). As to the reason why mechani-
cal philosophers proposed such tautological explanations, it was because of their situation of
competition with the Aristotelians: Aristotle’s worldview, to which they intended to offer
a plausible alternative, continued to shape their agenda (Gabbey, 1985, pp. 13–14, 1990a,
p. 279; see also Clarke, 1989, p. 189).
Although Gabbey had a nuanced judgment concerning the value of mechanical explana-
tions, his papers were perceived as reinforcing the de jure criticisms of mechanical explanations.
When these were studied at all, they were not considered as an epistemologically legitimate
practice, but as an outdated phenomenon to be studied from a contextualized point of view—
much like the remnants of a lost civilization, the meaning of which we cannot perceive. This
contextualized treatment of mechanical explanations was all the more striking given that, at
about the same time, historians of science claimed that practices that had been until that point
neglected and even rejected as irrational, for example alchemical practices, had a rationality
of their own (Principe, 1998; Newman, 2001, 2006). In the last ten years, however, while
the new mechanism literature emerged in philosophy of science, there was a reevaluation
of early modern mechanical explanations in a domain that had been until then considered
peripheral, namely the domain of biology and biology-related disciplines. Comparing early
modern explanations with the new mechanism literature will bring to light a number of points
of epistemological interest.
3. Reevaluating mechanical explanations

As we have seen, the de jure negative evaluation of mechanical explanations was grounded on
a specific picture of science, according to which the science par excellence would be a physics
relying on mathematical laws; not surprisingly, when McMullin and Gabbey began to evaluate
them more positively, it was through the notion of structural explanation, which was intro-
duced by contrast to the notion of nomological explanation. Similarly, if one sets aside some
forerunners (Haugeland, 1978; Salmon, 1984; Glennan, 1996, 2002), the mechanistic literature
of the 2000s began with the observation that the standard philosophy of science, being modeled
after parts of physics that rely on mathematized laws, did not say a word about mechanisms, not-
withstanding the fact that the use of mechanisms is more than frequent in contemporary biology
and biology-related disciplines (Machamer, Darden, and Craver, 2000, pp. 7–8, 23; Bechtel and
Abrahamsen, 2005, pp. 422–3; Bechtel and Richardson, 2010, pp. xvii–xviii; Bechtel, 2011,
33
Sophie Roux
pp. 533–4, 537, 539). Such a similarity suggests that we draw a parallel between contemporary
mechanistic explanations and early modern mechanical explanations.
This parallel is all the more justified given that the word “mechanism” itself was a neologism
first coined by the very members of the Royal Society who, in the 1660s in England, began to
use the term “mechanical philosophy.” In the Immortality of the Soul, More opposes Descartes’
assertion that we may blink without the intermeddling of the soul,
for if one . . . can keep himself from the fear of any hurt . . . he may easily abstain from
winking: But if fear surprise him, the Soule is to be entitled to the action, and not the
meer Mechanisme of the Body.
(1659, p. 103)
Similarly, Power claimed that the microscope would allow “the illustrious wits of the Atomical
and Corpuscularian Philosophers” to see “the curious Mechanism and organical Contrivance
of those Minute Animals, with their distinct parts, colour, figure and motion” (1664, p. 5),
the Philosophical transactions reported that Thomas Willis’ Pharmaceutice rationalis exposes the
“mechanism and power” by which “medicaments do their works,” the “mechanical way”
being contrasted to “specifique vertues” (1673, pp. 6166–7), and Nehemiah Grew presented as
a “mechanism” a membrane that functions as a bow (1682, p. 13). In his Micrographia, Hooke
used often the word “mechanism” in association with “curious,” “stupendious,” and “excel-
lent” to designate an exquisitely framed structure or contrivance which is apt to accomplish
certain functions (1665, pp. 91, 95, 102, 134, 152, 154, 165, 170–1, 173). There are differences
between the philosophical backgrounds of More, Power, Willis, and Hooke, but under their
pen, “mechanism” referred, most notably with regards to the anatomy and physiology of plants,
animals, and human beings, to a delicate material structure that accomplishes a function without
the intervention of any internal vital or spiritual principle (Bertoloni Meli, 2011, pp. 12–16,
to whom I owe some of the preceding occurrences). Here again Leibniz can be considered
as a go-between in as much as he introduced the word on the Continent (1923, vol. II–1,
p. 713, II–3, pp. 452, 713). However, in French “la mécanique” was for a long time preferred
to “le mécanisme” for designating, especially in the case of animals, a structure that accomplishes
a function (Fontenelle, 1740, vol. I, pp. 175, 249, vol. II, pp. xii, xxi). It was only in Stahl’s
Disquisitio de mechanismi et organismi diversitate and in his ensuing controversy with Leibniz that
mechanism began to be opposed to organism (1737).
Before drawing a parallel between early modern mechanical explanations and contempo-
rary mechanistic explanations, a few caveats are necessary. First, it should not cover up the
differing intellectual contexts in which the two enterprises were formulated (Theurer, 2013,
pp. 913–14). The mechanical philosophy was a reaction against Aristotelian natural philosophy
and various Renaissance natural philosophies drawing a sharp distinction between artificial and
natural beings; against them, the mechanical philosophy insisted on the homogeneity of nature
and on the universality of its laws. The contemporary mechanistic literature is directed against
the 1960s philosophers of science who prioritized physics over the other sciences and conse-
quently put the stress on a nomological account of the sciences; against them, this mechanistic
literature insisted that one should take into account mechanisms to capture the specificity of
biology with regard to the other sciences.
Second, although the contemporary mechanistic literature includes sometimes some ref-
erences to the early modern period (Glennan, 1996, pp. 50–1; Bechtel, 2011, pp. 535–6;
Nicholson, 2012; Theurer, 2013) and, conversely, although some historians of early modern
natural philosophy were involved in this literature (Des Chene, 2005; Hutchins, 2015), the
34
development of a mechanistic literature in the philosophy of science and the reevaluation of

early modern mechanical explanations in the history of science were largely independent.
Third, it is out of the question to dwell on the details of the proliferating contemporary
mechanistic literature. It is all the more out of the question given that this literature is not com-
pletely unanimous on what is a mechanism, since it includes two distinct traditions, the first
one focusing on mechanical systems, the second one on causal processes (for other attempts of
disambiguation of the notion of mechanism, see Glennan, 2002; Nicholson, 2012; Theurer,
2013). To overcome this problem, I will focus primarily on an especially relevant paper of the
first tradition (Bechtel and Abrahamsen, 2005).
Bechtel and Abrahamsen give the following definition of mechanism:
A mechanism is a structure performing a function in virtue of its component parts,

component operations, and their organization. Moreover:
• The component parts of the mechanism are those that figure in producing a
phenomenon of interest.
• Each component operation involves at least one component part.
• Operations can be organized simply by temporal sequence.
• Mechanisms may involve multiple levels of organization.
(2005, p. 423)
It is a definition that immediately makes sense if one thinks about mechanical devices like
clocks and automata that were constructed from the medieval period to early modern times. It
is indeed by reference to machines that Bechtel and Richardson first introduced the notion of
mechanistic explanation, albeit noting that this notion was extended when the technological
context evolved (Bechtel and Richardson, 2010, pp. 17–18; Bechtel, 2011, pp. 534–6; but note
that Craver and Darden, 2013, pp. 15–16 suggest we should distinguish machines and mecha-
nisms). Thus it is important to ask the question of what stays of the initial inspiration when the
notion is extended. While Machamer, Darden, and Craver answer this question by associating
mechanism with entities and activities in general (2000, pp. 3–4), Bechtel and Abrahamsen
say, not incompatibly but more precisely, that a mechanism is a system composed of parts that
interact to perform functions.
Discovering a mechanism is to bring to light various relations thanks to different cognitive
strategies. First, a strategy of decomposition that reveals the mereological relation between the
parts and the whole structure, a relation which is relative, since something which is a part with
respect to the whole structure can be considered as a whole with respect to its component
parts. Second, there is a strategy of localization, since these parts are arranged in specific spatial
relationships in order to interact, while the operations they perform follow successive tempo-
ral orderings. Finally, there are relations of causality between parts and operations as well as
between structure and function: parts produce operations, while the whole structure performs
a function. Discovering a mechanism implies decomposing a structure in parts, identifying their
localizations and interactions, and describing them as causes of the phenomenon at stake.
Since mechanisms are said to be composed of localized parts that perform operations in
time, it is only natural to represent them in diagrams that make use of spatial relations to con-
vey information both on the situation of parts in space and on the succession of operations in
time. Bechtel and Abrahamsen notice that, considering what our cognitive faculties are, the
information that is to be found in diagrams representing mechanisms could not be conveyed
verbally, or at least could not be conveyed so easily, inference procedures being different
35
Sophie Roux
whether one starts with a proposition or with a diagram. Moreover, they add that, to reason
about diagrams, the cognitive agent probably refers to former perceptions to mentally animate
the diagram and imagine how a part may act on another part. When looking at the different
parts of the diagram and following the temporal order, usually indicated by arrows, she will
be able to simulate the sequences of operations—provided, of course, that the mechanism at
stake is simple enough (Bechtel and Abrahamsen, 2005, pp. 426–31; Bechtel and Richardson,
2010, xix; Chapter 18, this volume; Machamer, Darden, and Craver, 2000, pp. 8–11 observe
that mechanisms are represented in diagrams as well, but they cannot easily justify it because of
their larger notion of mechanism).
When confronted with such a notion of mechanism, the historian of early modern science
could react negatively and say that, when the new mechanists refer to the early modern period,
they do it crudely. But the contemporary notion of mechanism and its correlation with diagrams
may nonetheless illuminate the epistemic strategies at stake in some early modern mechanical
explanations. Considering that the contemporary notion of mechanism appeared in the phi-
losophy of biology, the obvious move is to turn to biological writings; and considering that
Descartes played the role of the mechanical philosopher par excellence, it is natural to focus in par-
ticular on Descartes’ biological writings. While the mechanical philosophy has been for a long
time censured as irrelevant to biology (Westfall, 1977, pp. 94–104), it is indeed in this domain
that historians of early modern thought began to see mechanical explanations with a new eye.
Spirits and Clocks, the book that Dennis Des Chene devoted to Descartes’ biology, was a
starting point in this respect. Des Chene focused on machines as they were represented in the
so-called Theaters of machines, huge books displaying pictures of machines that the early mod-
ern engineers boasted they were able to build. It has been often noted since Jurgis Baltrusaïtis
and Geneviève Rodis-Lewis that Descartes may have referred to these books, more especially
to Salomon de Caus’ Des raisons des forces mouvantes, when, at the beginning of his treatise On
Man, he alluded to the marvelous machines that were found in the Gardens of the Greats
(De Caus, 1615; Descartes, 1964–74, vol. XI, pp. 130–2; Baltrusaitis, 1955; Rodis-Lewis,
1956). And it has been consequently argued that he may have been inspired by these books to
think that animals are only more complex machines than those machines that we construct.
Des Chene did much more than pick up on some comparisons between animal bodies
and machines: he showed that Descartes explained and represented organisms as the engineers
explained and represented machines in Theaters of machines. De Caus and other engineers
considered machines as spatial structures that can be decomposed into parts, each part producing
an operation and all the operations being coordinated to perform a function. According to Des
Chene, Descartes explained the human body in the very same way: he isolated parts (the heart,
the circulatory system, the lungs), analyzed the effect produced by operations of a localized part
(the heart) on the neighboring parts (the circulatory system and the lungs), and explained the
functions that this succession of operations composes for the organism as a whole (circulation of
the blood, respiration). Des Chene insisted also that it is essential for the machines displayed in
these Theaters to be visible and to exhibit through pictures how their component mechanisms
function. Likewise, Descartes’ biological treatises not only described with words the sequences
of operations that occur in human bodies, but also made use of pictures following the same
pictorial conventions as De Caus (2001, pp. 71–89). One cannot but think that De Caus and
Descartes bring into play a notion of mechanism and a use of pictures that is similar to what
Bechtel and Abrahamsen highlighted.
To push this further, considering that the new mechanist program appeared because the
nomological and reductionist account of the sciences was inadequate in the case of biology, it
can be asked what Descartes makes of laws and corpuscular reduction. Des Chene noted that
36
Descartes’ anatomical descriptions in terms of mechanisms were independent from any quanti-
tative assessment of forces and motions (2001, pp. 83–4), but he maintained that, in principle,
the reduction to simpler mechanisms should end up in descriptions formulated in terms of laws
applied to extended things (2001, pp. 71–2, 154). In a recent paper, Barnaby Hutchins argued
further that the reductionist ontology of the mechanical philosophy does not show up when
Descartes analyzes the operations of the human body; systems of mechanisms would be the only
entities at play. Hutchins’ general argument is that, if ontological commitments do not intervene
in practice when a phenomenon is explained, it is as if they did not exist. He has a more spe-
cific argument: the explanation of the beating of the heart in the Description of the Human Body
would be both systemic and non-reductionist. Systemic because the explanation of the beating
of the heart calls for the explanation of other associated functions (circulation of the blood and
respiration), which themselves involve explanations of other bodily functions still (nutrition and
assimilation). Non-reductionist because, in the explanation of the beating of the heart, Descartes
has recourse not only to corpuscles, but also to higher-level entities (flesh, pores, fibers) and
higher-level operations (the lengthening of the heart). The lengthening of the heart is especially
important: Descartes, hypothesizing that it is because of this lengthening that some corpuscles
of blood are expulsed, does not reduce the higher level to the lower level, but, on the contrary,
makes the higher level intervene on the lower level (Hutchins, 2015).
The new literature on mechanisms and the history of early modern thought converge also
in the case of diagrams. As we have seen, Bechtel and Abrahamsen insist that there is an affinity
between mechanisms and diagrams that helps us understand what would be difficult to capture
in verbal propositions alone, and Des Chene showed that the same pictorial conventions were
implemented by De Caus for representing machines and by Descartes for representing organs.
More generally, the numerous pictures that are to be found in Descartes’ writings have recently
been taken more seriously than before. Christoph Lüthy was one of the first scholars to recog-
nize Descartes’ innovation when he introduced pictures in books of natural philosophy, given
that there were none in scholastic textbooks. However, he judged severely their function;
according to him, Descartes used pictures to bridge the gap between logical deduction and
rhetorical persuasion, or between his general program of natural philosophy and his particular
explanations: they appeared when logical demonstrations failed (Lüthy, 2006, pp. 103, 110–11).
Although Lüthy is correct that a picture does not make a demonstration, it is still worth
following the logic of the pictures and understanding their cognitive function. Not only were
the Essays densely illustrated, but Descartes took great care of their pictures. He was a poor
painter himself, but he followed the advice of his friend Constantyn Huygens who favored
woodcuts over copperplates and preferred to place each picture in front of the corresponding
text. Huygens recommended an engraver who was “rather a philosopher” and had “the under-
standing as quick as the chisel” and the choice fell on Franz Schooten the Younger (Descartes
1964–74, vol. I, pp. 589, 607, 614).
Huygens anticipated that, among the pictures of the Essays, Figure 3.1, taken from Météores,
would be the most difficult to draw (Descartes, 1964–74, vol. I, pp. 607, 614, discussed in
Zittel, 2009, pp. 209–13). The first difficulty was probably that the picture had to convey the
idea that there are corpuscles of water that have an elongated figure, although nobody has ever
seen them. The trick of this picture is to include realistic elements, like the big rock on the
right side or the cloud above it, that are so to speak able to transfer by contiguity some of their
reality to the invisible corpuscles that they stand alongside. The second difficulty was that the
idea of a transformative motion should be conveyed through a static picture, since Descartes
thought corpuscles of water are progressively transformed into vapors, and vapors into clouds.
The reader is invited to see this transformation that happens in time with the help of the letters
37
Sophie Roux
Figure 3.1 Météores, in Descartes, 1964–74, vol. VI, p. 242
that guide her gaze from the water corpuscles above (A) to the vapors in which these corpus-
cles occupy more space because they rotate on themselves (B), and finally to different species
of clouds that move at the top of the pictures (C, D, E, F, G). It goes without saying that the
existence of such corpuscles was never experimentally confirmed. Still, such a picture functions
like a diagram that has the cognitive function of helping the reader to see nature as composed
of corpuscles that are transformed by motion.
The pictures of the posthumously published treatise On Man (discussed in Wilkin, 2003; Des
Chene, 2001, pp. 74, 84–6; Zittel, 2009, pp. 306–6, 2011) will help us to make a further point.
Because Descartes left only two illustrations while his treatise continuously referred to pictures,
his executor Claude Clerselier began to look for an engraver in 1657, then in 1659 launched
a “call for pictures” that happened to stay open for almost five years until he received at last
proposals from both Louis de La Forge and Gérard von Gutschoven (Clerselier, 1659 and 1664,
in Descartes 1964–74, vol. V, p. 764, vol. XI, pp. xiii–xvii). In between Frans Schuyl had pub-
lished in 1662 a Latin translation of the treatise, with other pictures still. Schuyl’s copperplates
were much more detailed, realist, and expressive than the diagrammatic woodcuts by La Forge
and Gutschoven that were published in Clerselier’s edition.
Figure 3.2 gathers two pictures from Clerselier’s edition representing the brain while awake
and while asleep, the corresponding picture from Schuyl’s edition being Figure 3.3. Or again,
compare the representation of the pineal gland by Gutschoven (Figure 3.4) with its represen-
tation as a small organ in the middle of the brain by Schuyl (Figure 3.5), where, as Wilkins
(2003) observed, the realism went so far as representing the pineal gland on an independent
bit of paper that can be moved by the reader. Although Claus Zittel has recently contested
the pre-eminence given by all editors and commentators to Clerselier’s pictures over Schuyl’s
pictures (2009, pp. 306–46, 2011), one does not need to decide who to support to understand
that two conceptions of pictures were at stake. Clerselier, who published his edition two years
38
Figure 3.2 The brain awake (above) and asleep (below), Clerselier’s edition, in Descartes, 1964–74,
vol. XI
after Schuyl’s edition and was probably embarrassed by being so late, justified his own mode
of representation. He could not but recognize that, as far as the engraving and the printing are
concerned, Schuyl’s pictures were better than those of La Forge and Gutshoven, but he added
that they were nevertheless not so good to make things intelligible.
One should not be astonished if these figures [of Clerselier’s edition] bear no resem-
blance to nature, because the purpose was not to make a book of anatomy . . . , but
only to explain through these figures what M. Descartes advanced in his book, where
he speaks more often of things that are not to be sensed, but that had to be rendered
sensible in order to be more intelligible. But there is nothing more easy than to put
them back in nature and to conceive them as they are, after having considered them
other than as they are.
(Clerselier, 1664; this passage is not reproduced in Descartes, 1964–74, vol. XI)
39
Figure 3.3 The brain awake (above) and asleep (below), Schuyl’s edition, in Descartes, 1662, pp. 77–8
Figure 3.4 The pineal gland, Clerselier’s edition, in Descartes, 1964–74, vol. XI
Figure 3.5 The pineal gland, Schuyl’s edition, in Descartes, 1662, p. 1188
Thus, while Schuyl had recourse to the conventions used in anatomy books because he
considered that Descartes’ explanations of animal functions would be best conveyed if the
reader was first persuaded that the pictures would conform to nature, Clerselier ended
up defending more schematic diagrams because he thought that one should first make
Descartes’ explanations intelligible, which required depicting parts and operations that are
unseen. While Schuyl favored empirical adequacy, Clerselier preferred intelligibility, which
he equated to visibility.
4. Conclusion
When they turn to history, some new mechanists suggest that mechanistic thinking began
in the seventeenth century with a “restrictive ontology” and an “austere worldview” and
that this view was progressively diversified and enriched (Machamer, Craver, and Darden,
2000, p. 15; Craver and Darden, 2005, p. 234; Craver and Darden, 2013, pp. 4–5). As has
been shown, the mechanical ontology was probably more diversified in the early modern
period than is usually thought. But the most important lesson to draw from this chapter
is that this ontology should be distinguished from the search for mechanical explanations
(Des Chene, 2001, p. 14, 2005, p. 246 put forward a similar distinction between mecha-
nism as a doctrine and mechanism as a method). The mechanical ontology was crucial to
the early modern sciences because it preliminarily excluded Aristotelian and Renaissance
doctrines that seemed to be incompatible with decomposition and localization. But, when
the matter at hand was the explanation of specific phenomena, even Descartes the arch-
mechanist made the search for particular mechanisms the primary focus; and, after him,
41
Sophie Roux
natural philosophers or physicians did not hesitate to search for mechanisms without being
committed to a mechanical ontology (Des Chene, 2005; Bertoloni Meli, 2011). In that
respect, focusing on specific mechanical explanations while drawing inspiration from the
new mechanistic literature is probably a more fruitful perspective on the early modern
period than trying to spell out the inexhaustible varieties of the mechanical philosophy.
References
Primary sources
Boyle, R. (1999–2000) The Works of Robert Boyle, ed. M. Hunter and E.B. Davis, 14 vols. London:
Pickering and Chatto.
Caus, S. de (1615) Les raisons des forces mouvantes, avec diverses machines tant utiles que plaisantes. Aus quelles sont
adionts plusieur desseings de grotes et fontaines. Frankfurt: J. Norton.
Charleton, W. (1654) Physiologia Epicuro-Gassendo-Charltoniana: or a Fabrick of Science Natural, Upon the
Hypothesis of Atoms, Founded by Epicurus, Repaired by Petrus Gassendus, Augmented by Walter Charleton.
London: T. Newcomb for T. Heath.
Clerselier, C. (1659) “Préface,” Lettres de Monsieur Descartes. Tome second, ed. C. Clerselier. Paris: C. Angot.
Clerselier, C. (1664) “Préface,” L’Homme de René Descartes et la formation du foetus avec les remarques de Louys
de La Forge, ed. C. Clerselier. Paris: C. Angot.
Descartes, R. (1662) Renatus Des Cartes de homine, figuris et latinaitate donates a Florentio Schuyl, ed. F. Schuyl.
Leiden: F. MOyardus et P. Leffen.
Descartes, R. (1664) L’Homme de René Descartes et la formation du foetus avc les remarques de Louis de La Forge,
ed. Claude Clerselier. Paris: C. Angot.
Descartes, R. (1964–74) Œuvres de Descartes, ed. C. Adam and P. Tannery (1897–1913), new presentation
by B. Rochot and P. Costabel, 13 vols. Paris: Vrin.
Fontenelle, B. Le Bovier de (1740) Éloges des académiciens, avec l’Histoire de l’Académie royale des sciences, 2
vols. La Haye: I. van der Kloot.
Gassendi, P. (1658) Petri Gassendi Opera Omnia, 6 vols. Lyon: L. Anisson and J.B. Devenet.
Glanvill, J. (1671) A Praefatory Answer to Mr Henry Stubbe, the doctor of Warwick wherein the malignity, hypocrisie,
falshood of his temper, pretences, reports, and the impertinency of his arguings & quotations in his animadversions
on Plus ultra are discovered. London: J. Collins.
Grew, N. (1682) The Anatomy of Plants, with an Idea of a Philosophical History of Plants, and Several other
Lectures, Read before the Royal Society. London: W. Rawlings.
Hooke, R. (1665) Micrographia or Some physiological descriptions of minutes bodies made by magnifying glasses:
with observations and inquiries thereupon. London: J. Martyn and J. Allestry.
Leibniz, G.W. (1923) Leibniz. Sämtliche Schriften und Briefe, ed. Preussische (then Deutsche) Akademie der
Wissenschaften, Darmstadt (then Leipzig, then Berlin).
More, H. (1653a) An antidote against atheisme, or An appeal to the natural faculties of the minde of man, whether
there be not a God by Henry More. London: R. Daniel.
More, H. (1653b) An antidote against atheisme, or, An appeal to the natural faculties of the minde of man, whether
there be not a God. London: R. Daniel.
More, H. (1659) The Immortality of the Soul. London: W. Morden.
Parker, S. (1666) A Free and Impartial Censure of the Platonick Philosophy. Oxford: W. Hall for R. Davis.
Perrault, C. (1680) La Mécanique des animaux, in Essais de physique. Tome III. Paris: J.-B. Coignard.
Philosophical Transactions of the Royal Society of London. London: J. Martyn.
Poli, M. (1706) Il Trionfo degli Acidi Vendicati dalle Calunnie di molti Moderni. Roma: G. Placo.
Power, H. (1664) Experimental philosophy, in three books containing new experiments microscopical, mercurial,
magnetical: with some deductions, and probable hypotheses, raised from them, in avouchment and illustration of the
now famous atomical hypothesis. London: J. Martyn and J. Allestry.
Stahl, G.E. (1737) Theoria medica vera. Physiologiam & Pathologiam, tanquam doctrinae medicae partes vere con-
templativas, e naturae & artis veris fundamentis, intaminata ratione, & inconcussa experientia sistens. 2nd edition
(1st edition, 1708). Halle: Orphanotropheum.
42
Secondary sources
Anstey, P. (2014) “Philosophy of Experiment in Early Modern England: The Case of Bacon, Boyle and
Hooke,” Early Science and Medicine, vol. 19, no. 2, pp. 103–132.
Anstey, P., and Vanzo, A. (2012) “The Origins of Early Modern Experimental Philosophy,” Intellectual
History Review, vol. 22, no. 4, pp. 499–518.
Anstey, P., and Vanzo, A. (2016) “Early Modern Experimental Philosophy,” in Sytsma, J. (ed.)
A Companion to Experimental Philosophy, Malden: Blackwell, pp. 87–102.
Bachelard, G. (1951) L’activité rationaliste de la physique contemporaine. Paris: Presses universitaires de
France.
Baltrusaitis, J. (1955) Anamorphoses ou perspectives curieuses. Paris: O. Perrin.
Bechtel, W., and Richardson, R.C. (2010) Discovering Complexity. Decomposition and Localization as Strategies
in Scientific Research. 2nd edition (1st edition, 1993). Cambridge, MA: The MIT Press.
Bechtel, W., and Abrahamsen, A. (2005) “Explanation: A Mechanist Alternative,” Studies in History and
Philosophy of Biology and Biomedical Sciences, vol. 36, pp. 421–441.
Bechtel, W. (2011) “Mechanism and Biological Explanation,” Philosophy of Science, vol. 78, no. 4,
pp. 533–557.
Bennett, J.A. (1986) “The Mechanics’ Philosophy and the Mechanical Philosophy,” History of Science,
vol. 24, no. 1, pp. 1–28.
Bertoloni Meli, D. (2011) Mechanism, Experiment, Disease: Marcello Malpighi and Seventeenth–Century
Anatomy. Baltimore: The Johns Hopkins University Press.
Boas, M. (1952) “The Establishment of the Mechanical Philosophy,” Osiris, vol. 10, pp. 412–541.
Chalmers, A. (1993) “The Lack of Excellency of Boyle’s Mechanical Philosophy,” Studies in History and
Philosophy of Science, vol. 24, no. 4, pp. 541–564.
Chalmers, A. (2012) “Intermediate Causes and Explanations: The Key to Understanding the Scientific
Revolution,” Studies in History and Philosophy of Science, vol. 43, no. 4, pp. 551–562.
Clarke, D.M. (1989) Occult Powers and Hypotheses: Cartesian Natural Philosophy under Louis XIV. Oxford:
Oxford University Press.
Clericuzio, A. (1990) “A Redefinition of Boyle’s Chemistry and Corpuscular Philosophy,” Annals of
Science, vol. 47, no. 6, pp. 561–589.
Clericuzio, A. (2000) Elements, Principles and Corpuscles: A Study of Atomism and Chemistry in the Seventeenth
Century. New York, Dordrecht, Boston, London: Kluwer Academic Publishers.
Clerselier, C. (1659) “Préface,” in Lettres de Monsieur Descartes. Tome second, ed. Claude Clerselier. Paris:
Charles Angot.
Craver, C., and Darden, L. (2013) Thinking about Mechanisms: Discovery across the Life Sciences. Chicago:
University of Chicago Press.
Des Chene, D. (2001) Spirits and Clocks. Machine and Organism in Descartes. Ithaca, NY: Cornell University
Press.
Des Chene, D. (2005) “Mechanisms of Life in the Seventeenth Century: Borelli, Perrault, Régis,” Studies
in History and Philosophy of Biological and Biomedical Sciences, vol. 36, pp. 245–260.
Dijksterhuis, E.J. (1961) The Mechanization of the World Picture. Oxford: Oxford University Press.
Gabbey, A. (1982) “Philosophia Cartesiana Triumphata: Henry More (1646–1671),” in Lennon, T.,
Nicholas, J.M. and Davis, J.W. (eds.), Problems of Cartesianism, Kingston, Ontario: McGill-Queen
University Press, pp. 171–250.
Gabbey, A. (1985) “The Mechanical Philosophy and its Problems: Mechanical Explanations, Impenetrability,
and Perpetual Motion,” in Pitt, J.C. (ed.), Change and Progress in Modern Science, Dordrecht, Boston,
Lancaster: D. Reidel Publishing Company, pp. 9–84.
Gabbey, A. (1990a) “Explanatory Structures and Models in Descartes’ Physics,” in Belgioioso, G., Cimino,
G., Costabel, P. and Papuli, G. (eds.), Descartes: il Metodo e I Saggi, Roma: Istituto della Enciclopedia
Italiana, pp. 273–286.
Gabbey, A. (1990b) “Henry More and the Limits of Mechanism,” in Hutton, S. (ed.), Henry More
(1614–1687): Tercentenary Studies, New York: Kluwer Academic Publishers, pp. 19–35.
Gabbey, A. (2001) “Mechanical Philosophies and their Explanations,” in Lüthy, C., Murdoch, J.E.
and Newman, W.R. (eds.), Late Medieval and Early Modern Corpuscular Matter Theories, Leiden: Brill,
pp. 441–466.
43
Sophie Roux
Gabbey, A. (2004) “What was ‘Mechanical’ about ‘The Mechanical Philosophy?’,” in Palmerino, C.R.
and Thijssen, J.M.M.H. (eds.), The Reception of the Galilean Science of Motion in Seventeenth-Century
Europe, New York: Kluwer Academic Publishers, pp. 11–24.
Garber, D. (2013) “Remarks on the Pre-history of the Mechanical Philosophy,” in Garber, D. and Roux, S.
(eds.), The Mechanization of Natural Philosophy, New York: Kluwer Academic Publishers, pp. 3–26.
Gaukroger, S. (2006) The Emergence of a Scientific Culture: Science and the Shaping of Modernity, 1210–1685.
Oxford: Oxford University Press.
Glennan, S.S. (1996) “Mechanisms and the Structure of Causation,” Erkenntnis, vol. 44, pp. 49–71.
Glennan, S. (2002) “Rethinking Mechanistic Explanation,” Proceedings of the Philosophy of Science Association,
vol. 3, pp. S342–S353.
Haugeland, J. (1978) “The Nature and Probability of Cognitivism,” The Behavioral and Brain Science, vol. 2,
pp. 215–260.
Henry, J. (1986) “Occult Qualities and the Experimental Philosophy: Active Principles in Pre-Newtonian
Matter Theory,” History of Science, vol. 24, no. 66, pp. 335–381.
Henry, J. (1989) “Robert Hooke, the Incongruous Mechanist,” in Hunter, M. and Schaffer, S. (eds.),
Robert Hooke. New Studies, Woodbridge, UK: Boydell, pp. 149–180.
Hutchins, B.R. (2015) “Descartes, Corpuscles and Reductionism: Mechanism and Systems in Descartes’
Physiology,” The Philosophical Quarterly, vol. 65, no. 261, pp. 669–689.
Leijenhorst, C. (2002) The Mechanisation of Aristotelianism: The Late Aristotelian Setting of Thomas Hobbes’
Natural Philosophy. Leiden, Boston and Köln: Brill.
Lennon, T. (1993) The Battle of the Gods and Giants: The Legacies of Descartes and Gassendi, 1615–1755.
Princeton, NJ: Princeton University Press.
Lolordo, A. (2007) Pierre Gassendi and the Birth of Early Modern Philosophy. Cambridge: Cambridge
University Press.
Lüthy, C. (1997) “Thoughts and Circumstances of Sébastien Basson: Analysis, Micro-History, Questions,”
Early Science and Medicine, vol. 2, pp. 1–73.
Lüthy, C. (2001a) “An Aristotelian Watchdog as Avant-Garde Physicist: Julius Caesar Scaliger,” Monist,
vol. 84, no. 4, pp. 542–561.
Lüthy, C. (2001b) “David Gorlaeus’s Atomism, or: The Marriage of Protestant Metaphysics with Italian
Natural Philosophy,” in Lüthy, C., Murdoch, J.E. and Newman, W.R. (eds.), Late Medieval and Early
Modern Corpuscular Matter Theories, Leiden: Brill, pp. 245–290.
Lüthy, C. (2006) “Where Logical Necessity Becomes Visual Persuasion: Descartes’ Clear and Distinct
Illustrations,” in Maclean, I. and Kusukawa, S. (eds.), Transmitting Knowledge: Words, Images and
Instruments in Early Modern Europe, Oxford: Oxford University Press, pp. 97–133.
Osler, M.J. (1994) Divine Will and the Mechanical Philosophy: Gassendi and Descartes on Contingency and
Necessity in the Created World. Cambridge: Cambridge University Press.
Machamer, P., Darden, L., and Craver, C.F. (2000) “Thinking about Mechanisms,” Philosophy of Science,
vol. 67, pp. 1–25.
McMullin, E. (1978) “Structural Explanations,” American Philosophical Quarterly, vol. 15, no. 2, pp. 139–147.
Meyerson, É. (1951) Identité et réalité. Paris: Vrin.
Murdoch, J.E. (2001) “The Medieval and Renaissance Tradition of minima naturalia,” in Lüthy, C.,
Murdoch, J.E. and Newman, W.R. (eds.), Late Medieval and Early Modern Corpuscular Matter Theories,
Leiden: Brill, pp. 91–141.
Newman, W.R. (1994) “The Corpuscular Theory of J. B. Helmont and Its Medieval Sources,” Vivarium,
vol. 31, no. 1, pp. 161–191.
Newman, W.R. (1996) “Boyle’s Debt to Corpuscular Alchemy,” in Hunter, M. (ed.), Robert Boyle
Reconsidered, Cambridge: Cambridge University Press, pp. 107–118.
Newman, W.R. (2001) “Experimental Corpuscular Theory in Aristotelian Alchemy: From Geber to
Sennert,” in Lüthy, C., Murdoch, J.E. and Newman, W.R. (eds.), Late Medieval and Early Modern
Corpuscular Matter Theories, Leiden: Brill, pp. 291–331.
Newman, W.R. (2006) Atoms and Alchemy: Chymistry and the Experimental Origins of the Scientific Revolution.
Chicago, London: The University of Chicago Press.
Nicholson, D.J. (2012) “The Concept of Mechanism in Biology,” Studies in History and Philosophy of
Biological and Biomedical Sciences, vol. 43, pp. 152–163.
Principe, L.M. (1998) The Aspiring Adept: Robert Boyle and his Alchemical Quest. Princeton, NJ: Princeton
University Press.
44
Rhigini Bonelli, M.L. and Shea, W.R. (eds.) (1975) Reason, Experiment and Mysticism in the Scientific
Revolution. New York: Science History Publications.
Rodis-Lewis, G. (1956) “Machinerie et perspectives curieuses dans leur rapport avec le cartésianisme,”
XVIIe siècle, vol. 32, pp. 461–474.
Roux, S. (1996) La philosophie mécanique (1630–1690). Unpublished PhD, EHESS, 2 vols, p. 808.
Roux, S. (2000) “Descartes atomiste?”, in Gatto, R. and Festa, E. (eds.), Atomismo e continuo nel XVII secolo,
Napoli: Vivarium, pp. 211–274.
Roux, S. (2004) “Cartesian Mechanics,” in Palmerino, C.R. and Thijssen, J.M.M.H. (eds.), The Reception
of the Galilean Science of Motion in Europe, Dordrecht: Kluwer Academic Publishers, pp. 25–66.
Roux, S. (2008) “Les Recherches métaphysiques de Gassendi: vers une histoire naturelle de l’esprit,” in
Taussig, S. (ed.), Gassendi et la modernité, Turnhout: Brepols, pp. 105–140.
Roux, S. (2009) “À propos du colloque The Machine as Model and as Metaphor,” Revue de synthèse, vol. 30,
no. 1, pp. 165–175.
Roux, S. (2012) “Quelles machines pour quels animaux? Jacques Rohault, Claude Perrault, Giovanni
Alfonso Borelli,” in Gaillard, A., Goffi, J.-Y., Roukhomovsky, B. and Roux, S. (eds.), L’automate.
Machine, métaphore, modèle, merveille. Bordeaux: Presses universitaires de Bordeaux, pp. 69–113.
Salmon, W. (1984) Scientific Explanation and the Causal Structure of the World. Princeton, NJ: Princeton
University Press.
Sargent, R.-M. (1994) “Learning from Experience: Boyle’s Construction of an Experimental Philosophy,”
in Hunter, M. (ed.), Robert Boyle Reconsidered, Cambridge: Cambridge University Press, pp. 57–78.
Sargent, R.-M. (1995) The Diffident Naturalist: Robert Boyle and the Philosophy of Experiment. Chicago:
The University of Chicago Press.
Theurer, K. (2013) “Seventeenth-Century Mechanism: An Alternative Framework for Reductionism,”
Westfall, R.S. (1977) The Construction of Modern Science: Mechanisms and Mechanics, 2nd edition. Cambridge:
Westman, R.S., and McGuire, J.E. (eds.) (1976) Hermeticism and the Scientific Revolution. Papers read at a Clark
Library Seminar, March 9, 1974. Los Angeles, William Andrews Clark Memorial Library: University of
California at Los Angeles.
Wilkin, R. (2003) “Figuring the Dead Descartes: Claude Clerselier’s Homme de René Descartes (1664),”
Representations, vol. 83, pp. 38–66.
Zittel, C. (2009) Theatrum philosophicum: Descartes und die Rolle ästhetischer Formen in der Wissenschaft. Berlin:
Akademie-Verlag.
Zittel, C. (2011) “Conflicting Pictures: Illustrating Descartes’ Traité de l’homme,” in Dupré, S. and Lüthy,
C. (eds.), Silent Messengers: The Circulation of Material Objects of Knowledge in the Early Modern Low
Countries, Berlin, Lit-Verlag, pp. 217–260.
45
4
THE ORIGINS OF THE
REACTION MECHANISM
William Goodwin
1. Introduction
Chemists have characterized reaction mechanisms in a variety of different ways, which vary
by sub-discipline and time period. For the purposes of this chapter, however, I will consider
a chemical reaction mechanism to be “a specification, by means of a sequence of elemen-
tary chemical steps, of the detailed process by which a chemical change occurs” (Lowry and
Richardson, 1987, p. 190). These elementary chemical steps may themselves be described by
providing “a picture of the participating species at one or more crucial instants during the
course of the reaction” (Gould, 1959, p. 127). These characterizations of mechanism are drawn
from relatively contemporary texts in theoretical organic chemistry, which limits how robustly
applicable they are. However, it is appropriate, in a chapter such as this, to consider the notion
of mechanism applicable in this subfield because it was the field in which reaction mechanisms
first emerged in modern chemistry.
Before going on to highlight some of the major episodes in the development of the chemical
reaction mechanism, it will be worth considering the extent to which this notion of mechanism
fits characterizations of mechanism in the philosophical literature. Minimally, “[a] mechanism
for a phenomenon consists of entities (or parts) whose activities and interactions are organized
so as to be responsible for the phenomenon” (Introduction to this volume). In the case of the
organic reaction mechanism, the phenomenon for which a mechanism is provided is a chemical
reaction, typically characterized by a chemical equation. A chemical equation represents a reac-
tion by providing a (stoichiometrically balanced) description of the reactants and the products
of the transformation. The reactants and products are themselves represented by chemical for-
mulas of one form or another. In the organic case, these are typically structural formulas, which
model1—in varying levels of detail—the composition, bond connectivity, and spatial distribu-
tion of atoms in a molecule.
A mechanism may be provided for a chemical reaction by decomposing the original chemical
equation into parts, each of which is itself a chemical equation (see Figure 4.1). In such cases, a
sequence of chemical equations, each of which represents a step in the original transformation,
collectively add up to produce the original transformation. Some products of the intermediate
steps in such a system of chemical equations, which are represented with their own structural for-
mulas, are more or less stable chemical entities that then function as reactants for subsequent steps.
46
The origins of the reaction mechanism
O O
Overall Reaction: + NucH + LGH
R LG R Nuc
H
O +
O
First Step: + H+
R LG
R LG
H H
+ +
O O
Second Step:
+ NucH + LGH
R LG R Nuc
H
+ O
O
Third Step: + H+
R Nuc
R Nuc
Figure 4.1 The decomposition of a chemical reaction into steps. This mechanism depicts an acid-
catalyzed Addition–Elimination reaction on a carboxylic acid derivative. H+ is a catalyst,
while the other charged species in the second and third steps are intermediates. Nuc
represents a nucleophile, LG a leaving group, and R a generic group
These are the intermediates of the reaction, and though they don’t occur in the original chemical
equations, their structures often do a lot of the explanatory work of the reaction mechanism.
Additionally, some chemical species may be reactants in some earlier step of the system of equa-
tions, but products later on, thereby canceling out and again not occurring at all in the original
chemical equation. These are catalysts of the reaction, and they too may play a crucial role in the
explanations provided by the mechanism. So a mechanism of this sort provides a decomposition
of a chemical reaction into a sequence of steps that include descriptions of the intermediates and
catalysts of a chemical transformation by providing their structural formulas. The sequence of
steps and their relative rates can be crucial to the mechanism because they allow for the identifi-
cation of the bottleneck step, which determines the rate at which the overall process takes place.
Additionally, the entities (intermediates and catalysts) occurring in these elementary steps can be
subject to structural analysis to explain important features of the original chemical reaction, often
including its relative rate and product distribution.
Not all organic reaction mechanisms proceed by decomposing the original reaction into
parts, in the sense of a sequence of sub-reactions. Many chemical reactions, including, typi-
cally, the elementary reactions into which more complex reactions are decomposed, proceed
in one step, without forming any more or less stable intermediates. Still, reaction mechanisms
are sought for these reactions. The entities invoked in such mechanisms are quasi-structural
descriptions of certain crucial points in the transition between reactants and products including,
typically, the highest-energy configuration along the reaction coordinate.2 And, much as with
47
William Goodwin
the intermediates in the more complex case, the representations of the crucial points of the
transition can be subjected to structural analysis, allowing the chemist to explain important facts
about the chemical reaction. Many contemporary reaction mechanisms combine both strategies
for describing the process by which a chemical reaction occurs: first, they decompose the reac-
tion into its sub-reactions, and then they provide structural characterizations of the transition
states of at least some of those sub-reactions (most crucially the bottleneck, or rate-determining
step of the reaction system).3
Chemical reactions usually involve the breaking and/or making of bonds: certain bonds in
the reactants must be broken so that the bonds in the product can be made. Mechanisms must
therefore describe stages in this process of bond breakage and formation. As a result, mechanisms
generally presuppose or invoke some account of chemical bonding. Furthermore, to explain why
bonds break and form between particular atoms, a mechanism typically relies on some account
of inter and/or intramolecular interactions. Chemists’ understanding of both chemical bonding
and the interactions between or within molecules underwent significant changes during the same
period of time that the reaction mechanism emerged as a useful theoretical tool in organic chemis-
try. As a result, there is no simple answer to the question of what sorts of activities and interactions
are attributed to the entities invoked in chemical mechanism. Early mechanisms invoke accounts
of chemical bonding and interaction that look quite different from our modern accounts, though
they are often recognizable in retrospect as approximations to these more modern accounts.
A reaction mechanism is a model of a chemical transformation. It is a model in just about
all of the relevant philosophical senses of the term: it is a primarily non-linguistic partial rep-
resentation, intermediate between theory and observation, that allows for both the concrete
application of theory and the refinement of novel explanatory concepts.4 In the first place, it
is a staged description of a chemical transformation that allows for the theoretical apparatus of
reaction kinetics or Transition State Theory to be applied, either quantitatively or qualitatively,
to the transformation. Second, it leaves out as much chemical detail as is possible (including,
in many cases, the solvent in which the reactants are dissolved), given the phenomena that it
hopes to explain. Third, it is a depiction of the microscopic actors in a macroscopic process, so
that at best it characterizes one of the innumerable ways that such transformations actually occur
at the microscopic level. And lastly, the depictions of these microscopic actors are themselves
structural models of the chemical species that they depict. Chemists are generally quite explicit
about the provisional and instrumental character of reaction mechanisms. For instance, in the
foundational textbook of physical organic chemistry, the author claims:
A mechanism . . . is a scientific tool by which to obtain verifiable relationships between

measureable quantities; it is to be judged by its utility in correlating known facts and
predicting new ones, not by its agreement with some unknowable absolute truth.
(Hammett, 1940, p. 98)
Thus, reaction mechanisms are theoretical tools for explaining and predicting chemical phenomena,
not primarily attempts to describe in some realistic way the actual processes by which chemical
reactions take place.
2. Setting the stage for mechanisms

Reaction mechanisms, in something like their modern form, first appeared in the organic chem-
istry literature at the beginning of the twentieth century (Hammett, 1966). By the end of the
nineteenth century, organic chemistry had developed an extremely successful theory—structure
48
theory—that allowed it to individuate, classify, and rationalize much of the chemical behavior
of organic compounds. The goal of structure theory, as articulated by Aleksandr Butlerov, was
to have one structural formula for each distinct compound and then “when the general laws
governing the dependence of chemical properties on chemical structure have been determined,
this formula [would] express all of those properties” (Brock, 2000, p. 256). Structural formu-
las depicted organic compounds in terms of the “bonds” between the elements (or atoms)
that composed them. Some structural chemists interpreted chemical structure realistically, as
something like a picture or model of the arrangements of atoms in organic molecules. Others,
particularly early on, resisted this realistic interpretation, opting instead to think of chemical
structures as abstract symbols that encoded the possible transformations of chemical kinds (see
Duhem, 2002; Nye, 1993, ch.6). This ambiguity was facilitated, at least in part, by the fact that
organic chemists had no consensus account of the nature of chemical bonding; thus the very
connections depicted by a chemical structure were a black box. Gradually, perhaps because of
the emerging awareness that the newly discovered electron had a role in chemical bonding,
not only were chemical structures interpreted increasingly realistically, but organic chemists
began to aspire to provide causal accounts of the processes by which one chemical structure was
transformed into another.
In 1908, we find in an annual report on the state of organic chemistry the following claim:
The great advances made in recent years in our knowledge of the structure of organic
compounds have directed attention all the more vividly to the gaps which remain.
This is especially the case with regard to reactions. The chemical equation only rep-
resents, as a rule, the initial and final stages, the mechanism by which the result is
obtained remaining obscure, and only to be filled by the assumption of hypothetical
intermediate products, the nature of which it is often very difficult to define.
(Desch and Morgan, 1909, p. 74)
This quote clearly brings out what was at this time a quite legitimate worry about the doubly
hypothetical “intermediate products” that might be invoked to fill in the mechanisms of chemi-
cal reactions. These intermediates are hypothetical not only in the sense that they are postulated
merely to explain features of the reaction and are not measureable or observable on their own,
but also because the very terms in which they are described are conditioned upon some hypo-
thetical account of chemical bonding and interaction. The speculative character of accounts of
reaction mechanisms was tolerated because of their potential payoff. The annual report of the
following year makes it clear what that payoff was:
The application of physico-chemical methods to the study of organic compounds is

largely responsible for the increasing interest shown in the relation of physical prop-
erties to chemical constitution, and in the mechanism of reactions. Whilst the older
structural organic chemistry, aiming principally at the synthetical formation of com-
pounds and the determination of constitution, was mainly concerned with the final
products of a given reaction, more and more attention is now being devoted to the
dynamical aspects of reactions, embracing the quantitative study of velocity, etc., as
well as the molecular mechanism by which the interchange is effected.
(Desch and Lapworth, 1910, p. 57)
The recently developed methods of what would later be called physical organic chemistry had
provided new kinds of information about both organic molecules and chemical transformations.
49
William Goodwin
It was in an effort to explain these newly recognized features of chemical reactions that organic
chemists early in the twentieth century resorted to speculation about the mechanisms of reac-
tions. As suggested in this quote, it was the rates of chemical reactions that primarily motivated
speculations about mechanisms.
The most fundamental result from physical chemistry that is relevant to the rates of organic
reactions is what is sometimes called the generalized Law of Mass Action:
In dilute systems the rate of every chemical reaction is proportional to the product of
the concentrations of the substances which are reacting, and is independent of the con-
centrations of all other substances and the presence or absence of all other reactions.
(Hammett, 1966, p. 465)
Physical chemists accumulated evidence for this principle over the course of the 1890s so that,
by the first decade of the twentieth century, “the generalized Law of Mass Action [was] accepted
without reservation or hesitation and [was] used with complete confidence as the basis for judg-
ments about . . . mechanism” (Hammett, 1966, p. 465). To see how this law could be used to
underwrite speculations about reaction mechanisms, I will describe one of the first studies on
organic reaction mechanisms that helped to establish “the whole technique of the kinetic inves-
tigation of reaction mechanism on a firm basis” (Hammett, 1966, p. 466).
3. Lapworth’s kinetic studies of reaction mechanism

Hydrogen cyanide (HCN) will add to the carbon-oxygen double bond in many aldehydes
(RCOH) and ketones (RCOR') to form a cyanohydrin (R2C(OH)(CN)). In 1903, Lapworth
published an account of the mechanism of this addition reaction supported by kinetic studies.
The great innovation in his account was to work backwards from an apparent violation of the
generalized Law of Mass Action to a multistep mechanism for the reaction that reconciled the
observed rate behavior with this general principle.
The stoichiometrically balanced equation for cyanohydrin formation is as follows:
R2CO + H+ + CN-  R2C(OH)CN
And according to the generalized Law of Mass Action, this would mean that the rate of reac-
tion should be proportional to the products of the concentrations of the reactants. But this
was not what Lapworth had found. Instead he reported that adding acid (H+) to the reaction
reduced its rate, contrary to the increase expected by the general principle. Adding bases, which
don’t occur among the reactants at all, actually increased its rate. Furthermore, adding cyanide
ions (CN-) to the reaction mixture led to dramatic increases in rate. Since hydrogen cyanide
is a weak acid, addition of acid would reduce the amount of HCN that dissociated to form
cyanide ions, while addition of base would promote cyanide ion formation. The upshot was
that “the reaction velocity depended mainly on the concentration of cyanogen ions [CN-]”
(Lapworth, 1903, p. 997).
To accommodate this rate data, Lapworth proposed that the reaction actually took place
in two steps. First the cyanide ion united, in a reversible process, with the neutral carbonyl
carbon, which was polarized because the oxygen possessed “residual affinity.” The result was
a complex ion, which then went on to react with hydrogen ions (H+) to form the cyanohy-
drin (see Figure 4.2). If it were supposed, in addition, that the second step was “very rapid in
comparison with the first” (Lapworth, 1903, p. 1001)—that is, that the first step was what is
50
O O−
Slow Step: + CN− R1 R2
R1 R2
CN
O− OH
Fast Step: R1 R2 + H+ R1 R2
CN CN
Figure 4.2 Lapworth’s decomposition of cyanohydrin formation. In the first, slow step, a cyanide ion
combines with a ketone to form a complex ion. In the second, faster step, the complex
ion combines with a hydrogen ion to form the cyanohydrin. The overall rate should be
proportional to the cyanide ion concentration, but not to the hydrogen ion concentration
now called the “rate-determining step”—then the prediction of the generalized Law of Mass
Action would be borne out by the rate data. When applied to the first equation, this general
principle entails that the rate should be proportional to the concentration of CN-, which is
just what the rate results had indicated.
Following up on his proposal, Lapworth tested his theory on a variety of different com-
pounds. He speculated that in certain circumstances, it should be possible to run the reaction not
with hydrogen cyanide, but instead with a cyanide salt, which would supply the CN- required
in the first step but not the hydrogen ions required in the second. In such cases it might be pos-
sible to produce and isolate the potassium salt of the complex ion that Lapworth speculated was
the product of the first step of the reaction. He reports, “[i]t was found that crystalline products
could be prepared by the direct action of potassium cyanide on benzaldehyde and camphor qui-
none, and that these products have approximately the composition of hydrated potassium salts of
the corresponding cyanohydrins” (Lapworth, 1904, p. 1207). Lapworth produced several other
kinetic studies of this general sort, each proposing a multistep mechanism proceeding though
organic ions, that reconciled measured rate data with the generalized Law of Mass Action.5
There are some important things to note about the mechanism that Lapworth provides.
First, though he proposes and defends his account of the mechanism on the basis of rate data,
he is concerned to show that the organic ion that he proposes as the product of the first step
of the reaction is a stable entity by finding a way to isolate it as a salt. Perhaps because the role
of organic ions in the transformations of organic compounds was still controversial, it was evi-
dently advantageous for the species appealed to in a mechanism to be stable and isolable. More
modern researchers, however, do “not find it necessary to isolate an intermediate in order to
establish a mechanism” (Hammett, 1966, p. 466). Lapworth, furthermore, made no attempt to
speculate about any transitory, unstable intermediate stages in the transformations of one organic
species into another. This is just to say that Lapworth’s mechanism decomposes a chemical
reaction into parts, where those parts are sequentially arranged sub-reactions, but there is no
attempt to look inside those sub-reactions themselves to describe transitory structures, such as
the transition state.
Additionally, Lapworth is clearly working with some conception of both how organic reac-
tions happen and why ions unite with organic compounds at some particular places rather
51
William Goodwin
than others. This is to say, he has theoretical views about bonding and interaction that inform
his account of mechanisms. In fact, Lapworth endorsed the general view that most changes in
organic compounds were due to “electrolytic dissociation” and thus proceeded through organic
ions of the sort proposed in this mechanism. In addition, he later developed a broader theory
that assigned “relative polarities” to the atoms in an organic molecule by first identifying a “key
atom” (what we would now identify as the most electronegative atom in the molecule) and
then “alternating polarities” of the atoms as they were further and further removed from the
key atom. Finally, he classified reagents as “anionoid” or “cationoid,” and these reagents could
be expected to attack atoms in an organic molecule of the opposite polarity. In the case of the
attack of cyanide ions on a carbonyl carbon, the cyanide ion would be an anionoid reagent and
thus directed at atoms with a positive polarity. The carbon in a carbonyl compound would have
a positive polarity because it was one atom removed from the oxygen, the key atom, with a
negative polarity.
Lapworth’s speculations about reaction mechanisms clearly show how rate data can be used
to decompose a chemical reaction into elementary reactions which themselves are consistent
with the guiding principle of reaction kinetics. He managed to do this while insisting on “an
electrochemical point of view at the molecular level” (Robinson, 1947, p. 995), which was
eventually recast and subsumed within the electronic theory of valence (the idea that covalent
chemical bonds are the result of the sharing of pairs of electrons, recognized around 1920). The
electronic theory of valence itself was eventually recast in turn within quantum mechanical
accounts of chemical bonding in the early 1930s. Still, Lapworth’s speculations about how to
think about and rationalize chemical reactions “forged a necessary link in the chain of theory”
(Robinson, 1947, p. 995) which grounds modern accounts of chemical mechanism. He is
generally credited with having helped to establish both the central role of ions in chemical
change, and the importance of polarization in understanding the interactions of chemical spe-
cies (Schofield, 1995; Russell, 1971, pp. 287–91). Oddly, however, Hammett (1966, p. 466)
reports that after Lapworth’s early work on mechanisms:
[A] remarkable thing happened; the whole development stopped almost completely
for twenty years or more. There were a few kinetic investigations, but they were
isolated, out of the main stream of chemical activity and unnoticed by the great mass
of chemists.
Indeed reaction mechanisms don’t return to center stage in the annual reports on progress in
chemistry until 1938, when the author reports (Watson, 1938):
The study of the mechanisms of chemical reactions, which has not received much
attention in recent Reports, continues to provide a fruitful field of research, and
our knowledge of many changes in which organic compounds participate has been
enriched by the results of a large number of recent investigations.
This leads to the interesting question of why the study of mechanisms was effectively
stopped for twenty years, but, following Hammett, I will not try to answer this question.
Instead I will concentrate on the sorts of developments that allowed organic chemists to
begin to look inside the elementary reactions, and thus to expand the power of the reaction
mechanism as an explanatory tool. After briefly discussing some of these developments, I
will go on to provide an example of the extended explanatory power of this second coming
of reaction mechanisms.
52
4. Moving inside the elementary reaction

As has already been suggested, one important change in the twenty years between Lapworth’s
early work and the work in the 1930s that ushered in the contemporary reaction mecha-
nism was in theories of chemical bonding and the description of the “activities” and “inter-
actions” of the species described in a chemical mechanism. Both the electronic theory of
the chemical bond and its quantum mechanical reinterpretations helped organic chemists to
coherently and consistently rationalize both the bonding and interactions of these species. For
instance, Lapworth’s “anionoid” or “cationoid” reagents came to be called nucleophiles and
electrophiles, according to a “predominating constitutional affinity either for atomic nuclei
or for electrons” (Ingold, 1953, p. 198). These reagents would then interact with the polar-
ized covalent bonds of organic substrates, which had uneven distributions of electrons ranging
over the positive atomic centers. Ingold, who was a central figure in the renewed inter-
est in reaction mechanisms, developed a system for classifying the sorts of intra and inter-
molecular interactions between organic species. This system attempted to take into account
both the electronic nature of chemical bonding and the delocalization of electrons (eventu-
ally) described by quantum mechanics. These interactions could then be used to explain not
only why a particular reaction occurred at a particular place on a molecule, but also relative
differences between the “energies” of such interactions. Something like this description of
the internal dynamics of chemical bonding and interaction had been present in Lapworth’s
accounts of the interactions between ions and polarized organic molecules, but “[i]t was only
when the picture of a chemical bond became something electrical and hence potentially fluid
that fruitful correlations were sought and found between electronic displacements affecting
bond character and modes and ease of reaction” (Bartlett, 1956, p. 10). By the time of the
second coming of the reaction mechanism, Ingold and others had “rationalized a large body
of organic experience in terms of polarizations transmitted through the molecule by essentially
two mechanisms,” basically “inductive displacement of electrons” and what would later (after
the quantum mechanical reinterpretation of the chemical bond) be called resonance (Bartlett,
1956, p. 10). Though organic chemists’ tools for thinking about bonding and interaction
have been articulated in the meantime, these are still fundamental concepts for understanding
organic reaction mechanisms.
A second crucial series of developments that allowed chemists to finally move inside the
elementary reactions was the development of theories of reaction rates. There were two distinct
theories of this sort developed in the 1930s. Both of them were attempts to model the tempera-
ture dependence of reaction rates in a way that drew upon the newly developed mathematical
machinery of chemical thermodynamics, canonically presented in (Lewis and Randall, 1923).
The Collision Theory, as (Hammett, 1940, p. 112) calls it, was principally applicable to reac-
tions in the gas phase.6 The Transition State Theory, on the other hand, was taken up by organic
chemists who are principally interested in reactions that take place in solution. The Transition
State Theory, described in (Hammett, 1940, pp. 115–18) and (Ingold, 1953, pp. 43–52), is an
extension or articulation of the thermodynamics of chemical equilibria onto the problem of
reaction rates. To appreciate its significance, it is first important to see how organic chemists
were able to use chemical thermodynamics, in conjunction with their newly refined under-
standing of chemical bonding, to understand or rationalize the effects of changes in chemical
structure or medium on a chemical equilibrium.
According to chemical thermodynamics, the eventual balance between reactants and prod-
ucts in a chemical reaction depends on their difference in a thermodynamic quantity known as
the “free” energy. Subject to some important caveats and exceptions, it is often possible in the
53
William Goodwin
sorts of circumstances relevant to the organic chemist to make predictions (typically qualitative
or relative) about how this free energy difference will change as a result of changes in either the
chemical structures of the species in equilibrium or the medium in which the equilibrium takes
place. So, for instance, by appealing to Ingold’s inductive effect, one might predict that replac-
ing one atom in a reactant structure with a more electronegative atom would result in products
that are, relatively speaking, of even lower energy. This would result in changes to the overall
free energy change in the reaction, thereby changing the eventual balance between reactants
and products in an anticipatable way. In other words, chemical thermodynamics, in concert
with the updated accounts of chemical bonding and interaction, allows for a structural analysis
of the eventual balance between reactants and products. This is a further step on the path of
fulfilling Butlerov’s ambition for structural chemistry in that features of a chemical reaction (the
eventual balance between reactants and products) can be explained in terms of the structures of
the reacting species.
Unfortunately, this sort of thermodynamics “has no concern with the rate of reaction, which
depends on reaction mechanism: it has no concern with mechanism” (Ingold, 1953, p. 40).
Only by a creative application of this thermodynamic machinery to a model of chemical transi-
tions could the sort of structural analysis facilitated by thermodynamics be used to think about
reaction rates and therefore mechanisms. The model of chemical transitions behind Transition
State Theory abstracts from the actual collisions that are a “necessary part of the process of
reaction” (Hammett, 1940, p. 116) and instead focuses on the changes in energy (in the first
instance, potential energy) that accompany an idealized reaction trajectory. The idealized reac-
tion trajectory is something like the least-energy path by which the reactants of an elementary
reaction can come together, breaking some old bonds and making new ones in the same pro-
cess, and then leave as newly formed products. If the potential energy of the reacting chemical
system is plotted versus a measure of progress along this reaction trajectory (typically called the
reaction coordinate), then there must be an energy maximum corresponding to some point in
the idealized reaction trajectory. This highest-energy intermediate configuration, which would
typically be at some midpoint in the simultaneous breaking and forming of bonds, is called the
transition state, and the energy difference between the reactants and the transition state is called
the activation energy.
The transition state is not an intermediate in the traditional sense because it is not stable
and is extremely short lived as a result; furthermore, the transition state does not have a typical
chemical structure because some of the bonds in it are partially broken while others are only
partially formed. Still, the trick of the Transition State Theory is to treat the reactants and the
transition state as if they were in chemical equilibrium. Following this strategy, the eventual bal-
ance between reactants and transition state is governed by the “free” energy difference between
them. And after some additional assumptions and approximations, the reaction rate can be taken
to be proportional directly to “the number of systems in the transition state” (Ingold, 1953,
p. 47). The free energy difference which, according to this approach, determines the rate of
the reaction can be related—again, subject to some important caveats and exceptions—to the
activation energy of the reaction.
The upshot is that the same sort of structural analysis that was possible in the case of chemical
equilibria can now be applied to questions about the rates of reactions, and thereby to mecha-
nisms, by considering how changes in structure or medium will affect the activation energies of
a class of reactions. As Ingold puts it, “we can often profitably apply the ideas of the [transition
state] theory in qualitative discussions of effects of molecular structure, or of solvent, on the heat
and entropy of activation, and therefore on rate of reaction” (Ingold, 1953, p. 48). Because of
these developments in the theory of reaction rates, therefore, the organic chemist “could treat
54
the effect of structure, of medium, and the like on rates and on equilibrium in the same terms”
(Hammett, 1966, p. 467) and thereby extend even further Butlerov’s vision of the explanatory
powers of chemical structure.
5. Explaining the Walden inversion

One of the great triumphs of the second coming of the reaction mechanism was to finally bring
clarity to a stereochemical puzzle first identified in the 1890s by Paul Walden.7 Stereochemistry
refers to the arrangement of atoms in space, and most particularly to the “handedness” of chemi-
cal compounds. When a carbon atom is bound to four distinct groups, those groups are arranged
about the carbon as if they were at the vertices of a tetrahedron with the carbon at its center. As
a result, there are two distinct such arrangements possible—these are non-super-impossible mir-
ror images of one another called enantiomers. Enantiomers have very similar chemical structures
and so their chemical and physical behavior is also very similar.
However, there are some chemical differences between them (some enzymes, for instance,
will only react with one enantiomer) and they have different optical activities; they rotate plane-
polarized light in opposite directions. Walden had shown that it was possible, by a series of
reactions, to convert an optically active substance into its enantiomer. Chemists at the time had
assumed that substitution reactions of the sort Walden employed would proceed in a way that
either kept the spatial arrangement the same, or perhaps randomized the spatial arrangement.
Walden showed that at some point during his series of reactions, the groups around carbon
had to systematically shift from one spatial arrangement to another. In the forty years after the
discovery of Walden’s inversion, many additional inversions were discovered; however, not all
substitution reactions on optically active starting products led to inversion, some resulted in the
retention of configuration, and others led to a mix of enantiomeric products.
Prior to the mechanistic approach described below, attempts to explain the Walden inver-
sion “were generally quite unsatisfactory since they were unable to predict what conditions or
reagents might produce the inversion” (Ramsay, 1981, p. 109). For our purposes, we can think
of the mechanistic explanation of the Walden inversion, worked out by about 1940 by Ingold
and Hammett and others, as consisting of three parts. First, the description and justification of
a substitution mechanism that would result in the inversion of stereochemical configuration.
Second, the description and justification of a substitution mechanism that would result in the
mixing (and possible retention) of configurations. And lastly, a scheme for deciding either which
mechanism a particular substitution was likely to undergo or, failing that, how changes to the
reaction setup were likely to influence the mechanism of a substitution.
The general form of a substitution or displacement reaction,8 which is “by far the most
important type of reaction in organic chemistry” (Hammett, 1940, p. 131), is:
A + B-C  A-B + C where A, B, and C are atoms or groups of atoms.
There are, prima facie, three ways that such a substitution could occur, corresponding to the
stages in which bonds are broken and formed. Either the bond between A and B forms first,
followed by the rupture of the bond between B and C, or the bond between B and C ruptures
first, followed by the formation of the bond between A and B, or lastly, the bond breakage and
formation might take place simultaneously. It turns out that for the kind of substitutions that are
relevant to the Walden inversion, only the second two occur. Furthermore, one of these pos-
sible mechanisms results in inversion of configuration, while the other results (most typically) in
a mixture of configurations.
55
William Goodwin
The synchronous mechanism, named SN2 by Ingold, is an elementary reaction. It takes place
in one step that proceeds through a transition state in which the bond between A and B is being
formed at the same time that the bond between B and C is breaking. If one assumes that A and
C repel one another, then (if A, B, and C all stand for single atoms) it stands to reason that the
lowest-energy path for this transformation would have A approaching B-C along the line of the
bond but on the opposite side of B. Then, as the new bond forms, C would exit along this same
line on the opposite side from A. This would result in a transition state that minimizes repulsion,
is lowest in energy, and therefore is the most likely reaction path. The case of substitution on an
asymmetric carbon9 by the SN2 mechanism is similar except that an asymmetric carbon has three
other groups to which it stays linked during the substitution process.
The behavior of these three linkages during the substitution “is like that of the ribs of an
umbrella in a gale” (Hammett, 1940, p. 159); they invert, resulting in a change of configuration
during the process of converting to the product (see Figure 4.3). That this “backside attack”
mechanism would result in the lowest-energy transition state for a simultaneous substitution
was supported by several arguments resting upon the newly available electronic and quantum
mechanical accounts of chemical bonding and interaction. Furthermore, one of the first signifi-
cant uses of isotopic labeling in chemical kinetics (see Ingold, 1953, pp. 377–9) established not
only that could such inversion occur, but also that it was universal for substitution of this sort.
The sequential mechanism, named SN1 by Ingold, proceeds in two steps. In the first step,
which is rate determining, the bond between the asymmetric carbon and the leaving group,
with some help from the solvent, breaks, resulting (typically) in an intermediate ion (much like
Lapworth’s mechanism). In the second step the new group forms a bond with this intermediate
resulting in the final product. Because the intermediate in this process is a (solvated) carbon ion
in which the carbon is bound to only three other groups (the ones that don’t change during
the substitution), this intermediate no longer has a “handedness.” The prior configuration is
lost, and when the new group adds to this intermediate it can do so in two different ways, each
resulting in a different enantiomer. Thus when a substitution occurs by this mechanism, it will
typically result not in an inversion, but rather in a mixture of the possible enantiomers.10
The last element in “solving” the Walden inversion was to develop a scheme for deciding by
which mechanism a particular reaction had proceeded or would proceed. Ingold and co-workers
showed how to distinguish whether a substitution occurred by the synchronous or the sequen-
tial route by kinetic criteria (measuring how the rate of the reaction varied with the changing
concentrations of certain reactants). They then went on to explore the structural features and
reaction conditions that promoted one or the other of these mechanistic routes. Basically, which
mechanism occurs is decided by which transition state (the one in the synchronous reaction or
the one in the rate-determining step of the sequential reaction) is lower in energy. This can be
analyzed in structural terms, considering things such as the energetic cost of breaking the initial
R1 R1
R1
Nuc + R2 LG Nuc LG Nuc R2 + LG

R2 R3
R3 R3
Figure 4.3 The “backside attack” mechanism, in which a nucleophile (Nuc) displaces a leaving group
(LG) from an asymmetric carbon (originally bonded to groups R1, R2, and R3 in addition
to LG), thereby resulting in an inversion of configuration. The structure in brackets depicts
the transition state
56
bond, the steric environment of the transition states, and the potential stabilizing effects of the
proposed solvent, etc. The upshot was that one could plan substitution reactions that should, on
structural grounds, go by one or the other of these mechanisms (typically in synthesis one wants
to employ the stereospecific SN2 mechanism because it gives a single, predictable product); and
furthermore, one could tell, based on the kinetics of the actual reaction, whether the reaction
had gone as planned. As a result of this second coming of the reaction mechanism, then, rational
planning of stereospecific synthesis first became possible.11
I hope to have shown, in this example, how moving inside the elementary reactions allowed
organic chemists to explain new features of chemical reactions. Theories of chemical bonding
and reaction rates, along with the techniques of reaction kinetics, allowed for the solution of
a recalcitrant structural puzzle. Furthermore, the extrapolation of structural analysis into the
theory of reaction rates, made possible by the Transition State Theory, allowed not only for
structural accounts of why a reaction proceeds by a particular mechanism, but also rational
prediction of how novel reactions were likely to proceed.
6. Conclusion
Attempts to explain the relative rates of chemical reactions initially led chemists to introduce
a new theoretical tool—the reaction mechanism. Initially, these were decompositions of a
chemical reaction into sequential steps containing hypothetical intermediates. Eventually,
facilitated by developments in theories of chemical bonding and reaction rates, chemists
moved inside these elementary steps, describing the transitory structures theoretically relevant
to the rates of chemical reactions. With this move, the explanatory power of the reaction
mechanism dramatically increased. Structural features of the products of the most important
types of reaction in organic chemistry were now explicable in mechanistic, and therefore
structural, terms. Furthermore, this new mechanistic understanding could be leveraged to
facilitate the rational planning of novel chemical syntheses, and this has always been the coin
of the realm for the organic chemist.12
Notes
1 See Goodwin (forthcoming) for an account of structural formulas as the primary models of organic
chemistry.
2 The reaction coordinate is the lowest-energy path from reactant to products, over the course of which
some bonds in the reactants are broken, while new bonds in the product are formed. The highest-
energy point along this path is called the transition state.
3 Bartlett (1956, pp. 9–10) says, “If in any organic reaction one could write reliable structural formulas
for the n intermediates and the n+1 transition states, the problem of its mechanism might be considered
solved.”
4 See Goodwin (forthcoming) for the roles of models in chemistry.
5 See Hammett (1940, pp. 96–9) for an account of Lapworth’s reasoning as applied to the halogenation
of ketones.
6 Hammett (1966, p. 466) speculates about some of the ways that work in the kinetics of gas reactions
helped trigger the “renaissance in the study of kinetics and mechanisms of reactions in solution,” but I
am not able to discuss these here.
7 My description of the Walden inversion and its mechanistic explanation draws on the accounts in
Ramsay (1981, pp. 107–15), Ingold (1953, pp. 372–400), and Hammett (1940, pp. 157–83).
8 It is nucleophilic substitutions that are relevant to the Walden inversion. There are also electrophilic
substitutions, which I will not discuss in this chapter.
9 An asymmetric carbon is one with four distinct groups bound to it, which would generally make a
molecule containing such a carbon optically active.
57
William Goodwin
10 It is, not surprisingly, substantially more complicated than this. There are some circumstances in which
the SN1 mechanism will result in retention of configuration and others when it favors inversion, but
these exceptions are also rationalizable in mechanistic terms.
11 This is not meant to suggest that the only significant role of mechanistic studies is in synthetic design.
As a helpful referee has pointed out, mechanistic understanding of organic chemistry has also facilitated
rational drug design, which is increasingly replacing discovery by trial and error.
12 See Goodwin (2011) for more on how mechanisms contribute to synthesis.
References
Bartlett, Paul D. (1956). “Reaction Mechanisms” in Perspectives in Organic Chemistry, ed. Sir Alexander
Todd. Interscience Publishers, New York, pp. 9–27.
Brock, W. H. (2000). The Chemical Tree: A History of Chemistry. Norton, New York.
Desch, C. H. and Morgan, G. T. (1909). “Organic Chemistry.” Annual Reports on Progress in Chemistry
1908 5: 74.
Desch, C. H. and Lapworth, A. (1910). “Organic Chemistry.” Annual Reports on Progress in Chemistry 1909
6:68.
Duhem, Pierre. (2002). Mixture and Chemical Combination. ed. and trans. Paul Needham. Dordrecht,
Kluwer Academic Publishers.
Goodwin, W. (2011). “Mechanisms and Chemical Reaction” in Handbook of the Philosophy of Science:
Philosophy of Chemistry, ed. A. Woody, R.F. Hendry, and P. Needham. Elsevier Science, the
Netherlands, pp. 309–27.
Goodwin, W. (forthcoming). “Models of Chemical Structure” to appear in the Springer Handbook of
Model-Based Science.
Gould, E. (1959). Mechanism and Structure in Organic Chemistry. Henry Holt and Company, New York.
Hammett, Louis P. (1940). Physical Organic Chemistry: Reaction Rates, Equilibria, Mechanisms. McGraw-Hill
Book Company, New York.
Hammett, Louis P. (1966). “Physical Organic Chemistry in Retrospect.” Journal of Chemical Education
43(9): 464–9.
Ingold, C. K. (1953). Structure and Mechanism in Organic Chemistry. Cornell University Press, Ithaca, NY.
Lapworth, A. (1903). “Reactions Involving the Addition of Hydrogen Cyanide to Carbon Compounds.”
Journal of the Chemical Society 83: 995–1005.
Lapworth, A. (1904). “Reactions Involving the Addition of Hydrogen Cyanide to Carbon Compounds.
Part II. Cyanohydrins Regarded as Complex Acids.” Journal of the Chemical Society 84: 1206–1214.
Lewis, G. N. and Randall, M. (1923). Thermodynamics. McGraw-Hill, New York.
Lowry, T. H. and Richardson, K. S. (1987). Mechanism and Theory in Organic Chemistry, Third Edition.
Harper and Row, New York.
Nye, Mary Joe. (1993). From Chemical Philosophy to Theoretical Chemistry: Dynamics of Matter and Dynamics
of Disciplines 1800–1950. University of California Press, Berkeley.
Ramsay, O. B. (1981). Stereochemistry. Heyden, London.
Robinson, R. (1947). “Arthur Lapworth, 1872–1941.” Journal of the Chemical Society 989–996.
Russell, C. A. (1971). The History of Valency. Humanities Press, New York.
Saltzman, Martin (1972). “Arthur Lapworth: The Genesis of Reaction Mechanism.” Journal of Chemical
Education 49(11): 750–752.
Schofield, K. (1995). “Some Aspects of the Work of Arthur Lapworth.” Ambix 42(3): 160–186.
Watson, H. B. (1938). “Organic Chemistry.” Annual Reports on Progress in Chemistry 1938 35: 208.
58
5
MECHANISM, ORGANICISM,
AND VITALISM
Garland E. Allen
1. Introduction
The term “mechanism” has been used widely in the life sciences at least since the seventeenth
century. Embodied in the “mechanical philosophy” of Thomas Hobbes (1588–1679), Pierre
Gassendi (1592–1655), René Descartes (1596–1650), and Robert Boyle (1627–91), among
many others, it came to dominate both the epistemology and ontology of virtually all modern
western science (Durbin, 1988). In this sense “Mechanism” refers to a world-view or philo-
sophical system that sees nature from a materialist perspective and to varying degrees, sometimes
literally, sometimes figuratively, marked with analogies to machines or machine-like processes.
It is also still referred to as the “mechanical philosophy” or “metaphysical Mechanism,” since
it is “concerned with questions about the constituents and organization of the natural world”
(Glennan and Illari, Chapter 1, this volume). However, as philosophers of science know, the
term “mechanism” has a second, related, and overlapping meaning designated as operative or
explanatory mechanism (hereafter, “operative mechanism”).
Operative mechanisms are processes in which a specific set of localized interactions of com-
ponent parts leads to some regular or predictable outcome, as in the mechanism of enzyme action
or the mechanism of genetic recombination. According to “new mechanist” philosophers such
as Peter Machamer, Lindley Darden, and Carl Craver, these mechanisms involve a set of enti-
ties, activities, and spatial-temporal interactions that produce and explain the phenomenon in
question (Machamer, Darden, and Craver, 2000; Craver and Darden, 2013; see also Bechtel
and Abrahamsen, 2005). Following the editors’ format laid out in Chapter 1, when the word
“mechanism” refers to a philosophical or world-view, it will be capitalized, while when it refers
to operative or local mechanisms it will not. This chapter will focus on the historical develop-
ment of philosophical Mechanism beginning with the seventeenth century, though focusing
primarily on the nineteenth and twentieth centuries.
While focusing primarily on the philosophical aspects of Mechanism in the life sciences, it
is also important to emphasize that this world-view was not merely an abstract philosophical
discourse, but one that evolved in various social and economic contexts, and was influenced
not only by changing technologies available in other sciences, but also by economic, social, and
political developments in society at large. Its uses and meaning have always been very much
context-dependent.
59
Garland E. Allen
2. Mechanism and materialism in the life sciences

As pointed out in Chapter 1, Mechanism derives from the mechanical philosophy of the six-
teenth and seventeenth centuries, and in various forms has been one of the dominant philo-
sophical and methodological approaches in the life sciences up to the present day (Boas, 1952;
Dijksterhuis, 1961: 3, 431; Durbin, 1988: 179; Shapin, 1996). Mechanism is just one of several
forms of a broader philosophy of materialism and encompasses, though is not necessarily lim-
ited to, the associated concepts of atomism and matter-in-motion tracing back to the writings
of Democritus and Lucretius in the classical world (Durbin, 1988). Throughout its history
“Mechanism” has been used in both an ontological and epistemological sense. For the most
part, however, working biologists have been less prone to engage in extended ontological dis-
cussions about whether organisms really are nothing more than complex machines or aggregates
of chemical reactions, preferring to use mechanical or chemical analogies as a heuristic device to
pose questions and test hypotheses experimentally.
While differing in focus, earlier Mechanisms and new Mechanism have overlapping com-
mitments. Deriving from the machine analogy, both expect that any process can be described in
terms of material, physical components that work together in an organized way harmoniously
and consistently. For example, just as the cogs in a gear-shift mechanism fit together to produce
the motion of a vehicle, so during enzyme-mediated catalysis a substrate fits into the active site
of its enzyme in such a way that specific chemical bond rearrangements are facilitated, yielding
a new end-product.
Although Mechanism has tended to dominate much of western science, especially biology,
over the past 350+ years, periodically voices have been raised in protest against what seemed like
mechanists’ over-simplistic way of viewing organisms. The alternative has been one or another
forms of what has been referred to as “holism” or “organicism” that claimed living organisms
could never be fully understood as, or reduced to, simple mechanical or chemical processes, but
had to be looked at as “wholes,” as entities in themselves, not merely as an additive aggregation
of parts. In the period 1700–1850 many of these alternative views were mostly couched in mys-
tical terms, as the result of some immaterial, “vital,” or “directive” force (“vitalism”) that had
no counterpart in the non-living world. Many vitalistic theories were considered to be serious
alternatives to standard mechanistic explanations in that they posited law-like activities, but just
not the sort that arose from ordinary physical or chemical processes.1 These mechanist-vitalist
debates punctuated the history of biology, producing a pendulum swing at times away from
Mechanism, at times back toward it, with each swing, however, altering in various ways the
mechanistic research programs that were under dispute. The periods of the most blatant promo-
tion of mechanistic thinking include the mid-to-late seventeenth and eighteenth centuries, the
mid-nineteenth century, and much of the twentieth century.
Mechanism is a form of materialism that gained ascendancy during the scientific revolution
of the seventeenth century as the Mechanical Philosophy, or later, mechanistic materialism
(Boas, 1952; Westfall, 1971; Shapin, 1996). Materialism is the view that all phenomena in
the universe depend upon the interactions of some sort of material particles (derived from
Greek “atomism”) in continuous motion, whose collisions generate events/phenomena in
the world. It also embodies the assumption, or metaphysical claim, that our ideas about
phenomena derive from interaction with the material world through our senses. This means
that matter is primary and our ideas about it are secondary, or derivative (in contrast to
non-materialist, or “idealistic” philosophies that claim ideas are primary, or exist as abstract
categories, and the material world is derived from these pre-existing and non-material cat-
egories). The mechanistic world-view also claims that nature can best be understood as a
60
Mechanism, organicism, and vitalism
mosaic of separate parts, a detailed description of which, when combined together, yields a
complete description/understanding of the phenomena in question.
From the mechanistic materialist point of view, the proper way to study any system is to
take it apart (analysis) and determine the characteristics of the individual, isolated parts under
as controlled a set of conditions as possible. For some mechanists, such as biochemists studying
enzyme kinetics, the lowest level of analysis might be a solution of purified enzyme and its
substrate. For others, such as cell biologists studying respiration, the lowest level of organiza-
tion might be a group of molecules attached to a specific cell membrane system. To understand
a given biological process, the mechanistic materialist strategy has been to start with isolating
those parts by some analytical procedure and identifying them structurally and functionally.
Mechanistic materialists generally follow some kind of reductionist strategy, that is, to investi-
gate higher-level processes by reducing them to their lower levels of organization: for example,
cells in terms of molecules, organs in terms of cells, organisms in terms of organ-systems, and
the like. In this formulation mechanistic materialism is downward looking, proceeding ulti-
mately to the most general and basic material-level explanation as possible. By treating living
processes in material, molecular, or atomistic terms, reductionism brings biology into complete
accord with physics and chemistry.
In discussing the relationship between parts and wholes, it is sometimes assumed that mecha-
nistic materialists have no appreciation of complexity and no interest in how components in a
system interact to produce an overall effect. This is not the case for most mechanists. For exam-
ple, new mechanists identify the components of a complex system (muscle cells, for example) to
understand how they work together in relation to the whole (coordinated muscle contraction).
For mechanistic materialists, investigating the components separately reveals their individual
characteristics more precisely than is possible when studying them as part of a complex whole.
However, a tacit assumption of most mechanistic materialists is that once the characteristics of
each component are known, their relationship to each other and to the whole will become
apparent. It has been a cardinal principle of Mechanism that the properties of living systems
can be understood in terms of the laws of physics and chemistry (a view sometimes referred to
as “physicalism”). Thus, the lowest level of organization to which mechanists have tended to
reduce complex phenomena has been that of atoms and molecules (Gilbert and Sarkar, 2000;
Bechtel and Richardson, 1993; Schaffner, 1967, 1976; Allen, 1975).
In their attempt to model biology after the physical sciences in the later nineteenth and early
twentieth centuries, mechanists emphasized the importance of experimentation, which allowed
the biologist to distinguish between alternative hypotheses; hypotheses that were incapable of
experimental test—for example, that living systems were organized by a non-chemical, non-
physical “vital” force—were considered of no scientific value. Experimentation thus served
epistemologically as a corrective against unbridled speculation and idealistic metaphysics.
Standing in opposition to mechanistic materialism in the later nineteenth and early twen-
tieth centuries were a variety of views that can be clustered under the general term “holism”
(also called “organicism” or “emergentism”) which, as the names imply, were concerned with
how complex systems function organically, as a unified “whole.” Holistic biologists sought to
provide an approach that differed from what they saw as the piecemeal, oftentimes naïve and
simplistic views promoted by zealous mechanists such as Carl Vogt or Jacques Loeb (see below).
In the period starting around 1880, two categories of holistic thinking emerged, especially
in Europe: materialist and non-materialist, though the two were often mistakenly conflated.
Among the former are included holistic materialism and dialectical materialism, and among
the latter vitalism and some versions of organicism. Holistic and dialectical materialism share a
materialist epistemology, seeking to account for living processes as functioning wholes within
61
Garland E. Allen
the framework of known physical and chemical laws. Vitalism, on the other hand, claims that
living organisms defy description in purely physico-chemical terms, because organisms possess
some non-material, non-measurable forces or directive agents that account for their complexity.
Vitalism was regarded by mechanistic materialists as fuzzy-minded and subjective, offering no
real guidelines for practical investigation.
Gilbert and Sarkar (2000) have pointed out that even though other versions of holistic think-
ing in the early twentieth century shared with Mechanism a strictly materialist epistemology,
they were nonetheless largely ignored or rejected because they were perceived as keeping “very
bad company”—either with vitalism of one sort or another, Nazi organicism, or Marxist dia-
lectical materialism—all of which advanced some kind of holistic/organicist view (Gilbert and
Sarkar, 2000: 4–5). Yet, in substance, holistic materialism, and the more formally developed sub-
category, dialectical materialism, stand in sharp contrast to vitalism in not postulating any forces
that cannot be understood in terms of the known laws of physics and chemistry. What united
all forms of holism was the clear recognition that living organisms were capable of activities that
had no counterpart in the machine world: self-replication, purposeful or ordered response to
stimuli, elaborate self-regulatory capabilities, and the incredible efficiency of their energy trans-
duction. To understand these complex interactive processes, it was necessary to get beyond
the individual parts to look at the whole system or process. In the period 1900–40, finding
methods—conceptual and experimental—to accomplish these investigations was a major goal of
holistically oriented materialist biologists (including the embryologists Hans Spemann, Ludwig
von Bertalanffy, Joseph Needham, Paul Weiss, and the geneticist Richard Goldschmidt).
As part of their opposition to mechanistic thinking, holistic materialists emphasized the
importance of distinguishing between levels of organization in a complex system (in studying
organisms this could include the atomic, molecular, cellular, tissue, organ-system, organismic,
population, or ecosystem levels), proceeding to investigate each level on its own terms. The
holistic approach does not preclude starting with a reductionist, analytical breakdown of a com-
plex system into its component parts, but it does emphasize that this is not enough. For example,
in studying the function of the mammalian kidney, holistic materialists might first determine all
of the tissue and cell types of which the kidney is composed. These might be studied initially in
isolation, using methods of histology, cytology, and cell physiology. However, holistic material-
ists argue that it is also necessary to study the kidney in the context of the whole, functioning
organism where its response to variables such as blood pressure, ion concentration, and hor-
mones could be revealed. Isolating and studying separately the ten or so individual tissue types
that make up the kidney could not be expected to provide a full picture of how the kidney
functions as a whole within the intact organism.
The nephron, the main filtration site in the kidney, depends for its function completely on its
interaction with cells in different regions of the kidney, from the cortex, where the first major
filtration of numerous ions occurs, to the medulla, where selective reabsorption takes place.
Each level of organization within the organ—cellular, tissue, and organ—has its own properties
and characteristics that cannot be understood only by examining them separately. In addition,
these tissues types are organized anatomically in such a way that their position within the kidney
is a crucial aspect of their function. It is in their appreciation of the concept that each level of
organization in a complex system has its own special properties, and that these must by studied
by techniques appropriate for that level, that holists have differed in one significant way from
mechanists in the past.
For holistic biologists, complex systems (even relatively simple ones) show emergent prop-
erties that are the product of the individual parts plus their interactions (what we call today
synergistic effects are an example of emergent properties). To those holistic thinkers committed to
62
materialistic explanations, emergent properties were not mystical, since they resulted from the
interaction of the parts within the whole. But emergent properties could not be predicted from
knowing only the individual components making up a given system. Explanations relevant to a
given level had to be derived from studying that level itself and not merely by isolating, describ-
ing, and then extrapolating upward to higher levels.
A more formalized version of holism, known in the twentieth century as dialectical materi-
alism, became the official philosophy of science (and social science) in the Soviet Union after
1917 and the People’s Republic of China after 1949. Dialectical materialism shares all the basic
characteristic of holistic materialism: concern with levels of organization, interaction of parts,
and emergent properties, but it added several important features. The most unique and defining
of these is the dialectical insistence itself: that all processes can be best understood in terms of
the interaction of opposing forces, or agents within a system, and between any system and its
external environment. A classic example of dialectical materialist thinking applied to biology has
been Darwinism, where the dynamics of evolutionary change is constantly moved forward by
the interaction of two opposing tendencies: heredity (non-variable reproduction) and variation
(variable reproduction), adjudicated by natural selection (Prenant, 1943: 138).
With either process by itself there is no evolution; with both present, evolution becomes
inevitable. A chief feature of dialectical materialism, its advocates point out, is that it provides
a way to investigate and understand the dynamic change that characterizes all systems in the
universe. Random events may of course affect the way in which any system changes, but more
constant factors driving change lie at a deeper level internal to the system itself. Further discus-
sion of the characteristics of dialectical materialism take us too far afield from the discussion of
mechanistic materialism. The point of mentioning it here is to indicate that from the 1860s
onward another version of holistic thinking existed that was overtly and self-consciously mate-
rialistic and devoid of mystical or vitalistic overtones. That it was not more consciously pursued
was a result of the political climate following the Bolshevik revolution and the ensuing gulf
between Soviet and western science and philosophy (Graham, 1986).
3. Mechanistic philosophy in the nineteenth and twentieth centuries

Early mechanistic approaches to understanding living organisms in the seventeenth and eight-
eenth centuries were highly simplistic and literally cast in terms of familiar mechanical devices
and processes. They were a way to explain biological processes in terms and models that were
familiar. The use of such mechanistic analogies grew in part out of the increasing presence
of machinery in daily life (water pumps, mining hoists, and the like). Organisms, like the
Newtonian universe, operated like machines, with precision, regularity, and predictability, fea-
tures that life scientists wanted to extend to living organisms in the same way physicists had done
for planetary systems. Even in popular culture of the later eighteenth century, the mechanistic
view found expression in a fascination with mechanical models of animals, such as Vaucanson’s
Duck, a wind-up model that could feed, drink, digest, and defecate. Mechanistic models were
expanded and made more sophisticated as the nineteenth century progressed
Historians have dubbed the mid- and later nineteenth century as the “age of materialism,” in
part because of the rise of industrial capitalism, which emphasized the production and distribu-
tion of material commodities, a concern for the acquisition and movement of capital, and the
increased presence of machines in the workplace. It was in this period that Marx worked out
the laws of capitalist development, placing them in the context of a materialist theory of history.
This widespread materialist culture also had its own versions of mechanistic theories in biology,
particularly in medicine and physiology.
63
Garland E. Allen
Medical materialism developed in the mid-nineteenth century, particularly in France and

Germany, in the wake of the revolutions of 1848. Ideologically, materialism was associated with
the attempt to eliminate the Cartesian dualism between body (viewed mechanistically) and the
mind/soul (viewed idealistically, non-materially), and as part of the growing anti-clericalism
of the period (Temkin, 1946: 324). In France, Henri Dutrochet (1776–1847) investigated
the passage of materials across membranes as a purely physical process and Francois Magendie
(1783–1855) pioneered experimental studies of nutrition (using animal models) and the struc-
ture and function of the nervous system. Magendie’s successor at the College de France, Claude
Bernard (1813–78), pursued a quasi-mechanistic research program in his studies on the glyco-
genic function of the liver and the maintenance of the milieu intérieur, or internal environment.
He did, however, leave some room for vitalistic processes that could not be explained solely by
physical or chemical means (Temkin, 1946: 326).
In Germany, too, physiology was at the center of a new mechanistic biology, especially
embodied in the influential Berlin medical materialists of the late 1840s and 1850s: Ludwig
von Helmholtz (1821–94), Emil Du Bois-Reymond (1818–96), Carl Ludwig (1816–95),
and Ernst von Brücke (1819–92). Interested particularly in the electrical properties of nerve
and muscle tissue, this group pioneered a thorough physical and chemical research pro-
gram that explicitly made no distinction between organic and inorganic matter. In his
“Investigations concerning Animal Electricity” (1848) Du Bois-Reymond looked forward
to the day when “physiology would dissolve completely into biophysics and biochemistry”
(Temkin, 1946: 326). It was in this same vein that Karl Vogt (1817–95) could write (1845):
“thoughts have about the same relation to the brain as bile has to the liver or urine to the
kidney” (quoted in Temkin, 1946: 324).
Toward the end of the century mechanistic explanations found clear expression in another
area of life sciences: embryology. Wilhelm His (1831–1904) attempted to account for the folds
that occurred in the various germ layers in vertebrate development by mechanical processes
that he illustrated with the folding properties of a series of rubber sheets and tubing. Slightly
later (1881) Wilhelm Roux (1850–1924) advanced his mosaic theory of embryonic develop-
ment (also known as the Roux-Weismann theory, after August Weismann, 1834–1914, who
had come up independently with a similar view). Their focus had been on the mechanism of
differentiation during embryogenesis. Roux claimed with each cell division hereditary units
were parceled out in such a way that each cell generation received increasingly specialized
particles, so that by the time differentiation was complete, each cell type (muscle, nerve,
skin) contained only the particles determining that cell’s specific characteristics. To test his
theory, Roux killed one of the first two blastomeres of the frog egg with a hot needle. True
to his prediction, he got some half-embryos that developed through an early embryonic
(the gastrula) stage. From this work Roux developed an entire research program, known as
Entwicklungsmechanik, aimed at promoting an experimental and mechanistic understanding of
embryonic differentiation.
However, in 1891 a young German embryologist, Hans Driesch (1867–1941), working at
the Naples Zoological Station, carried out a similar series of experiments using the sea urchin.
Instead of killing one of the blastomeres, Driesch separated them by vigorously shaking the
cultures. Driesch expected half-embryos also, but found that the separated blastomeres each
developed into a complete, though smaller, embryo. Driesch, who was also initially a mecha-
nistic materialist, interpreted the results as contradicting Roux’s strictly mechanical model. No
machine could reconstruct the whole out of an incomplete set of individual parts. The sea
urchin embryo acted as what Driesch called a “harmonious equipotential system,” and for sev-
eral subsequent years he carried out experiments to study the characteristics of self-regulation
64
and adjustment of embryos to altered conditions. Eventually, by the early 1900s, however, he
despaired of finding a mechanistic, physico-chemical solution to the problem, and adopted an
increasingly vitalistic interpretation (see below).
4. The early twentieth century: Jacques Loeb and the

mechanistic conception of life
In the twentieth century Mechanism found its most explicit and forceful exponent in a German
émigré to the United States, physiologist Jacques Loeb (1858–1924). Loeb’s work illustrates
more clearly than many others how one form of mechanistic materialism was put into practice
as a scientific research program.
Born in Germany in the same year as the publication of the influential essay “The Mechanistic
Conception of Life” by Rudolf Virchow (1821–1902), Loeb was, in the words of historian
Donald Fleming, “a child of his time” (Fleming, 1964: xi). Virchow’s views were exemplars of
an explicitly materialistic interpretation of the natural and social worlds (Allen, 1992). Interested
in the problems of human volition and the expression of the will, Loeb received an M.D.
degree from the University of Strasbourg, working subsequently in the laboratory of Adolf Fick
(1829–1901), who had been a student of the Berlin medical materialists and espoused their view
that life was nothing more than an expression of physical and chemical laws. Embedded in this
mechanistic view was the assumption, as in physics, of a strictly deterministic universe where, if
one knew all the inputs, it would be theoretically possible to predict the precise outcome of any
process, including human behavior. At Würzburg Loeb also came under the influence of plant
physiologist Julius Sachs (1832–97), who gave him an appreciation for the importance of experi-
mentation and an interest in the specific problem of plant and animal movements, or tropisms.
It was only after his emigration to the United States in 1891 that Loeb’s mechanistic philosophy
became explicit and took its most concrete shape. Loeb’s research on artificial parthenogenesis
(the development of an unfertilized egg into an adult organism) provides a useful illustration of
what Mechanism meant as a biological philosophy.
No process in late nineteenth- or early twentieth-century biology remained more of a bastion
for metaphysical explanations than that of fertilization of the egg by a sperm and the subsequent
developmental events that it triggered. Once the sperm has penetrated the egg, a cascade of
events takes place, rapidly in most cases, leading to the first cleavage, in which the single-celled
zygote becomes a two-celled embryo (the blastomeres in Roux’s experiment). This remarkable
series of events raised many questions: How does the entrance of sperm into the egg cytoplasm
initiate division? Does the point of penetration determine the plane of division, and if so how
does the plane of the first division affect the future axis (anterior-posterior, dorsal-ventral) of the
organism? What role, if any, does the egg cytoplasm play in the course of development? And, of
course, there is the fundamental question of how differentiation takes place.
Although a strong proponent of Roux’s Entwicklungsmechanik program, Loeb at first saw no
contradiction between Driesch’s results and a mechanistic interpretation. He, too, had experi-
mented with sea urchin eggs at Woods Hole, and found that fragments of embryos, or even
eggs, could be stimulated to produce full embryos under the right conditions. It was when
Driesch declared himself an unabashed vitalist that Loeb parted ways, since he saw vitalism as an
unacceptable metaphysics that served no function other than stifling research.
Loeb focused on the problem of fertilization, since nothing seemed to be more basic to
the distinction between living and non-living matter than the moment in the initiation of a
new life. It was also known, however, that under natural conditions, the eggs of some species
(ants, bees, wasps, aphids) undergo normal development without being fertilized, i.e., natural
65
Garland E. Allen
parthenogenesis (Loeb, 1900). Loeb asked whether parthenogenetic development could be

induced by physical or chemical means, which, if possible, could provide a new mechanistic
approach to a fundamental, vital process.
Loeb’s experiments involved placing sea urchin eggs in sea-water of varying osmolarities
(concentration of ions), and of varying types of salt combinations. The ideas behind these
experiments came from Sachs, who had studied the effects of varying salt concentrations on
water uptake, transpiration, and photosynthesis in plants, and from the physical chemist Svante
Arrhenius (1859–1927), whose theories had convinced Loeb that biological phenomena were
controlled by precise ionic concentrations. Living phenomena such as fertilization thus could be
approached from the standpoint of physical chemistry. He hoped to replicate that normal process
by known physico-chemical agents (Loeb, 1912 [1964]: 6).
When Loeb placed unfertilized sea urchin eggs in hypertonic sea-water, he found that on
returning them to normal sea-water, the eggs started to divide (though at first none developed
beyond early blastula stage) (Loeb, 1899: 326–7). By manipulating variables such as types and
concentrations of various ions or the length of time the eggs were immersed, Loeb determined
the proper mixture and concentration of ions and immersion time (2 hours) that would not only
yield a large number of cleavages, but would also support development through the swimming
larval stage (Loeb, 1899: 329). As he wrote:
I consider the chief value of the experiments . . . to be the fact that they transfer the
problem of fertilization from the realm of morphology [which meant at the time
descriptive and often speculative work] into the realm of physical chemistry.
(Loeb, 1899: 332)
Others soon showed that purely physical stimuli, such as pricking the egg with a needle, could
also initiate development. These were all positive confirmations that life could be understood
by employing the concepts of physics and chemistry, and testing them empirically by controlled
experiments. Loeb’s interest in this work was not merely in finding the operative mechanism
by which fertilization actually triggered development, but rather in demonstrating that liv-
ing processes followed the laws of physics and chemistry, and through proper experimental
investigation could be brought under human control. The “Mechanistic Conception of Life”
represented progress for Loeb precisely because it led to control, to the engineering of nature
(Pauly, 1987: 114).
Loeb’s mechanistic view was cast in the reductionist language of physics and chemistry.
Loeb’s form of reductionism sought to explain complex processes by examining their increas-
ingly lower levels of organization. What Loeb meant by reductionism (and he did use the term)
was to trace higher-level processes down to particular physical and chemical reactions. As he
wrote in 1888:
Whatever appear to us as innervations, sensations, psychic phenomena, as they are

called, I seek to conceive through reducing them—in the sense of modern physics—to
the molecular or atomic structure of the protoplasm, which acts in a way that is similar
to (for example) the molecular structure of an optically active crystal.
(Pauly, 1987: 38)
Loeb’s commitment to Mechanism was part of a larger program he shared with many colleagues
to make biology a hard science, and thus to place it on an equal professional footing with the
physical sciences.
66
5. Mechanistic approaches in the later twentieth century

Loeb’s mechanistic philosophy was highly influential in a variety of areas of biology outside of
general physiology. It became part and parcel of Thomas Hunt Morgan’s (1866–1945) research
program in genetics, and was codified in the “beads-on-a-string” representation of genes as
atomistic parts of chromosomes (Morgan et al., 1915: 60). Others saw genes as the biologist’s
equivalent to the atoms or molecules of physics and chemistry, and the organism as a mosaic of
traits determined by the individual, particulate gene. Genetics became an exemplar of the mech-
anistic approach for the period 1915–50. Mechanistic thinking was also highly influential in the
population genetics of the founders of the evolutionary synthesis, particularly in the writings of
Ronald A. Fisher (1890–1962) in 1930. Building on work in laboratory genetics by Morgan
and others, Fisher saw evolution as the process of selection acting on discrete Mendelian genes,
not the organism as a whole. Fisher’s claim, in his path-breaking The Genetical Theory of Natural
Selection (Fisher, 1930), was that by reducing populations of organisms to individual, freely
recombining genes, he could accomplish for evolutionary biology what the kinetic theory of
gases had accomplished for physics and chemistry.
Neurobiology continued the mechanistic trend started by the Berlin medical materialists of
the previous century with studies of neuronal conduction not as the flow of electrons along a
wire, but of ions across the axon membrane down a concentration gradient to produce a self-
perpetuating action potential. Conduction or inhibition across synapses between neurons could
also be reduced to molecules released by the pre-synaptic and absorbed by receptors on the post-
synaptic membrane. Even in psychology, in the first half of the century, mechanistic thinking
was prominent, especially in the behaviorism of John B. Watson (1878–1958) and B.F. Skinner
(1904–90). Their views of operant conditioning were often couched in machine-like language,
and saw the organism and its behavior as the result of a set of input-output processes that was
both reductionistic and deterministic.
The rise of biochemistry and molecular biology in the middle and second half of the cen-
tury provided a new stimulus to mechanistic approaches. Biochemists in the 1940s and 1950s
referred to cells as “bags of enzymes” while elucidation of the molecular structure of deoxyribo-
nucleic acid (DNA) provided one of the most far-reaching revolutions in mechanistic thinking
in recent times. The fact that similar mechanisms for DNA replication, protein synthesis, and
gene control seemed to apply to all organisms was exemplified in the claim by molecular biolo-
gist Gunther Stent (1924–2008) that “What is true of E. coli [as exemplar of all bacteria] is true of
the elephant.” The implication is not only that there is a unity among diverse living organisms
on earth, which biologists already recognized, now extended to the molecular level, but also
that multicellular organisms can be viewed simply as large collections of unicellular organisms.
This brand of Mechanism ignored the emergent properties that characterize the different levels
of organization in complex organisms.
6. Beyond mechanistic materialism: (w)holism and organicism

Since the rise of mechanistic materialism in the seventeenth century, there have been
reactions to what seemed like the overly simplistic machine models of living organisms. Initially,
these were expressed as some form of vitalism in which living beings were thought to possess
non-material, non-measurable energies or forces that were qualitatively different from those
described in non-living entities. Echoes of this vitalism can be found in the writings of French
anatomist and physiologist Xavier Bichat (1771–1802), who saw the ability to experience sensa-
tions and thought as evidences of vital properties very different from those of the inorganic world.
67
Garland E. Allen
However, by the mid-nineteenth century, Driesch notwithstanding, vitalism as such had lost
much of its significance, as it offered no prospects for a meaningful research program.
However, one aspect of the vitalistic perspective that did attract attention, especially in the
early twentieth century, was the holistic or organismic movement, which focused attention on
organisms, populations, and even ecosystems as “wholes”; that is, not as machines atomized into
discrete and separable components. Traces of holistic or organismic thinking are apparent in the
Naturphilosophie movement in Germany in the late eighteenth and early nineteenth centuries,
which emphasized the importance of looking at organisms whose properties were something
more than just a sum of their individual parts. A concern with holistic thinking remained a part
of many aspects of German biology well into the twentieth century, although it had some pro-
ponents in Europe and North America as well.
Historian Anne Harrington has suggested that holism in biology in the early twentieth
century grew out of not only reaction to the biological Mechanism of Roux, Loeb, and oth-
ers, but also as a reaction to the cultural fragmentation associated with “modernism,” World
War I, urbanization, industrialization, and the perceived intrusion of machines into everyday
life. The fragmentation was most painfully obvious in Germany following the harsh political
and financial terms forced on it by the Versailles Treaty after World War I. In combina-
tion with disillusionment and embittered national pride, Germany produced a romanticized
nationalism and nostalgia for a past in which it was believed that humans were more harmo-
niously integrated into the natural world (Harrington, 1996: 19–33). Thus the holistic view
was intimately tied up with a variety of scientific and cultural trends that set it against the
mechanistic world-view that many thought had led to an undermining of morality, the social
fabric and “man’s place in nature.”
In the period between 1900 and 1950, holistic philosophy included a wide spectrum of
approaches with widely differing views on both the philosophical meanings of holism and
the practical applications of holistic thinking to scientific research programs. Driesch was,
of course, the major and most extreme exponent. His claim that embryonic development
was guided by an “entelechy,” an organizing, directive force that consumed no energy, was
immaterial, but it was the factor that distinguished living from non-living matter. There were
many others, however, including Jakob von Uexküll (1864–1944), Ludwig von Bertalanffy
(1901–72), and Henri Bergson (1859–1941), who espoused similar views. In one way or
another each of these individuals advocated an anti-mechanistic, holistic approach, replacing
the machine analogy with an organicism that repudiated the reduction of all living processes
to simple matter-in-motion (Harrington, 1996).
But in the twentieth century there were two sorts of holistic approaches: those that still advo-
cated some sort of vitalistic property of living systems, and those that claimed that living systems
obeyed all the laws of physics and chemistry, yet because of their various levels of organization,
displayed properties that were greater than the sum of their parts. Holists of either type argued
that even if ions in solution in the cell cytoplasm, or in blood plasma, obeyed Arrhenius’ laws of
electrolytic dissociation, that was not what made an organism “alive.” What was important was
that organisms displayed the wholeness of the organic process, characterized by their organiza-
tional relationships. For all holists the whole was always greater than the sum of the parts, thus
showing emergent properties that were not inherent in any of the parts by themselves.
Holists with a vitalistic tinge, such as von Uexküll, interpreted emergent properties as an
expression of a vital property that could never be understood from a purely analytical approach.
In a tribute to Driesch and his concept of the embryo as a “harmonious equipotential system,”
von Uxeküll wrote:
68
Driesch succeeded in proving that the germ cell does not possess a trace of
machine-like structure, but consists throughout of equivalent parts. With that fell
the dogma that the organism is only a machine. Even if life occurs in the fully
organized creature in a machine-like way, the organization of the structureless
germ into a complicated structure is a power sui generis, which is found only in
living things and stands without analogy. . . . It is not to be denied that vitalists are
the victors all along the line.
(Quoted in Harrington, 1996: 51)
To mechanists, this interpretation represented a resurgence of German romanticism, a new

Naturphilosophie that was not only philosophically backward-looking but from a pragmatic point
of view had no significant research potential. At the same time, however, a new form of holis-
tic thinking, what came to be called holistic materialism, was emerging. Holistic materialism
attempted to approach biological systems with an appreciation of their emergent properties in
materialist terms that provided the basis for an experimental research program. However, so
powerful were the forces of mechanistic thinking in the first half of the twentieth century that
attempts at holistic alternatives, even when expressed in materialistic, non-mystical ways, were
easily misunderstood or viewed with considerable suspicion.
Among those who tried to develop a holistic materialist, experimental framework was
German embryologist Hans Spemann (1869–1941) and his students in the 1920s–40s. Trained,
like Loeb, at Würzburg in morphology, Spemann took up experimental work in embryology as
an alternative to the purely descriptive and speculative approach of his mentors. Although appre-
ciating Driesch’s motivations to treat the embryo holistically, Spemann could not agree with
his overtly vitalistic interpretation. Starting in the late 1890s through the mid-1920s, Spemann
developed the concept of embryonic induction to account for the sequence of events in tissue
differentiation. His approach was holistic, materialistic, and experimental, based on remov-
ing tissues during different stages of development and transplanting them to various regions of
younger or older embryos. His initial example was the induction of the lens in the developing
vertebrate (frog) eye. Spemann noted that the lens began to form when the optic vesicle, an out-
growth of the underlying brain tissue, first made contact with the overlying ectodermal tissue.
Spemann showed experimentally that if the optic vesicle were removed, the overlying ectoderm
did not differentiate into a lens; conversely, if the optic vesicle were transplanted from the ante-
rior to the flank or posterior region of the embryo, it produced lenses in these otherwise eyeless
regions. Spemann hypothesized that induction required physical and (presumably) chemical
contact between the inducing and the induced tissue. For Spemann this process became a para-
digm for how differentiation could be viewed in the embryo as a whole: as a series of inductive
cascades of increasingly greater specificity (Hamburger, 1988: 18).
Spemann then took the process of induction to even earlier embryonic stages: to the gastrula
stage after the blastula has begun to invaginate (push inward) to create a two-layered embryo.
The point of invagination was known as the blastopore. Spemann, and his graduate student
Hilde Mangold (neé Proescholdt, 1898–1924), found that transplanting tissue from the dorsal
(upper) side of the blastopore into the ventral (underside) of an older embryo induced the for-
mation of a whole secondary embryo. The dorsal lip tissue appeared to be the primary inducer,
or “organizer” as it came to be called, of the entire vertebrate embryonic process. Working out
details of factors influencing the function of organizers became the focus of work in his labora-
tory for over a decade (1924–35). (For this work Spemann was awarded the 1935 Nobel Prize
in Physiology or Medicine.)
69
Garland E. Allen
As a holistic materialist, Spemann focused his experimental work only at the tissue and
organ-system levels of organization. He was not interested in reducing the process of induc-
tion to the chemical, physical, or even genetic level. Spemann’s view was similar to Driesch’s
concept of the embryo as a “harmonious equipotential system” (a term that Spemann himself
adopted and used), but not in a vitalistic way. Spemann thought that induction was a process
based on interaction of material components and that while molecules were certainly involved,
reduction to the molecular level would lose the essence of the process itself. For Spemann, the
most interesting processes could only be revealed at those higher levels at which their emergent
properties occurred. The basis of the organizer was to be found in its wholeness: inducing tissue
had no meaning except in its interaction with inducible tissue. The whole, then, was greater
than the sum of the parts in isolation.
Other embryologists followed suit. Spemann’s student, Viktor Hamburger (1900–2001), car-
ried on his mentor’s holistic approach, but with a greater appreciation for the necessary interplay
between mechanistic and reductionist approaches. From the mid-1930s onward, Hamburger
pioneered work on development of the nervous system in the chick, noting that while some
inductive process triggered neurons to begin growing out from the spinal cord toward devel-
oping limb buds, if the limb buds were removed nerve growth was significantly reduced (this
work led to the later discovery of nerve growth factor, or NGF). Both tissues were involved in
a reciprocal interaction and thus led to the emergence of fully innervated and functional limbs.
Other developmental biologists who were greatly influenced by Spemann in the 1940s–70s
included Paul A. Weiss (1898–1989) and Conrad H. Waddington (1905–75). Weiss developed
a holistic view of the embryo as a series of “fields” in which various inductions and interaction
took place at different times during embryogenesis. Waddington made significant advances in
integrating genetics with development in a metaphorical “landscape” in which both systems
co-evolved through the action of natural selection. More than in most other fields, embryolo-
gists promoted a holistic materialism that went beyond older mechanistic conceptions of life.
7. Conclusion
This chapter has focused on characterizing the mechanistic materialist philosophy that has been
at the center of attempts to understand living systems from the early seventeenth through the
end of the twentieth century, and to distinguish it from various emerging holistic or organismic
philosophies, often associated with some form of vitalism. Looking beyond the latter years of
the twentieth century, however, it is apparent that there is a flowering of holistic materialist
and even dialectical materialist views in biology that is taking the life sciences in new directions.
Gilbert and Sarkar have provided an exemplary survey of organicist thinking in widely differ-
ent fields, from developmental biology, evo-devo (evolutionary developmental biology), ecol-
ogy, neurobiology, and other interdisciplinary attempts to see biological systems as composed
of multiple levels of organization with emergent properties at each new level of organization
(Gilbert and Sarkar, 2000).
Richard Burian has argued that the older Mendelian concept of the gene as a discrete atom-
istic unit was grossly oversimplified in light of new work in molecular genetics, where segments
of DNA are not functionally or even structurally discrete, and thus the Mendelian gene cannot
be reduced to the molecular gene. In this sense, perhaps nowhere has the shift from a mecha-
nistic approach and mechanistic research strategy between 1920 and 2000 been more apparent
than in genetics as a whole. Yet nowhere in this changing philosophical approach has there been
any tendency to return to forms of vitalism or non-material processes. The old mechanist-vitalist
debates of the 1890s through the 1930s have been purged from biology at every level.
70
There has also been a small but significant resurgence of formalized dialectical materialist
views in the life sciences, particularly in the work of evolutionary biologists such as Richard
Lewontin (1929–) and Richard Levins (1930–2016). Their elaboration of a dialectical philoso-
phy in various fields of evolutionary biology, population genetics, and ecology in the book
The Dialectical Biologist suggests how a dialectical materialist approach can be highly fruitful as
a way to understand the dynamics of living processes (Levins and Lewontin, 1985). They also
suggest that if you really observe nature holistically, you cannot help but develop a dialectical
materialist approach, even if not expressed as a formal philosophical system. This point has
been made with respect to Darwin, who was not very philosophically inclined but couched his
whole theory of evolution by natural selection in terms of the interaction of opposing forces,
or processes (Allen, 1989, 1991, 1992).
Today, with the advent of computer technologies and fields such as bioinformatics with
large data sets, systems, and “complexity” theory, holistic thinking is no longer so suspect in
a variety of fields. As biology has begun to supplant physics as the pre-eminent science of the
day, its philosophical underpinnings are not only becoming more sophisticated, but also less
rigid. Like organisms themselves, biological approaches to understanding their functions require
a variety of approaches, and must be flexible depending on what system is being studied, what
questions are being asked, and what methods are available for research. Both analytical (Bechtel
and Richardson’s “decomposition”) and synthetic, holistic approaches are necessary and com-
plementary. New mechanists such as Craver and Darden argue that the search for mechanisms
is not reductionist in the classical sense, as mechanisms can be studied at various different levels
of organization. These new approaches are less monolithic and more flexible. As evolutionary
biologist and historian Ernst Mayr (1904–2005) argued, biology is not bound to try to subsume
all its theories under general laws, but can be free to develop its generalizations in a statistical
sense, allowing for variation not only in the organisms being studied themselves but also in the
messy world of lab and field observation and experimentation.
Note
1 As a movement in the philosophy of science, vitalistic claims should not be confused with religious argu-
ments such as Special Creationism or Intelligent Design, which also appeal to metaphysical explanations,
but couch them in terms of an individual, unknowable creator or deity.
References
Allen, G. (1975) Life Science in the Twentieth Century, New York: Cambridge University Press.
Allen, G. (1989) “Dialectical Materialism in Modern Biology,” Science and Nature 3: 43–57.
Allen, G. (1991) “Mechanistic and Dialectical Materialism in 20th Century Evolutionary Theory: The
Work of Ivan I. Schmalhausen,” in L. Warren and H. Koprowski (eds), New Perspectives on Evolution,
New York: John Wiley, 15–36.
Allen, G. (1992) “Evolution and History: History as Science and Science as History,” in M. Nitecki and
D. Nitecki (eds), History and Evolution, Albany, NY: State University of New York Press, 211–239.
Allen, G.E. (2001) “The Classical Gene: Its Nature and Its Legacy,” in L.S. Parker and R. Ankeny (eds),
Mutating Concepts, Evolving Disciplines: Genetics, Medicine and Society, Dordrecht: Kluwer Academic
Publishers, 11–41.
Ash, M. (1995) Gestalt Psychology in German Culture, 1890–1967: Holism and the Quest for Objectivity,
Cambridge: Cambridge University Press.
Bechtel, W. and A. Abrahamsen (2005) “Explanation: A Mechanistic Alternative,” Studies in History and
Philosophy of Biology and Biomedical Sciences 36: 421–441.
Bechtel, W. and R. Richardson (1993) Discovering Complexity, Princeton, NJ: Princeton University Press.
Boas, M. (1952) “The Establishment of the Mechanical Philosophy,” Osiris 10: 412–541.
71
Garland E. Allen
Clark, A. (1997) Being There. Putting Brain, Body and World Together, Cambridge, MA: MIT Press.
Commoner, B. (2002) “Unraveling the DNA Myth,” Harper’s Magazine (February): 39–47.
Craver, C. (2001) “Role Functions, Mechanisms and Hierarchy,” Philosophy of Science 68: 53–74.
Craver, C. and L. Darden (2013) Thinking about Mechanisms: Discovery across the Life Sciences, Chicago:
Dijksterhuis, E.J. (1961) The Mechanization of the World Picture, Oxford: Oxford University Press.
Durbin, Paul T. (1988) Dictionary of Concepts in the Philosophy of Science, New York: Greenwood Press.
Engels, F. (1940 [1935]), The Dialectics of Nature. “Preface” to the English translation by J.B.S. Haldne,
New York: International Publishers.
Fisher, R.A. (1930) The Genetical Theory of Natural Selection, Oxford: Clarendon Press.
Fleming, D. (1964) “Introduction,” in Jacques Loeb, The Mechanistic Conception of Life, Cambridge, MA:
Harvard University Press, vii–xlii.
Gilbert, S. and S. Sarkar (2000). “Embracing Complexity: Organicism for the 21st Century,” Developmental
Dynamics 219: 1–9.
Graham, L. (1986) Science, Philosophy and Human Behavior in the Soviet Union (2nd ed.), New York:
Columbia University Press.
Hamburger, V. (1984) “Hilde Mangold, Co-discoverer of the Organizer,” Journal of the History of Biology
17: 1–11. Also reprinted as an Appendix to Hamburger, 1988: 173–180.
Hamburger, V. (1985) “Hans Spemann, Nobel Laureate 1935,” Trends in NeuroSciences 8: 385–387.
Hamburger, V. (1988) The Heritage of Experimental Embryology, New York: Oxford University Press.
Harrington, A. (1996) Reenchanted Science: Holism in German Culture from Wilhelm II to Hitler, Princeton,
NJ: Princeton University Press.
Holtfreter, J. (1933) “Nachweis der Induktionsfähigkeit abgetöter Keimteile.” Roux’ Archiv für
Entwicklungsmechanik 128: 584–633.
Levins, R. and R. Lewontin (1985). The Dialectical Biologist, Cambridge, MA: Harvard University Press.
Loeb, J. (1890) Der Heliotropismus der Tiere und seine Übereinstimmung mit dem Heliotropismuus der Pflanzen.
Würzburg: Georg Hertz. Translated in Loeb, J. (1905) Studies in General Physiology, Chicago: University
of Chicago Press, 1–88.
Loeb, J. (1899) “On Some Facts and Principles of Physiological Morphology,” Biological Lectures Delivered
at the Marine Biological Laboratory of Woods Hole, Summer, 1893, Boston: Ginn & Company, 37–61.
Reprinted in Maienschein (1986).
Loeb, J. (1900) “On the Nature of the Process of Fertilization,” Biological Lectures Delivered at the Marine
Biological Laboratory of Woods Hole, Summer, 1899, Boston: Ginn & Company, 273–282. Reprinted in
Maienschein (1986).
Loeb J. (1903) “Sind die Lebenserscheinungen wissenschaftlich und vollständig erklärbar?” Die Umschau
7: 21; presented also as a lecture at Berkeley, as “Phenomena of Life,” and reported in The Daily
Californian, March 6, 1903. Quoted in Pauly, 1987.
Loeb, J. (1904) “The Recent Development of Biology,” Science 20: 781.
Loeb, J. (1909) “On the Nature of Formative Stimulation (Artificial Parthenogenesis),” in D. Fleming
(ed.), The Mechanistic Conception of Life, Cambridge, MA: Harvard University Press.
Loeb, J. (1912) “The Mechanistic Conception of Life,” Popular Science Monthly (January). Reprinted in
Fleming, D., ed. (1964) The Mechanistic Conception of Life, Cambridge, MA: Harvard University Press,
5–34.
Machamer, P., L. Darden, and C. Craver (2000) “Thinking about Mechanisms,” Philosophy of Science 67:
1–25.
Maienschein, J. (1996) Defining Biology: Lectures from the 1890s, Cambridge, MA: Harvard University Press.
Morgan, T.H., A.H. Sturtevant, H.J. Muller, and C.B. Bridges (1915) The Mechanism of Mendelian Heredity,
New York: Henry Holt.
Oparin, A.I. (1964 [1938]) Life: Its Nature, Origin and Development. Translated by Ann Syng, New York:
Academic Press.
Pauly, P. (1987) Controlling Life: Jacques Loeb and the Engineering Ideal in Biology, Chicago: University of
Chicago Press.
Prenant, M. (1943) Biology and Marxism. Translated by C. Desmond Greaves, New York: International
Publishers.
Provine, W. (2001) The Origins of Theoretical Population Genetics, Chicago: University of Chicago Press.
Reprint of the 1972 edition, with a new “Afterword.”
Schaffner, K. (1967) “Approaches to Reductionism,” Philosophy of Science 34: 137–147.
72
Schaffner, K. (1976) “Reductionism in Biology: Prospects and Problems,” in R.S. Cohena and
A. Michalos (eds), Proceedings of the 1974 Meeting of the Philosophy of Science Association, Dordrecht:
D. Reidel, 613–632.
Schmalhausen, I. (1949) Factors of Evolution, Philadelphia: Blakiston; Reprint ed., Chicago: University of
Chicago Press (1986).
Shapin, S. (1996) The Scientific Revolution, Chicago: University of Chicago Press.
Spemann, H. (1923) “Zur Theorie der tierischen Entiwcklung,” Rektoratsrede 1–16. Freiburg im Breisgau:
Speyer und Kaerner.
Spemann, H. (1927) “Neuen Arbeiten über Organisatoren in der tierischen Entwicklung,” Naturwissenschaften
15: 946–951.
Spemann, H. (1938) Embryonic Development and Induction, New Haven, CT: Yale University Press.
Stein, G. (1988) “Biological Science and the Roots of Nazism,” American Scientist 76: 50–58.
Temkin, O. (1946) “Materialism in French and German Physiology of the Early Nineteenth Century,”
Bulletin of the History of Medicine 20: 322–327.
Uexküll, J. (1937) “Die neue Umweltlehre. Ein Binderglied zwischen Natur- und Kulturwissenschaften,”
Der Erziehung: Monatschrift für den Ausammenhang von Kultur und Erziehung im Wissenschaft und Leben 13:
185–199.
Westfall, R. (1971) The Construction of Modern Science, New York: John Wiley.
73
6
MECHANISMS AND THE
MENTAL
Marcin Miłkowski
1. Introduction: the notion of mechanism evolves with the

understanding of the mental
In this chapter, I sketch the history of mechanistic models of the mental, as related to the
technological project of trying to build mechanical minds, and discuss the changing use of
these models. In section 2, I introduce the Cartesian notion of mechanism, which shaped the
debate in the centuries that followed. Early mechanistic proposals are also connected with early
attempts to formulate the computational account of thinking and reasoning, upheld notably
by Hobbes and Leibniz. In section 3, associationist and behaviorist models of the mind are
sketched, along with attempts to understand the neural system in terms of connections and
associations. Early robotic models, built mostly by behaviorists and other students of animal
behavior, are also introduced. In section 4, the focus is on early computational and cybernetic
models of the mind. In section 5, I deal with computational models of mental mechanisms
as proposed by students of artificial intelligence and cognitive science. The history of uses of
mechanistic models sheds light on different kinds of explanations of the mental.
The notion of mechanism has played an immense role in the history of attempts to sci-
entifically explain mental phenomena. On the one hand, it has always been related to the
philosophical mind–body problem, and, on the other hand, to the possible scope of mechanistic
explanations. Mechanization of thinking is also one of the classical problems in philosophy of
mathematics, and the notion of mechanism, in a much more abstract form, appears in the history
of mathematics as well.
The philosophical mind–body problem is the question of how the mental and the physical
are related. One way to defend a naturalistic solution to the mind–body problem is to build a
mechanistic model of the mind. Such a model might show, even as a proof of an in-principle
possibility, that thinking can be instantiated in a physical mechanism. Conversely, the proof that
thinking may not be mechanized hints at the possibility that the mind is not physical, after all.
However, one may also contend that mechanistic models of thinking, even if successful, do
not decide the mind–body question at all. In other words, even if thinking is just the exercise
of computational mechanisms, there still could be some non-mechanistic aspects of thought
or perception that completely evade the mechanistic explanation. Indeed, this is the position
taken by computational dualists, from Leibniz to David Chalmers. Yet the existence of limits to
74
Mechanisms and the mental
mechanization may be interpreted not as evidence that human minds transcend mechanism but
that they are also subject to the same limits, as the debate over limiting theorems in mathematics
shows (see Chapter 33).
In short, the development of technological mechanisms has affected the ways one conceived
of possible explanations of the mental, and mechanists tried to describe, build, and elucidate
mechanisms that exhibit at least some cognitive capacities, including mathematical reasoning and
conceptual thought. Early mechanistic models were just proofs of possibility of mechanization;
however, current mechanistic explanations strive to go beyond the question of how it is possible to
display such-and-such a mental phenomenon, and (at least partially) explain actual mental processes.
2. Early mechanists: Descartes and Hobbes to Leibniz

and mechanical ducks
René Descartes is known for his defense of mind–body dualism, but he was also a staunch pro-
ponent of mechanistic explanations, which paved the way for modern science. His understand-
ing of mechanisms was strongly influenced by the technology available in his age, in particular
fountains and clocks. A mechanism for Descartes is a spatial system whose operation depends on
local interactions among its parts. These interactions can be further described geometrically.
Notably, mechanisms can be explained without recourse to teleology, and indeed, the Cartesian
account of the mechanistic explanation is strongly opposed to the Aristotelian and medieval
models of explanation, which used to appeal to final causes, theological considerations, angelol-
ogy, basic elements, and Pythagorean principles, all at the same time.
The significance of Descartes does not stop here. He also offered early mechanistic accounts
of the nervous system and psychological phenomena (see Chapter 28). The nervous system was
conceived of in hydraulic terms of flows of animal spirits. These are produced in the blood and
may exert influence over bodily parts. The basic building block of the nervous system, accord-
ing to Descartes, is the reflex arc: the stimulus pulls tiny wires of the nervous system, which in
turn open little valves in the brain, releasing animal spirits to hollow nerve tubes that lead to
appropriate muscles (see Figure 6.1). This general outline of the reflex arc as the basic operation
of the nervous system has remained immensely important, even if subsequent research rejected
hydraulic metaphors. The modern notion of reflex itself was introduced by a medical doctor
Thomas Willis in his treaty De motu musculari (Willis 1694). Willis was influenced by Descartes,
but his mechanism was framed in optical terms (reflection being an optical phenomenon);
the reflex is a pure reaction to external influence rather than something that relies on internal
resources (as in Descartes’ model (cf. Canguilhem 1955)).
However, Descartes argued that there are limits to mechanistic explanation of the mind. He
thought that the physiological machine just described can have functions such as
the reception by the external sense organs of light, sounds, smells, tastes, heat and
other such qualities, the imprinting of the ideas of these qualities in the organ of the
“common” sense and the imagination, the retention or stamping of these ideas in the
memory, the internal movements of the appetites and passions, and finally the external
movements of all the limbs.
(Descartes 1664/1985, 1: 100)
But in Discourse on the Method, he claimed that no mechanism can be built that could “use words,
or put together other signs, as we do in order to declare our thoughts to others” (Descartes
1637/1985, 1: 140). According to him, it is not conceivable that “a machine should produce
75
Marcin Miłkowski
Figure 6.1 The reflex arc of the painful stimulus according to Descartes. The fire A is a stimulus
afflicting the skin (B) and moving the fine thread (C), which goes to the valve (d, e). The
valve opens the cavity (F), from which the animal spirit is released, and this in turn makes
the head turn and move the hand and the foot
different arrangements of words so as to give an appropriately meaningful answer to whatever is

said in its presence.” Moreover, no machine could work for a variety of purposes, and, in con-
trast, reason is a universal instrument. A human being’s rational capacities must therefore come
from a non-physical rational soul, which Descartes believed to be connected with the material
brain through the pineal gland (Descartes 1649/1985, 1: 340). The Cartesian challenge is still
alive today (Wheeler 2008).
In contrast to Descartes, Hobbes was probably the first proponent of a thoroughly mecha-
nistic and computational theory of mind; as he claims, ratiocination is just computation: “to
compute is, either to collect the sum of many things that are added together, or to know
what remains when one thing is taken out of another” (Hobbes 1655/1839, 3). However,
Hobbes failed to deliver even a sketchy description of a mechanism that could indeed think
by computing. But soon Leibniz envisaged that any philosophical argument could be simply
settled by computing in a universal mathematical language (Leibniz 1666). The seventeenth
century witnessed important developments in the construction of various machines, including
Blaise Pascal’s arithmetic machine, but none of them was fit to perform complex logical com-
putations. Further development in engineering led to even more sophisticated machines, the
Digesting Duck, created in 1739 by Jacques de Vaucanson, being probably the most promi-
nent. The mechanical duck appeared to be able to digest food; however, it only collected
food in one container and produced artificial feces from another one. Another imaginary
76
Figure 6.2 Statua Humana Circulatoria, Salomon Reisel’s imaginary model of respiration
model of human respiration, Statua Humana Circulatoria, conceived by Salomon Reisel, is

depicted in Figure 6.2. At the same time, such contraptions aroused the imagination of pro-
ponents of mechanical reductionism.
The mechanical reductionism is mostly associated with Julien de La Mettrie, who
offered a purely mechanical account of the mental in his Man a Machine (La Mettrie
1748/1912). Note, however, that La Mettrie did not espouse the Cartesian notion of the
mechanism; he claimed that the matter is capable of movement and sensibility only when
appropriately organized. To him, organization is not just a matter of spatial arrangement,
as in Descartes, but also of inner complexity that may be quantified. In this, La Mettrie
embraced a notion of mechanism closer to the one defended by New Mechanists than the
Cartesian one. At the same time, he admitted that it was an open question what kind of
organization contributed to sensation.
77
Marcin Miłkowski
3. Associationistic models of the neural system and early robotic models

The most influential criticism of the Cartesian account of thought came from John Locke,
who criticized the theory of innate ideas defended by Descartes, and claimed that all thought
came from the senses. Yet for the development of the mechanistic understanding of the men-
tal, another empiricist claim is much more important: namely, that ideas form thoughts by
association. The associationist theory of ideas has roots as deep as in Plato’s Phaedo, and early
formulations can also be found in Hobbes. But it was David Hume, having identified three
principles of association—similarity, contiguity of time and place, and cause and effect—who
had the greatest influence on subsequent theorizing about mental mechanisms. For example,
having experienced similar impressions when touching fire, a stove, and hot metal, we may
associate all of them with the notion of something hot.
Soon after, associationism was linked together with hydraulic models of the mind. For example,
Sigmund Freud’s speculations about neurological mechanisms underlying psychological phe-
nomena relied on associationist and hydraulic hypotheses (compare Bilder and LeFever 1998).
In particular, a difficult question was how new associations were formed. Associationism sug-
gested that there had to be a certain mechanism, and as late as at the turn of the twentieth
century, physical models of association learning were offered (for a comprehensive review of
such models, see Cordeschi 2002).
In the second half of the nineteenth century, the first labs in experimental psychology were
formed by Wilhelm Wundt in Germany and William James in the US. Wundt’s work had deep
roots in psychophysical research, whose champions included physicists Herman Helmholtz (who
also posited unconscious inference) and Ernst Mach. However, Wundt’s goal was to explain
psychological phenomena with lawlike generalizations, just like in psychophysics, rather than
by appeal to organized mechanisms. His experimental paradigm, related largely to introspection,
did not uncover a large body of well-confirmed laws. In opposition to Wundt’s introspective
psychology and to its American competition in the work of Titchener, the behaviorist move-
ment started. Interestingly, behaviorists, such as E. Thorndike, also tried to discover general laws
of learning that would hold true of all animals.
Behaviorism is committed to a broadly associationist view of the learning organism. Hence,
at the turn of the twentieth century, there was a large interest in understanding the underlying
principles of association. This was also driven by new discoveries in neurobiology; in particular,
Sir Charles Sherrington proposed that elementary building blocks of the nervous systems, the
neurons, communicate via synapses (see Chapter 28). At the time, Santiago Ramón y Cajal
and Camillo Golgi debated over the general anatomy of the nervous tissue. Cajal argued that
the nervous tissue is composed of multiple cells while Golgi claimed that the brain is a single
continuous network, called the reticulum. Sherrington’s proposal of the synaptic communica-
tion, related to his detailed study of reflexes, was one of the tipping points in the debate (Burke
2007). Moreover, he also refined the Cartesian notion of the reflex arc as the basic operational
blueprint of the nervous system. The notion of the reflex arc—as composed of a receptor,
conductor, and effector—was compatible with the behaviorist psychology, which focused on
stimulus-response pairs.
Such is the background for the early connectionist theory defended by Edward Thorndike
(1911). According to him, “the animal mind is, by any definition, something intimately asso-
ciated with his connection system or means of binding various physical activities to various
physical impressions” (Thorndike 1911, 16). The “connection system” underlies the capacity of
animals to learn and consists of connections between stimuli coming from the external environ-
ment and collected by a “receptor system,” and motor responses.
78
Subsequently, artificial models of learning mechanisms were proposed. For example, Bent
Russell built hydraulic networks that demonstrated effects of learning over time, supported by
experimental results of reinforcement on animals (Russell 1913). This approach was then con-
tinued by a prominent behaviorist, Clark Hull, in his robotic models of learning in the 1920s
and 1930s. Hull wanted to build robots to prove the possibility that complex forms of behavior
were not limited to living organisms. Hull’s models were no longer hydraulic but electrochemi-
cal. His model of the conditioned reflex was supposed to free “the science of complex adaptive
mammalian behavior from the mysticism whichever haunts it” (Krueger and Hull 1931, 267).
Importantly, it was not an exact replica of animal behavior but a conceptual model that proved that
a number of features of behavior are mechanically replicable. This feature of simulative models
remained important throughout the rest of the twentieth century.
However, the time was not yet ripe for simulative research and later Hull shied away from
further robotic simulations, as they were not treated seriously. For Edward Tolman, with his
hypothetical robotic “schematic sowbug,” a simplified artificial animal that illustrated his theory
that all behavior was entirely stimulus driven (Tolman 1939), robotic models were not essen-
tial in studying behavior either. Not only behaviorists built robots at the time; one particularly
important exception was Alfred Lotka (1925), known for his work on mathematical predator/
prey models. His simple mechanical animal model—a “correlating apparatus”—was an impor-
tant predecessor of later cybernetic robots for two reasons. First, Lotka approached the evolution
of complex systems in terms of reciprocal equilibrium. No longer does the nervous system work
merely as a reflex arc; correlating actions with the environment is “essentially cyclic in charac-
ter: It has its origin in the external world, which becomes depicted in the organism, provokes
a response, the terminal step of which is usually, if not always, a reaction upon the external
world” (Lotka 1925, 340). Second, his model uses the notion of information as controlling the
behavior of an animal, and his simplistic model of representation (called “depiction” by Lotka)
predates later work on cognitive representations. This notion allows talk of how the animal’s
behavior can be influenced by future events—represented thanks to the antennas mounted on
the mechanical robot to detect the edge of a table—without appealing to the notion of the final
cause (Cordeschi 2002, 128).
4. Computation and cybernetics: from Babbage and

Turing to Ashby
Another crucial development in the nineteenth century was the blueprint for a general-
purpose computational machine. Charles Babbage argued that his Difference Engine (see
Figure 6.3) could serve as a simple model of the geological laws (Babbage 1837; Dolan 1998).
This was one of the first uses of numerical computation for simulation purposes, but the
implications of the Difference Engine do not end here. It was the predecessor of the Analytic
Engine, the first general-purpose programmable machine to be designed, as Babbage him-
self and Lady Ada Lovelace observed. And a general-purpose mechanism was exactly what
Descartes thought was impossible.
It was only Alan Turing (1937) who brought the consequences of the general-purpose com-
puting machines to light. Having formalized the notion of a computing machine, called later a
Turing machine, he defined a universal machine that was able to operate like any other Turing
machine just by operating on the description of that machine on its input tape (for more, see
Chapter 33). However, universal Turing machines can only compute effectively computable
procedures, so there are limits to their capacities. Turing referred to the important work by Kurt
Gödel (1933) who proved that in the first-order predicate logic with elementary arithmetic,
79
Marcin Miłkowski
Figure 6.3 Difference Engine by Charles Babbage. Fragment of the original blueprint
there are always true statements that cannot be proved in that very logic. This later led to an
antimechanistic argument that no machine could simulate a human mathematical mind (Lucas
1961; Penrose 1989). However, these arguments need to presuppose that the human mind is
logically consistent, which also cannot be decided using an effective procedure (Putnam 1960).
(The consistency presupposition is unjustified, and the proof used by Lucas leads to a contradic-
tion (Krajewski 2007).)
Turing suggested that artificial computational machines can be thought to be generally intel-
ligent as soon as they could converse with people without being detected as machines (Turing
1950). In this conversational test of intelligence—called later the Turing test—the criterion of
thinking is what Descartes considered impossible.
Even more importantly, Turing was not alone when he thought machines could be pro-
grammed to display intelligence. It was a common assumption among British and American
80
defenders of cybernetics—a general theory of control and communication in the animal and the
machine, as Norbert Wiener defined it in the eponymous title of his book (Wiener 1948).
Cybernetics made several important contributions to the development of mechanistic models
of the mental (Pickering 2009; Johnston 2008). First, it offered a new conceptual framework
for natural teleology, i.e. for describing the goals of physical systems without recourse to any
non-physical factors (Rosenblueth, Wiener, and Bigelow 1943). The notion of feedback, fun-
damental to cybernetics, was used to elucidate the relation of the mechanism to its environment,
showing how the environment can control the behavior of the mechanism. The feedback occurs
when the output of the system is directed back to the input of the system, which effectively cre-
ates a loop. There are two kinds of feedback: negative and positive. The negative feedback occurs
when some function of the output of the system is fed into its input to stabilize the output. It
promotes stability and is frequently used for control purposes (Sluckin 1954). The positive feed-
back, on the other hand, occurs when a small disturbance in the output of the system increases
the magnitude of the disturbance. It can lead to chaotic oscillations and exponential growth.
Note that machines with feedback had been invented much earlier than the 1940s. There were
other regulators or governors before (for example, the Watt governor used to control the steam
engine), but there was no precise mathematical theory of control.
Second, cybernetics used the newly developed theory of information (Weaver and Shannon
1964). Claude Shannon defined a mathematical notion of information as uncertainty of the
receiver of the information as to what information it will receive over the noisy channel from
a sender (for introductions to the theory of information, see Floridi 2010; Burgin 2009). It has
become possible to quantify the information sent. However, the open question was how to
relate this theory to meaningful (semantic) information. Semantic information is about some-
thing, in contrast to purely structural, probabilistic, or syntactic information, which boils down
to the existence of mere regularities. A number of theorists proposed their own approaches to
semantic information (MacKay 1969; Bar-Hillel 1955; Dretske 1981).
Third, the focus on complexity of control allowed the hypothesis that the main function of
the brain is to control—in a stable way—the overall behavior of the animal. One particularly
influential notion was that of homeostasis or keeping the internal conditions of the system rela-
tively constant and stable. W. Ashby built a model of a homeostatic machine—Homeostat—that
was supposed to show how complex systems adapt to the environment (Ashby 1960). Even if
the Homeostat does not seem particularly exciting (it has no effectors to move it), it shows what
Ashby called “ultrastable behavior”: it adjusts its internal organization to react appropriately to
the outside disturbance. As such, Homeostat goes beyond simple reflex arc models and simple
animal stimulus-driven behaviors.
Fourth, cybernetics promoted interdisciplinary research and asked integrative questions
about mechanisms of control and communication in animals and machines. This promoted
building robotic models of animals, and the use of cybernetic notions in biology and psychol-
ogy. For example, W. Grey Walter built robots (Walter 1950a, 1950b, 1953)—called tortoises
for their looks—as models of animal behavior: environmentally induced exploration (in a
robot jokingly called Machina speculatrix); and conditional reflexes (in Machina docilis). These
electromechanical tortoises were milestones in the history of biorobotics (the discipline that
explains biological behavior with robots; compare Webb 2002) and bionics (the discipline that
builds biologically inspired robots). The technological obstacles to building more complex
robots could not, however, be easily surpassed, and robots gave way to computational mod-
eling in information-processing psychology. Yet Grey Walter’s research remained influential
to defenders of classical cognitive research (Miłkowski 2016), and later, in the 1980s, inspired
behavioral robotics (Brooks 1999).
81
Marcin Miłkowski
5. Computational models and artificial intelligence

Cybernetics, information theory, and the advent of digital computers made it possible to con-
sider the development of computer simulation of mental processes. While in the nineteenth and
early twentieth century, Charles Sherrington, William James, and Alfred Lotka would talk of
the telephone switchboard as the model of the brain, in the second half of the twentieth century
the preferred mechanistic model of the mental was the programmable digital computer. There
were four reasons given for the general analogy between computers and minds (Apter 1970).
First, both are general-purpose machines. Second, both process information. Third, both can
create models of reality. In other words, one of the basic functions of the brain is to build models
of the environment; this was particularly stressed by Kenneth Craik (1967). Fourth, both com-
puters and brains have a large number of similar elementary building blocks that perform simple
operations, which, taken together, contribute to complex behavior.
In 1958, Herbert Simon and Allen Newell predicted that in ten years most psychological
theories will take the form of “computer programs, or of qualitative statements about the char-
acteristics of computer programs” (Simon and Newell 1958, 7–8). Their view was extreme,
and many cognitive psychologists did not share their enthusiasm (Baars 1986; Neisser 1963).
Nevertheless, psychologists used information-theory notions to design and analyze significant
experimental results (Sperling 1960; Miller 1956; Broadbent 1958). Even in 1967, almost ten
years after Simon and Newell made their initial prediction, psychological references to com-
puter programs were highly metaphorical and lacked experimental evidence (Miller, Galanter,
and Pribram 1967). In the 2000s, however, the prediction holds true: over 80 percent of articles
in theoretical journals focus on computational models (Busemeyer and Diederich 2010).
What made the prediction true? One factor was the new interdisciplinary field formed at the
end of the 1970s that came to be called cognitive science (Arbib et al. 1978). It may be a stretch
to claim that it was a scientific revolution in the strict sense (Kuhn 1970), as some initially
claimed (Gardner 1985): first, it does not seem incommensurable with previous psychological
research (for example, Tolman’s (1948) claim that rats use cognitive maps to navigate mazes
is easily translatable into cognitive terminology); and second, it’s deeply connected with mul-
tiple disciplines that already coalesced around cybernetics. Cognitive science made significant
contributions to mechanistic models of the mental even if some would deny that all cognitive
models are mechanistic.
In the 1950s, another important field was formed: Artificial Intelligence (AI). In AI, two
basic approaches were dominant: connectionist machine learning, initially in the Perceptron
machine (Rosenblatt 1958), and the rule-based approach, defended by Simon and Newell. The
Perceptron was an electromechanical, artificial neural network (ANN) with extremely simplified
neurons. It could learn by the so-called delta rule that allowed researchers to adjust the weights
of connections between neurons. However, it was soon discovered that using the delta rule,
Perceptrons could not learn elementary logical operations such as exclusive disjunction, or XOR
(Minsky and Papert 1969). This has led ANN research to stall until the 1980s. In their place,
rule-based approaches dominated, but they also failed to quickly deliver the promised results.
These two approaches have their counterparts in cognitive science. The rule-based approach,
also called the symbolic approach because of its reliance on a particular conception of cognitive
representation, was the mainstream approach until the early 1980s. For example, Allen Newell
and Herbert Simon (1972) simulated problem-solving processes with rule-based systems, which
were then evaluated by comparing the simulation with the behavioral data (such as verbal
reports of subjects and eye-tracking data). Newell and Simon stressed that their methodology
was ill-suited for motor and perceptual processing.
82
The connectionist approach has its roots in associationism, and it became much more
sophisticated in the 1980s when new methods of ANN learning appeared. Notably, the
backpropagation algorithm allowed the training of networks with multiple layers of hidden
neurons, and the ANNs were used to simulate parallel cognitive processing. Significant pro-
gress has been made in transcending the limitations of associationism: the association relation
is symmetrical while many cognitive classical structures are ordered (McClelland, Rumelhart,
and PDP Research Group 1986; Rumelhart and McClelland 1986; for a general assessment,
see Bechtel and Abrahamsen 2002). However, defenders of the symbolic approach claimed
that ANNs either do not offer a full account of cognitive structures, with such features as
compositionality or systematicity, or they are mere implementations of symbolic approaches
(Fodor and Pylyshyn 1988).
All this criticism notwithstanding, ANN modeling has matured significantly. While the early
ANNs were extremely simplified and idealized, the current models use much more biologi-
cally plausible models of neurons, sometimes with a large variety of neuron types. ANNs are
also routinely described using control theory, which allows using dynamical systems theory to
design their architecture (Eliasmith and Anderson 2003). They are also used in computational
neuroscience, for example to build large brain models (Eliasmith 2013). The large models strive
for biological plausibility (in the mechanistic terminology, they are how-plausibly models; see
Craver 2007), and while they are heavily idealized, they are able to offer novel predictions.
In other words, they are no longer just proofs of possibility that a mechanism implementing a
mental capacity may be built, but idealized models of how actual brains may work.
Because of the heavy stress on computational models, general objections against cognitive
science were raised. The first one is related to the notion of mental representation as used in
cognitive research. In most explanatory models, it was assumed that it’s sufficient to simulate the
syntactic structure of representations for the semantic properties to appear; as John Haugeland
quipped: “if you take care of syntax, the semantics will take care of itself” (Haugeland 1985,
106). John Searle raised an important objection against this formalist principle in his Chinese
Room thought experiment (Searle 1980): one can easily imagine a person with a special set of
instructions in English who could manipulate Chinese symbols and answer questions in Chinese
without understanding it at all. Hence, understanding is not reducible to syntactic manipulation.
While the discussion around this thought experiment is hardly conclusive (Preston and Bishop
2002), the problem was soon reformulated by Stevan Harnad as the “symbol grounding prob-
lem” (Harnad 1990). How can symbols in computational machines mean anything? Or, to spell
this question out a little more generally, what mechanism is able to mean anything?
Formulated this way, Searle’s question is yet another version of the mind–body problem
as related to intentionality, or aboutness of mental processes or states. In the nineteenth cen-
tury, Franz Brentano claimed that intentionality distinguishes the mental from the physical
(Brentano 1900; Chrudzimski 1999). In this, he followed Descartes: no mechanism could enter-
tain thoughts. However, with the advent of digital computers, a new argument was put forward:
thought and content are causally relevant in the physical world just because syntactic entities are
computed over. Searle argues that syntax is not sufficient for intentionality, and in this, most
theorists of mental representation agree. The most influential position, teleosemantics, defended
by Ruth Garett Millikan (1984) and Fred Dretske (1997), is that to represent is, roughly, to
have the biological function of making information available to cognitive and motor processes.
This position remains controversial (Ramsey 2007; Hutto and Myin 2013; Fodor and Pylyshyn
2015), especially for the followers of the cybernetic approach to cognition in terms of control.
They claim that the appeal to (cognitive) representation in control mechanisms plays no particu-
lar explanatory role (Van Gelder 1995; Chemero 2009), and that computational explanations
83
Marcin Miłkowski
may be replaced by dynamical explanations (essentially similar to the ones offered by Ashby in
his account of Homeostat). However, most contemporary criticisms of computational cognitive
science focus on two other properties of mental processes: qualities of experience and the ability
to use common sense.
It was argued that no computational mechanism can produce the qualitative states of con-
sciousness (called qualia) and that consciousness is not physical (Jackson 1986; Chalmers 1996).
These arguments are usually put forward in the following form: no amount of physical knowledge
(for example, about computational structures of the brain) is sufficient to know the qualitative
consciousness. The two most frequent replies to such arguments are the following: first, it is
claimed that qualitative knowledge of conscious states is not knowledge-that but knowledge-
how (Ryle 2009), which is in principle not available intersubjectively (Churchland 1985); and
second, it is argued that it’s just the lack of imagination of philosophers trying to understand the
epistemic situation of a person with all the relevant physical knowledge that creates the illusion
that there is something to qualia that is beyond all physical description (Dennett 2005).
The last important argument against computational cognitive science is that it cannot explain
how people smoothly cope with their everyday tasks, or to understand their common sense
(Dreyfus 1979). Indeed, it is still a Holy Grail of artificial intelligence to build a machine that
can make commonsensical inferences and act accordingly. The defenders of AI claim that the
actual scope of common-sense knowledge is vaster than critics of AI imagine.
6. Conclusion
The history of mechanistic models of the mental reflects theorizing about mechanisms as such.
First, mechanisms were considered to be entities in space, whose changes could be described
geometrically, and whose interactions could be only local. Then, they were considered causal
and temporal as well. La Mettrie subsequently stressed that mechanisms are organized systems.
At the same time, the understanding of the mental in terms of mechanisms of information
processing progressed. Descartes proposed a general outline of the nervous system in terms
of a reflex arc but considered important mental faculties to be exempt from the purview of
mechanism. He supposed that there could be no mechanism for thinking and drawing infer-
ences. Soon, however, Leibniz sketched a proposal of a universal, logical, and computational
calculus of thought. As soon as computing mechanisms were built, it was suggested that they
could be used to build not only rough approximations of the mind, but also precise causal
models, including also cyclical interactions with the environment. Subsequently, the gap
between mathematical, computational simulations of the mental and the neurobiologically
plausible models began to close.
The role of mechanistic models has also changed. Here, it is useful to distinguish mechanistic
simulations from theoretical descriptions of mechanisms. Until the second half of the twentieth
century, mechanistic simulations, usually owing to limits of technology, were mere proofs of
possibility that the mind might be, in spite of appearances, mechanical. Some modelers were
overly enthusiastic about the capacities of their artifacts; for example, Walter would claim that
his tortoises were capable of self-recognition, while they simply responded to light sources, and
just because they had lamps mounted on them, they would react to their own image in the
mirror. Theoretical descriptions of mechanisms were not bounded by technological limitations,
and they could offer some insight into how the mental processes actually occur. Hence, already
the model of the reflex arc was not just a mere how-possibly model that would prove that the
mechanical mind is possible (and Descartes would even retort that the mechanical mind is a con-
tradiction in terms). It was a how-plausibly model of the nervous system (not the mind) whose
84
role was to explain at least a large number of muscle movements. Only in the second half of the
twentieth century have mechanistic simulations been built for similar explanatory purposes. Yet
technological and theoretical limitations have not all been overcome, and they are still far from
complete models of how actual mechanisms work. But the underlying assumption of model-
ers in cognitive neuroscience today is that such models not only can but also should be built
(Piccinini and Craver 2011; Miłkowski 2013; Boone and Piccinini 2016).
Different technological artifacts have been treated as models of the mental, from a wax tablet,
through clocks and the automated telephone switchboard, to a digital computer. However, as
Margaret Boden stresses, the computer is not just another metaphor of the mind; it might also
be the last metaphor because it may offer a precise causal model of the mind (Boden 2008).
Most mechanistic models of the mental are no longer controversial. Even self-proclaimed dual-
ists rarely believe that it is impossible to understand intentionality mechanistically. The only
remaining bastion of anti-mechanists is qualitative consciousness, although in recent years, some
progress has been made in terms of measuring consciousness (Seth et al. 2008) and building
more plausible causal models of its processes (Boly et al. 2013). We still do not know how to
model consciousness mechanistically, but it is no longer obvious that we will never know.
References
Apter, Michael. 1970. The Computer Simulation of Behaviour. London: Hutchinson.
Arbib, Michael A., Carl Lee Baker, Joan Bresnan, Roy G. D’Andrade, Ronald Kaplan, Samuel Jay Keyser,
Donald A. Norman, et al. 1978. Cognitive Science, 1978: Report of The State of the Art Committee to
The Advisors of The Alfred P. Sloan Foundation. New York: The Alfred P. Sloan Foundation. http://
csjarchive.cogsci.rpi.edu/misc/CognitiveScience1978_OCR.pdf.
Ashby, W Ross. 1960. Design for a Brain: The Origin of Adaptive Behavior. New York: Wiley.
Baars, Bernard J. 1986. The Cognitive Revolution in Psychology. New York: Guilford Press.
Babbage, Charles. 1837. The Ninth Bridgewater Treatise. A Fragment. London: J. Murray.
Bar-Hillel, Yehoshua. 1955. “Information and Content: A Semantic Analysis.” Synthese 9 (1): 299–305.
doi:10.1007/BF00567416.
Bechtel, William and Adele Abrahamsen. 2002. Connectionism and the Mind. Oxford: Blackwell.
Bilder, Robert and F. Frank LeFever, eds. 1998. Neuroscience of the Mind on the Centennial of Freud’s Project
for a Scientific Psychology. New York: New York Academy of Sciences.
Boden, Margaret A. 2008. “Information, Computation, and Cognitive Science.” In Philosophy of
Information: Volume 8, edited by Pieter Adriaans and Johan Van Benthem, 8: 741–61. Elsevier B.V.
doi:10.1016/B978-0-444-51726-5.50023-6.
Boly, Melanie, Anil K. Seth, Melanie Wilke, Paul Ingmundson, Bernard J. Baars, Steven Laureys, David
Edelman, and Naotsugu Tsuchiya. 2013. “Consciousness in Humans and Non-Human Animals: Recent
Advances and Future Directions.” Frontiers in Psychology 4. Frontiers. doi:10.3389/fpsyg.2013.00625.
Boone, Worth and Gualtiero Piccinini. 2016. “The Cognitive Neuroscience Revolution.” Synthese, June
193 (3): 1509–1534. Springer Netherlands. doi:10.1007/s11229-015-0783-4.
Brentano, Franz. 1900. Psychologie Vom Empirischen Standpunkt. Hamburg: F. Meiner.
Broadbent, D. E. 1958. Perception and Communication. Oxford: Pergamon Press.
Brooks, Rodney A. 1999. Cambrian Intelligence: The Early History of the New AI. Cambridge, MA: The
MIT Press.
Burgin, Mark. 2009. Theory of Information: Fundamentality, Diversity and Unification. Singapore: World
Scientific Publishing Co.
Burke, Robert E. 2007. “Sir Charles Sherrington’s the Integrative Action of the Nervous System: A
Centenary Appreciation.” Brain: A Journal of Neurology 130 (Pt 4): 887–94. doi:10.1093/brain/awm022.
Busemeyer, Jerome R. and Adele Diederich. 2010. Cognitive Modeling. Los Angeles: Sage.
Canguilhem, Georges. 1955. La Formation Du Concept de Réflexe Aux XVIIe et XVIIIe Siècles. Paris: Presses
universitaires de France.
Chalmers, David J. 1996. The Conscious Mind: In Search of a Fundamental Theory. New York: Oxford
University Press.
Chemero, Anthony. 2009. Radical Embodied Cognitive Science. Cambridge, MA: The MIT Press.
85
Marcin Miłkowski
Chrudzimski, Arkadiusz. 1999. “Die Theorie Der Intentionalität Bei Franz Brentano.” Grazer Philosophische
Studien 57 (July): 45–66. doi:10.5840/gps1999574.
Churchland, Paul M. 1985. “Reduction, Qualia, and the Direct Introspection of Brain States.” The Journal
of Philosophy 82 (1): 8–28.
Cordeschi, Roberto. 2002. The Discovery of the Artificial. Studies in Cognitive Systems. Dordrecht: Springer
Netherlands. doi:10.1007/978-94-015-9870-5.
Craik, Kenneth. 1967. The Nature of Explanation. Cambridge: Cambridge University Press.
Craver, Carl F. 2007. Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience. Oxford: Oxford
University Press.
Dennett, Daniel C. 2005. Sweet Dreams: Philosophical Obstacles to a Science of Consciousness. Cambridge, MA:
The MIT Press.
Descartes, René. 1985. The Philosophical Writings of Descartes. Edited by John Cottingham, Robert
Stoothoff, and Dugald Murdoch. Vol. 1. Cambridge: Cambridge University Press.
Dolan, Brian P. 1998. “Representing Novelty: Charles Babbage, Charles Lyell, and Experiments in Early
Victorian Geology.” History of Science 36: xxxvi.
Dretske, Fred I. 1981. Knowledge and the Flow of Information. 2nd ed. Cambridge, MA: The MIT Press.
——. 1997. Naturalizing the Mind. Cambridge, MA: The MIT Press.
Dreyfus, Hubert. 1979. What Computers Still Can’t Do: A Critique of Artificial Reason. Cambridge, MA:
The MIT Press.
Eliasmith, Chris. 2013. How to Build the Brain: A Neural Architecture for Biological Cognition. New York:
Eliasmith, Chris and Charles H. Anderson. 2003. Neural Engineering: Computation, Representation, and
Dynamics in Neurobiological Systems. Cambridge, MA: The MIT Press.
Floridi, Luciano. 2010. Information: A Very Short Introduction. Oxford: Oxford University Press.
Fodor, Jerry A. and Zenon W. Pylyshyn. 1988. “Connectionism and Cognitive Architecture: A Critical
Analysis.” Cognition 28 (1–2): 3–71.
——. 2015. Minds without Meanings: An Essay on the Content of Concepts. Cambridge, MA: The MIT Press.
Gardner, Howard. 1985. The Mind’s New Science: A History of the Cognitive Revolution. New York: Basic
Books.
Gödel, Kurt. 1933. “Zum Entscheidungsproblem Des Logischen Funktionenkalküls.” Monatshefte Für
Mathematik Und Physik 40 (1): 433–43. doi:10.1007/BF01708881.
Harnad, Stevan. 1990. “The Symbol Grounding Problem.” Physica D 42: 335–46.
Haugeland, John. 1985. Artificial Intelligence: The Very Idea. Cambridge, MA: The MIT Press.
Hobbes, Thomas. 1839. The English Works of Thomas Hobbes of Malmesbury. London: J. Bohn.
Hutto, Daniel D. and Erik Myin. 2013. Radicalizing Enactivism: Basic Minds without Content. Cambridge,
MA: The MIT Press.
Jackson, Frank. 1986. “What Mary Didn’t Know.” The Journal of Philosophy 83 (5): 291–5.
doi:10.2307/2026143.
Johnston, John. 2008. The Allure of Machinic Life: Cybernetics, Artificial Life, and the New AI. Cambridge,
MA: The MIT Press.
Krajewski, Stanisław. 2007. “On Gödel’s Theorem and Mechanism: Inconsistency or Unsoundness Is
Unavoidable in Any Attempt to ‘Out-Gödel’ the Mechanist.” Fundamenta Informaticae 81 (1): 173–81.
Krueger, Robert C. and Clark L. Hull. 1931. “An Electro-Chemical Parallel to the Conditioned Reflex.”
Journal of General Psychology 5: 262–9.
Kuhn, Thomas S. 1970. The Structure of Scientific Revolutions. Chicago: University of Chicago Press.
La Mettrie, Julien. 1912. Man a Machine. Translated by Gertrude Carman Bussey and Mary Whiton
Calkins. Chicago: The Open Court Pub. Co.
Leibniz, Gottfried. 1666. Dissertatio de Arte Combinatoria, in qua Ex Arithmeticae Fundamentis Complicationum
Ac Transpositionum Doctrina Novis Praeceptis Extruitur, & Usus Ambarum per Universum Scientiarum Orbem
Ostenditur. Lipsiae: apud Joh. Simon Fickium et Joh. Polycarp. Seuboldum Literis Spörelianis.
Lotka, Alfred. 1925. Elements of Physical Biology. Baltimore: Williams & Wilkins Company.
Lucas, J. R. 1961. “Minds, Machines and Gödel.” Philosophy 9 (3): 219–27.
MacKay, Donald MacCrimmon. 1969. Information, Mechanism and Meaning. Cambridge, MA: The MIT
Press.
McClelland, James L., David E. Rumelhart, and PDP Research Group, eds. 1986. Parallel Distributed
Processing: Explorations in the Microstructures of Cognition, Volume 2: Psychological and Biological Models.
Cambridge, MA: The MIT Press.
86
Miller, George A. 1956. “The Magical Number Seven, Plus or Minus Two: Some Limits on Our Capacity
for Processing Information.” Psychological Review 63 (2): 81–97. doi:10.1037/h0043158.
Miller, George A., Eugene Galanter, and Karl H. Pribram. 1967. Plans and the Structure of Behavior. New
York: Holt.
Millikan, Ruth Garrett. 1984. Language, Thought, and Other Biological Categories: New Foundations for Realism.
Miłkowski, Marcin. 2013. Explaining the Computational Mind. Cambridge, MA: The MIT Press.
——. 2016. “Models of Environment.” In Minds, Models and Milieux: Commemorating the Centennial of
the Birth of Herbert Simon, edited by Roger Frantz and Leslie Marsh, 227–38. New York: Palgrave
Macmillan. doi:10.1057/9781137442505_13.
Minsky, Marvin and Seymour Papert. 1969. Perceptrons: An Introduction to Computational Geometry.
Neisser, U. 1963. “The Imitation of Man by Machine: The View That Machines Will Think as Man Does
Reveals Misunderstanding of the Nature of Human Thought.” Science 139 (3551): 193–7. doi:10.1126/
science.139.3551.193.
Newell, Allen and Herbert A. Simon. 1972. Human Problem Solving. Englewood Cliffs, NJ: Prentice-Hall.
Penrose, Roger. 1989. The Emperor’s New Mind. London: Oxford University Press.
Piccinini, Gualtiero and Carl F. Craver. 2011. “Integrating Psychology and Neuroscience: Functional
Analyses as Mechanism Sketches.” Synthese 183 (3): 283–311. doi:10.1007/s11229-011-9898-4.
Pickering, Andrew. 2009. The Cybernetic Brain: Sketches of Another Future. Chicago: University of Chicago Press.
Preston, John and Mark Bishop. 2002. Views into the Chinese Room: New Essays on Searle and Artificial
Intelligence. Oxford: Clarendon Press.
Putnam, Hilary. 1960. “Minds and Machines.” In Dimensions of Mind, edited by Sidney Hook, 148–79.
New York: New York University Press.
Ramsey, William M. 2007. Representation Reconsidered. Cambridge: Cambridge University Press.
doi:10.1017/CBO9780511597954.
Rosenblatt, F. 1958. “The Perceptron: A Probabilistic Model for Information Storage and Organization in
the Brain.” Psychological Review 65 (6): 386–408. doi:10.1037/h0042519.
Rosenblueth, Arturo, Norbert Wiener, and Julian Bigelow. 1943. “Behavior, Purpose and Teleology.”
Philosophy of Science 10 (1): 18. doi:10.1086/286788.
Rumelhart, D. E. and James L. McClelland. 1986. Parallel Distributed Processing: Explorations in the
Microstructure of Cognition. Volume 1. Foundations. Cambridge, MA: MIT Press.
Russell, S. Bent. 1913. “A Practical Device to Stimulate the Working of Nervous Discharges.” Journal of
Animal Behavior 3 (1): 15–35. doi:10.1037/h0070584.
Ryle, Gilbert. 2009. The Concept of Mind. New York: Routledge.
Searle, John R. 1980. “Minds, Brains, and Programs.” Behavioral and Brain Sciences 3 (3): 1–19. doi:10.1017/
S0140525X00005756.
Seth, Anil K., Zoltán Dienes, Axel Cleeremans, Morten Overgaard, and Luiz Pessoa. 2008. “Measuring
Consciousness: Relating Behavioural and Neurophysiological Approaches.” Trends in Cognitive Sciences
12 (8): 314–21. doi:10.1016/j.tics.2008.04.008.
Simon, Herbert A. and Allen Newell. 1958. “Heuristic Problem Solving: The Next Advance in Operations
Research.” Operations Research 6 (1): 1–10.
Sluckin, Wladyslaw. 1954. Minds and Machines. Harmondsworth: Penguin.
Sperling, George. 1960. “The Information Available in Brief Visual Presentations.” Psychological Monographs:
General and Applied 74 (11): 1–29. doi:10.1037/h0093759.
Thorndike, Edward L. 1911. Animal Intelligence. London: The Macmillan Company.
Tolman, Edward Chace. 1939. “Prediction of Vicarious Trial and Error by Means of the Schematic
Sowbug.” Psychological Review 46 (4): 318–36. doi:10.1037/h0057054.
——. 1948. “Cognitive Maps in Rats and Men.” Psychological Review 55 (4): 189–208.
Turing, Alan. 1937. “On Computable Numbers, with an Application to the Entscheidungsproblem.”
Proceedings of the London Mathematical Society s2-42 (1): 230–65. doi:10.1112/plms/s2-42.1.230.
——. 1950. “Computing Machinery and Intelligence.” Mind LIX (236): 433–60. doi:10.1093/mind/
LIX.236.433.
Van Gelder, T. 1995. “What Might Cognition Be, If Not Computation?” The Journal of Philosophy 92 (7): 345–81.
Walter, W. Grey. 1950a. “An Imitation of Life.” Scientific American. doi:10.1038/scientificamerican0550-42.
——. 1950b. “An Electro-Mechanical Animal.” Dialectica 4 (3): 206–13. doi:10.1111/j.1746-8361.1950.
tb01020.x.
87
Marcin Miłkowski
——. 1953. The Living Brain. New York: Norton.

Weaver, W. and C. E. Shannon. 1964. The Mathematical Theory of Communication. Urbana: University of
Illinois Press.
Webb, Barbara. 2002. “Can Robots Make Good Models of Biological Behaviour?” Behavioral and Brain
Sciences 24 (6): 1033–94. doi:10.1017/S0140525X01000127.
Wheeler, Michael. 2008. “God’s Machines: Descartes on the Mechanization of Mind.” In The Mechanical
Mind in History, edited by Phil Husbands, Owen Holland, and Michael Wheeler, 307–30. Cambridge,
MA: The MIT Press.
Wiener, Norbert. 1948. Cybernetics, or Control and Communication in the Animal and the Machine. New York/
Paris: J. Wiley/Hermann.
Willis, Thomas. 1694. Opera Omnia, Nitidius Quam Unquam Hactenus Edita, Plurimum Emendata. Coloniae:
Sumptibus Gasparis Storti.
88
PART II
The nature of mechanisms

7
VARIETIES OF MECHANISMS
1. Why explore mechanistic variety?

A common worry about the utility of philosophical analyses of mechanism is that the term
“mechanism” is applied to such a staggeringly diverse collection of things that nothing
informative can be said about it as a general category. This is especially true for those like us
(Glennan forthcoming; see also Chapter 1, Illari and Williamson 2012) who have argued for
a conception of mechanism that is expansive enough to allow most anything scientists have
called mechanisms to fall under its extension. In response to such worries, we have argued
that, while the things scientists (and philosophers) call mechanisms are heterogeneous, there
are enough commonalities to allow for an informative analysis of mechanisms. Such an analy-
sis can show what mechanisms are, and distinguish them from things that are not mechanisms.
Even if this analysis is successful, it is important for advocates of expansive conceptions to explore
the varieties of mechanisms. The class of mechanisms is a heterogeneous lot, and exploring the
nature and scope of these variations will give us insight into both metaphysical and methodological
questions about mechanisms. There are (at least) two ways we might explore this variety. One is to
look at exemplars of mechanisms that are studied in various scientific domains. The chapters in Part
IV collectively exemplify this approach, and they do much to illustrate the special techniques, chal-
lenges, and opportunities for mechanistic research within those domains. In this chapter, though,
we will take a different approach. Our goal will be to describe a set of taxonomic dimensions along
which we can characterize the varieties of mechanisms that exist within the natural and social world.
This system of classification will often cut across disciplinary and domain boundaries.
As we explore the varieties of mechanisms, we need to contrast our project with another
related project. Our goal in this chapter is focused on the varieties of mechanisms that exist in
the world, rather than the variety of philosophical or methodological approaches that have gone
by the name of “Mechanism” (or “mechanicism” or “mechanical philosophy”).1 There are
certainly varieties of mechanical philosophies just as there are varieties of mechanisms, and some
philosophers have offered taxonomies of those varieties (see also Chapters 3 and 5 for histori-
cal discussions). It is, moreover, easy to slip between varieties of mechanisms and varieties of
mechanical philosophies, because different kinds of philosophical approaches will often assume
or privilege certain varieties of mechanisms. We shall return to the issue of the relationship
between mechanisms and mechanical philosophies in the conclusion of our chapter.
91
2. The dimensions of mechanistic variety

Any attempt to provide a scheme for classifying things must begin with a specification of the
set of things to be classified. For us, that specification is embodied in minimal mechanism
(see Chapter 1):
We start with this permissive characterization, not because it is better or “more correct” than
less permissive ones, but because its permissiveness allows us to treat more restrictive concep-
tions as varieties of mechanisms.
How might one fruitfully sort out the vast array of things that count as mechanisms under
the minimal definition? Our proposal is that a set of classificatory dimensions can naturally be
read off the characterization of minimal mechanism.2 Specifically, a mechanism can be classified
according to (1) the kind of phenomena for which it is responsible, (2) the kinds of entities,
activities, and interactions that comprise the mechanism, and (3) the way in which these entities,
activities, and interactions are organized. To these three dimensions, we add another: (4) the
mechanism’s etiology, i.e. the way it came to exist and have the properties it does. A mecha-
nism’s etiology is logically independent of its current constitution and properties, but there are
interesting historical and causal dependencies between how a mechanism came to be, what it is
made of, how it is organized, and what it does.
Within each of these four dimensions, there is great variety, and we cannot hope in this short
piece to do more than offer a few examples of that variety, but it should be enough to illustrate
our main point—namely that attending to the varieties of mechanisms will help illuminate both
metaphysical and methodological debates. In identifying these four dimensions as dimensions,
we are suggesting that they are to a large extent independent. For instance, mechanisms with
very different kinds of entities, activities, and interactions can have similar kinds of organization,
or mechanisms with very different etiologies can be responsible for similar kinds of phenomena.
On the other hand, our decision to treat entities and activities as lying within a single dimension
reflects our understanding that entities or parts have a comparatively narrow range of activities
and interactions in which they can engage. In this and other ways, the independence of these
dimensions is not complete, and indeed understanding the ways in which, e.g., particular kinds
of organization are characteristic of mechanisms with certain sorts of parts is crucial to under-
standing the taxonomic space.
3. Varieties of phenomena
It is a truism among new mechanists that mechanisms are individuated by their phenomena or
behavior. All mechanisms are mechanisms for. This usage is common in scientific and techno-
logical contexts. For instance, one speaks of mechanisms for protein synthesis, thermal regula-
tion, or signal amplification. Mechanisms are classified as mechanisms of a certain kind if they
produce the same kind of phenomena, so for instance, artificial and natural hearts are both hearts
in the sense of being mechanisms for pumping blood within a circulatory system, even if the
material constitution and internal operations are quite different.
There is no non-pragmatic standard by which to determine if two mechanisms are respon-
sible for the same kind of phenomena. If two factories produce cars, one Fords and the other
Volvos, or one sedans and the other SUVs, do they produce the same kind of thing or different?
92
Varieties of mechanisms
Obviously there is no proper answer to this question apart from some criteria of sameness, and
such criteria will depend upon one’s interests. But while we cannot offer an absolute standard
for type-identity of phenomena, we can say much about structural features of phenomena by
which we might classify them. These will show us that very disparate phenomena may be, in
some respects, of similar kinds, and these similarities will bear on a variety of methodological
and metaphysical issues.
Consider, for example, the question of the regularity of the mechanism’s phenomenon (see
Chapter 12). Machamer, Darden, and Craver (MDC), in their much-quoted definition of
mechanism (Machamer et al. 2000), argued that mechanisms produce regular changes from start
or set-up conditions to finish or termination conditions. A paradigm of a mechanism of this kind
is protein synthesis. Given appropriate start conditions (presence of appropriate substrates, pro-
moter regions, etc.), transcription of DNA will be initiated, leading through a sequence of steps
to synthesis of proteins in the ribosome (the termination condition). Such a process is regular
in the sense that it happens many times in the life of a cell, but it is far from universal—because
various things can interfere with protein production. For instance, gene regulation can inhibit
production of certain proteins, or the cellular machinery can become damaged.
The regularity of a mechanism’s behavior provides an important way of classifying mech-
anisms by their phenomena. While early characterizations, like MDC, impose a regularity
requirement, our approach is to see that different varieties of mechanisms are different in the
degrees and respects of their regularity.
One sense in which mechanisms can be regular is that the phenomena they are responsible
for can recur. Recurrence comes in two basic varieties. The same token mechanism can exhibit
the phenomenon on multiple occasions, and the phenomenon can be exhibited by multiple
tokens of the same type of mechanism. Protein synthesis recurs in both of these ways. The
same cellular machinery continually produces proteins over its lifetime, and there are many cells
producing the same proteins. In other cases, you can have one kind of recurrence without the
other. Apoptosis is a kind of mechanism that is responsible for programmed cell death. Many
cells die this way, but each cell dies only once. On the other hand, token mechanisms, like the
Old Faithful Geyser in Yellowstone National Park, can repeatedly exhibit the same kind of
phenomena. Among mechanisms with recurrent phenomena, we can also distinguish between
the more or less regular. For instance, we can characterize mechanisms by the probability with
which start-up conditions actually lead to termination conditions (see Chapter 13).
Issues about regularity and recurrence are important in both methodological and metaphysi-
cal debates. For instance, one of the epistemological arguments for a regularity requirement
on mechanisms is that if mechanisms are not in certain ways regular, then we will not be able
to discover them. Experimental techniques in the sciences depend upon the fact that the same
phenomena recur across many tokens of the same kind of mechanism. Because of this, for
instance, one can study the mechanism of protein synthesis in a few cells of some species of
model organism and apply those conclusions to protein synthesis in cells of organisms in a wide
variety of taxa. Arguably, the distinction between the historical and experimental sciences is that
the experimental, much more than the historical, study recurrent phenomena. Historians (of
both natural and human history) are often concerned with the explanation of singular events,
and these events are often produced by mechanisms that are non-recurrent, or “one-off.” For
instance, natural historians may be interested in the causes of the Cambrian explosion, while
historians of human civilization will be interested in the causes of distinctive events in human
history, like the fall of the Roman Empire or the stock market crash in 1929. Investigation
of these “ephemeral mechanisms” require distinctive methodologies (Glennan 2010; see also
Chapter 31, this volume).
93
Regularity and recurrence also play roles in metaphysical debates about the relationship
between mechanisms, causes, and laws. Recurrent mechanisms are the foundation of the new
mechanist’s account of the nature of laws. Laws in the higher-level sciences, if such exist, are
thought to be effects—descriptions of the behavior of recurrent and regular mechanisms. At the
same time, the idea of the existence of one-off mechanisms is essential if one thinks, as Glennan
does, that causal claims are claims about the existence of mechanisms. This is required because
not all causally connected events are produced by mechanisms that are regular or recurrent.
Another useful way to classify mechanistic phenomena is by the kind of relationship a phe-
nomenon bears to the entities, activities, and interactions that are responsible for it. Craver and
Darden (2013) have suggested that sometimes mechanisms underlie phenomena, whereas other
times they produce them. When a mechanism underlies some phenomenon, the mechanism’s
phenomenon is constituted by the collective activities and interactions of its parts, and for this
reason such phenomena are called constitutive. In contrast, when a mechanism produces some
phenomenon, that phenomenon is a (causal) result of the activities and interactions of the
mechanism’s parts, rather than being made of them. For this reason, we call such phenomena
non-constitutive. The contrast between these two kinds of phenomena can be seen by comparing
protein synthesis to the mechanism responsible for muscle contraction. Protein synthesis can
be seen as a paradigm of a productive, non-constitutive mechanism, because the proteins are
made by, but not made of, the organized activities and interactions of the parts of the mecha-
nism. Muscle contraction seems quite different. Muscles consist of cells called myocytes, which
contain within them bundles of fibrils, which can shorten via the sliding of filaments made of
proteins. When a muscle contracts, it does this because the cells of which its tissues are made
contract, and this in turn happens because of the sliding filaments within the fibrils within the
cell. Here it seems that the result of the mechanism is not some product made by the mecha-
nism; the contracting of the muscle just is the contracting of its cells and the fibrils within them.
This “just is” relation is a constitutive one.
Attending to the distinction between these kinds of phenomena can be helpful in sorting
out a variety of metaphysical and epistemological issues. Metaphysical debates about the nature
of causation, production and difference making, and causal chains are concerned chiefly with
the “horizontal” production of non-constitutive phenomenon (Bogen 2004; Campaner 2013;
Glennan forthcoming, Chapter 7; see also Chapters 10, 11, and 12, this volume; Kincaid 2012;
Salmon 1984). On the other hand, mechanisms for constitutive phenomena have been at the
center of discussions of “vertical” relations between parts and wholes in science (Craver 2007;
Gillett and Aizawa 2016; see also Chapters 9 and 14, this volume; Leuridan 2012; Winther
2009), which offer a new perspective on both metaphysical and methodological versions of
debates about reduction and emergence. There has also been discussion of experimental methods
for identifying working parts of constitutive mechanisms (Craver and Darden 2013) and of the
challenges of integrating accounts at these different levels of mechanisms (Stinson 2016; see also
Chapter 28, this volume).
4. Varieties of entities, activities, and interactions

Perhaps the most obvious way to classify kinds of mechanisms is by the constituents of which
they are made. According to minimal mechanism, these constituents are of two kinds—first the
entities that are working parts of the mechanism, and second the activities and interactions in
which these entities engage. The varieties of entities in mechanisms are diverse, ranging from
DNA and mRNA in protein synthesis, through people and institutions such as Central Banks
94
in economic mechanisms, to electrons and star cores in astrophysics. The same mechanism can
include entities of notably different sizes. For instance, it appears that the mechanism responsible
for core collapse in supernovae involves interactions between parts that range in size and mass
between the subatomic and stellar cores with masses greater than our sun (Illari and Williamson
2012). Relatedly, the mechanisms responsible for some phenomenon may include entities of
quite different kinds. For instance, the mechanisms that account for the diversity and abundance
of flora in urban landscapes will involve artifacts in the built environment (like buildings and
bridges) and the social and physical actors that build them, as well as human and non-human
fauna in the area, the nutrients within the soil, and the sun and rainwater that sustains growth.
Just as mechanisms can be classified by the kinds of entities of which they are made, so too
can they be classified by the kinds of activities and interactions in which the parts of mecha-
nisms engage. It is one of the important insights of the new mechanist literature that concepts
like “activity,” “interaction,” and “cause” are abstract. Proper characterization of mechanisms
requires one to use more specific concepts like “pushing,” “folding,” “binding,” “trading,”
and “collapsing” (Bogen 2008; Machamer et al. 2000). These specific activity concepts tell one
much more about how and under what conditions mechanisms bring about their phenomena.
It is not enough, though, to distinguish between the abstract category of activities and inter-
actions, and specific sorts of activities and interactions. Rather, activities and interactions can
be characterized in a hierarchy of increasingly less abstract ways, corresponding to increasingly
concrete and determinate varieties of activities and interactions. For example, many activities
and interactions are involved in protein synthesis, but each one is an activity or interaction of
a particular kind. Transcription is a very general activity, but any particular transcription is a
transcription of a particular coding strand of DNA, into a matched strand of mRNA, which will
be more determinate. Similarly, protein folding is the folding of a polypeptide chain—a chain
of amino acids—into a functional protein. Specific foldings take many forms, going through at
least three stages to form secondary, tertiary, and quaternary protein structure.
Glennan (forthcoming, Chapter 2) has argued that the parallel point needs to be made about
entities. The concept of entity is extremely abstract, and classifications of kinds of entities will
form hierarchies of increasingly concrete and determinate forms. For instance, mRNA and pro-
teins are different kinds of entities, but each of these kinds have a variety of more determinate
kinds that fall under them. The same is true with functionally specified entities: all promoters
have features in common that make them the kind of entity they are, but they come in many
more determinate forms, promoting the transcription of different coding strands of DNA.
Entity and activity varieties provide us with one way of classifying mechanisms into kinds.
We can lump mechanisms by the kinds of entities which are their parts, and also by the kinds
of activities that those parts engage in. Some mechanism classifications emphasize similari-
ties in entities, while others emphasize similarities in activities. For example, biochemical
mechanisms are largely grouped together as similar based on their entities, and competi-
tive mechanisms are largely grouped based on their activities and interactions, while social
mechanisms share a bit of both.
The kinds of entities and kinds of activities involved in a mechanism constrain but do not
completely determine each other. This is crucial to discovery methods that involve using
knowledge of an entity to identify its activity, or vice versa (see Chapter 19, this volume).
Particular activities are activities of particular entities, and a particular entity can only take part
in some activities, while a particular activity can only be produced by some entities, but entities
don’t have to have proper activities. Generally quite different kinds of entities may engage in
similar activities, and particular kinds of entities may engage in very different sorts of activities
95
and interactions. Think of all the things that mouths or screwdrivers can do, or of the many
different kinds of entities that one can use to break a window or plug a leak.
Indeed, whether something counts as an entity or activity at all is not something that can be
answered except locally, in the context of particular phenomena. We take this to be Dupré’s
point when he argues that nothing is an object tout court, but only at a particular timeframe
(Dupré 2012). On a long timeframe, Dupré argues, even mountains are processes (which we
take to be activity-like), because on a geological timeframe, even mountains are constantly
changing. Something similar might be said for at least some activities and entities in mechanisms,
which exist at a particular timescale, depending on the phenomenon that the mechanism is for.
Cells and tissues are sufficiently stable that particular cells or tissues often count as entities in
many mechanisms for bodily phenomena. Nevertheless, they are sufficiently changeable that
they do not count in mechanisms for other phenomena, and may count instead as whole mech-
anisms or even systems—which are in their turn decomposable into both entities and activities.
5. Varieties of organization
What a mechanism does, and so how we discover it and use it in explanation, depends not just
on the entities and activities of which it is made, but also on how these constituents are organ-
ized. Electrical circuits nicely illustrate this fact: the very same resistors or capacitors can exhibit
very different resistance or capacitance depending upon whether they are wired in parallel or in
series. Similarly, the developmental processes by which fertilized eggs divide and develop into
embryos and ultimately mature organisms depend upon variations in concentrations of proteins
within different regions of the egg. The fact, for instance, that a mature fruit fly has its wings
and body segments where it does depends crucially upon the locations and rates at which gene
products express within the developing embryo.
We treat mechanistic organization as a separate dimension of mechanistic variety, because
it can vary largely independently of the kinds of entities and activities that constitute the
mechanism. Consider, for example, forms or organization like positive and negative feedback
loops. In positive feedback loops, a mechanism produces a change in some property of the
system, and this change in turn feeds back into the system, amplifying the effect. In contrast,
in negative feedback loops, the changing property feeds back in a way that dampens the effect.
These and other forms of organization can be found in all manner of mechanistic processes and
systems—electrical, molecular and chemical, genetic, climatological, economic, and social. It is
the fact that such forms of organization induce predictable patterns of behavior in mechanisms
that allows for the development of representational and modeling techniques like dynamical
systems theory, control theory, and information theory that are largely independent of the
particular kind of entities or activities involved (see Chapter 20, this volume).
These kinds of organizational varieties can be called topological and functional. They show
how parts of mechanisms and their activities are arranged—spatially, temporally, and causally.
These kinds of features are often captured in mechanism diagrams (see Chapter 18, this volume)
or via mathematical formalisms like structural equation models or Bayes nets, which indicate
which parts are connected to which, and the functional forms of such dependencies.
The varieties of topological and functional organizations of mechanisms are essentially limitless,
and we shall not try to survey them further here. We do, however, want to call attention to a few
more basic ways to classify mechanistic organization that are of both methodological and onto-
logical import. To begin, we can classify mechanism by the number of parts they have. Muscular
skeletal mechanisms, for instance, have relatively few parts (muscles, bones, cartilage, etc.)—
few enough that scientists can identify each of those parts along with its role in the mechanism.
96
In contrast, in molecular mechanisms like the mechanism of protein synthesis, there are many
parts—e.g., many segments of mRNA and many proteins—and operations are completed many
times over. Whether mechanisms have few or many parts has important consequences for the
techniques scientists use to describe, manipulate, and explain them. For instance, in mechanisms
of protein synthesis, diagrams represent token entities and activities of a process that occurs many
times over. Description of what the mechanism does will not be simply in terms of producing a
protein, but in producing proteins in various concentrations and at various rates. This is very dif-
ferent from a representation of the parts of a circuit in an electronic amplifier, where the wiring
diagram explicitly identifies and locates each part.
We can also classify mechanisms by the degree to which their parts are uniform. Some
mechanisms are composed of a small set of largely uniform parts, in the way that Lego struc-
tures may be made out of a large number of similar blocks. Other mechanisms have a variety
of different parts with different capacities and roles. For instance, at the lowest molecular level,
the components of DNA, RNA, and proteins are small in number, and each of these building
blocks (nucleotides, amino acids) is more or less identical, with the structural and functional
diversity of proteins arising from the many different ways in which these simple building blocks
can be combined. At higher levels of mechanistic organization in multi-cellular organisms, we
often see much more structural and functional diversity in parts. Locomotion in animals, for
instance, requires the orchestration of very different types of entities and activities—muscular,
skeletal, pulmonary, respiratory, neurological, and so on.
Finally, we can classify mechanistic organization by what brings about and maintains
relationships between parts. At one extreme, which we call induced organization, these rela-
tionships are imposed by an external agent, like when a host arranges the seating of guests
at a dinner party; at the other, which we call affinitive organization, these relationships arise
from the affinities that parts bear to each other, like when the guests at the dinner party seat
themselves. In such a case, the arrangements of the guests and who interacts with whom will
be determined in part by chance, but also by who each guest is disposed to sit with. All mecha-
nistic organization is to some degree affinitive, because of the simple fact that different entities
have different capacities to act and interact with each other. However, sometimes, as in many
chemical processes, reactions occur largely due to random interactions between molecules with
different kinds of affinities, whereas in other cases, like the case where protein concentrations
vary from the front to middle to back of a developing egg cell, the organization was induced
by prior processes that set the next stage in the developmental mechanism. All of these forms
of organization matter to how we mechanistically explain and discover mechanisms, especially
since we have become able to recognize and model forms of organization that recur in many
mechanisms (see Chapter 20, this volume).
6. Varieties of etiology
Because mechanisms are localized in space and time, they must have etiologies—that is to say,
there must be some causal process that led up to their existence. There is a difference between
what caused a mechanism to come into being and how it works now, though as we shall show
in this section, there are often connections between a mechanism’s etiology on the one hand,
and what it is made of and how it is organized on the other.
While mechanical philosophy owes much in its origins to analogies with machines of
human construction, the etiology of such mechanisms is typically unlike that of mechanisms
responsible for naturally occurring phenomena. Mechanical devices like windmills or cars
have what we call designed-and-built etiologies. This is to say that an agent, the designer,
97
identified some phenomena they wanted to produce (like the grinding of corn or moving
passengers over roads), and then set about to collect and arrange a set of parts so that their
activities and interactions would in fact bring about the desired result. It is not just the artifacts
of human engineering that can be designed and built. Many social, political, and economic
systems are also “engineered” in this way, and there is no requirement that designers and
builders must be human agents.
Most mechanisms, however, evolve over time. Evolved etiologies are by no means limited
to biological systems. In the most basic sense, evolved mechanisms are those that are built and
modified over time, so that the present characteristics of the mechanism have emerged gradu-
ally from earlier stages or versions of the mechanism. While the idea of evolved mechanisms
brings most immediately to mind biological mechanisms, abiotic systems and processes like stars,
volcanoes, and weather systems also have evolved etiologies in our sense, as do many social,
legal, and economic mechanisms—like property or commodity markets or systems regulating
property, marriage, and child-rearing.
There are a variety of different kinds of processes that can underlie the evolution of mecha-
nisms and mechanical systems. For instance, in stellar evolution, the gradual changes in stars
from one stage to another depend mainly upon the star’s mass and the concentrations of various
elements within the stellar core. As a star consumes hydrogen in fusion, gradually its properties
will change. Evolutionary biology is concerned with the etiology of populations of reproducing
organisms (and their traits), and there appear to be a number of processes—selection, mutation,
migration, drift—which can drive this evolution. We believe that these represent varieties of
etiological mechanisms, though whether and in what sense they are mechanisms is a matter of
some debate (see Chapter 22, this volume). Similarly, when we speak of developmental mecha-
nisms, we are describing mechanisms by which an organism and its component mechanisms
evolve in the developing embryo (see Chapter 25, this volume).3
One other way that a mechanism can come to be is by the operation of ephemeral mecha-
nisms. Ephemeral mechanisms are mechanisms in which the relationship between the parts of
the mechanism that are responsible for the phenomena are short lived, unstable, and thrown
together by happenstance. The mechanisms responsible for one-off events, like forest fires, car
crashes, and romantic liaisons, are typically ephemeral, but seldom do enduring mechanisms
arise from such processes. In a very different context, Donald Davidson imagined the possibility
of lightning striking a tree and somehow miraculously rearranging the tree to create an exact
molecule-by-molecule replica of himself, which he called Swampman. Were such miraculous
events to occur, the process by which the mechanisms within Swampman came to be would be
ephemeral in an extreme way.
While we distinguish these three kinds of etiologies, the etiologies of particular mechanisms
typically involve several of these elements in concert. For instance, while biological organisms
and the mechanisms that operate in them are the product of evolutionary processes, it is clear
that many events that impact the outcome of those processes will be the result of ephemeral
mechanisms—for instance, mechanisms responsible for one-off mutations or for environmen-
tal changes that lead to extinctions. Many artifacts and technologies have designed-and-built
etiologies, but the design and build process can be iterated, so present versions are either evo-
lutionary modifications of earlier versions, as when we upgrade the plumbing in our houses,
or newly designed and built tokens of a type, but where the design of the type itself has been
modified as the result of an evolutionary process. For instance, my new iPhone was designed
and built, but its design is deeply constrained by earlier iterations of iPhones.
While a mechanism’s constituents and organization are to some degree independent of its
etiology, there are clearly connections. What a mechanism does and how it does are often
98
constrained by its history, and traces of its history are evident in the present mechanism.
How we set about discovering the mechanism and using it in explanation and in attempts to
control the world by altering the operation of the mechanism can be affected by that history.
Perhaps most obviously, evolved and ephemeral mechanisms will tend to rely on affinitive
forms of organization, because there is no external agent to impose organization on them. We
could not effectively alter the operation of the mechanism without taking this into account.
Similarly, if Simon’s argument is correct, evolved systems will tend to have “modular” designs
(Simon 1996; see Chapter 14, this volume). Also, evolved systems will contain vestiges of
earlier versions of the system that have become entrenched; this is obviously true of popula-
tions of organisms evolving by natural selection, but it is equally true of my iPhone, or for that
matter the organization of Britain’s National Health Service. Effective policy or technological
change needs to consider this.
7. Upshots
After this too brief exploration of the varieties of mechanism, we would like to reflect on the
philosophical significance of this classificatory project, and to discuss how it is related to other
attempts to classify mechanisms and mechanical philosophy.
To begin, let us consider whether and in what sense these varieties of mechanisms are real.
Our language has been realist. Mechanisms are things in the world, and we can place them into
categories according to the different properties they have. Nonetheless, our realism is tempered.
The dimensions of mechanistic variety do not allow us to sort mechanism tokens neatly into
kinds. The properties by which we sort cross cut. Mechanisms may have similar patterns in the
kind of phenomena they produce (e.g., in being regular or constitutive) while involving very
different kinds of activities. Or, again, mechanisms made with very different kinds of entities
can be organized in similar ways, and mechanisms that produce the same kind of phenom-
ena may have very different etiologies. Another problem for realism is vagueness. How many
parts does one need to have to be a many-parted mechanism, or with what probability must
a start-up condition lead to a termination condition to have us count a mechanism as regular?
A third problem is that mechanisms will often be hybrids involving multiple varieties within
each of these dimensions. Consider, for instance, the varieties of entities and activities involved
in mechanisms responsible for mood disorders, or the variety of organizational motifs in gene
regulatory networks.
However, the fact that there are many ways to carve up the world does not mean that
there isn’t a world out there that constrains and makes sense of our carvings. We think our
account is consistent with what Mitchell calls a “pluralist realist approach to ontology, which
suggests not that there are multiple worlds, but that there are multiple correct ways to parse our
world, individuating a variety of objects and processes that reflect both causal structures and our
interests” (Mitchell 2009, p. 13; see also Wimsatt 1994 and Chapter 15, this volume). While
Mitchell mentions “interests,” we should also add abilities, since how we parse the world will
depend upon the tools we have. Take for instance the distinction between few and many-parted
mechanisms. This distinction is not only vague; it is largely determined by our cognitive and
computational resources. Many-parted mechanisms require different techniques for discovery,
representation, and control—but this is because of our abilities rather than any intrinsic fea-
ture of the mechanism. More generally, we think that the way we classify mechanisms into
varieties reflects the models we use to represent them, and that the suitability of such models
depends both on token mechanisms in the world and upon our interests and epistemic resources
(Glennan forthcoming, Chapter 4).
99
We turn now to a discussion of the relation of our account of mechanistic variety to other
taxonomic efforts in the literature. As we noted at the outset, care needs to be taken to dis-
tinguish kinds of mechanisms, which are things in the world, from kinds of (capital “M”)
Mechanisms, which are philosophical and scientific claims or views about the role of mecha-
nisms and mechanistic reasoning. This distinction is easily lost in taxonomic discussions, since
mechanistic philosophical projects typically implicitly or explicitly presuppose a conception of
what mechanisms are in the world.
Let us begin with Levy’s taxonomy, which is of big “M” Mechanisms. He identifies three
distinct but related philosophical projects that he calls “Causal Mechanism (CM),” “Explanatory
Mechanism (EM),” and “Strategic Mechanism (SM)” (Levy 2013). CM is a metaphysical pro-
ject intended to give an account of causation in terms of mechanisms. EM is a thesis about
explanation to the effect that good explanations must cite information about mechanisms, while
SM aims to draw attention to the heuristics used in mechanism discovery. The projects of
EM and SM are epistemological and methodological. Quite rightly, Levy observes that Causal
Mechanism of the sort proposed by Glennan must embrace a permissive conception of mecha-
nisms in the world (like our minimal mechanism) if it is to have any hope of making the
claim that causal connections are all or mostly mediated by mechanisms. Similarly, he observes
that Strategic Mechanism, which he associates most prominently with Bechtel and Richardson
(1993), adopts a narrower conception of mechanisms, in which they are machine-like, and
where decomposition and localization are powerful strategies. Explanatory Mechanism, which
he associates most prominently with MDC (2000) and Craver (2007), also focuses on a some-
what narrower conception of mechanism, especially those that are always or for the most part
regular, and have hierarchical organization.
Andersen (2014) offers a broadly similar distinction between two projects, which she calls
Mechanism2 (“Mechanisms as an ontology of the world”), which aims to use mechanisms to
give a complete account of what there is, including an account of causation, and Mechanism1
(“Mechanisms as integral to scientific practice”), which focuses on the practices of discovery,
representation, and explanation associated with hierarchically organized mechanisms. Roughly,
Andersen’s Mechanism2 and Levy’s CM identify metaphysical or ontological projects, while
Andersen’s Mechanism1 and Levy’s EM and SM identify explanatory, epistemological, and
methodological projects.4
Nicholson explores the concept of mechanism in biology, and argues that the word “mech-
anism” has principally been used to refer to machine-like structures in the world: “machine
mechanisms,” “a step-by-step explanation of . . . a causal process,” and to “Mechanicism”:
“the philosophical thesis that conceives organisms as machines” (Nicholson 2012, p. 153).
The first of the theses refers to a kind of small “m” mechanism, while the last of these is
clearly a big “M” Mechanism (see Chapter 5, this volume). The middle of these is neither,
but it is associated with what Nicholson calls “the Mechanismic Program,” and which most
philosophers now call “the New Mechanism” and is close to Andersen’s Mechanism1 and
Levy’s Explanatory Mechanism.
Kuorikoski’s (2010) taxonomy distinguishes two kinds of small “m” mechanisms—
“computational causal systems” and “abstract forms of interaction.” The former he takes
to be the sort of mechanisms that have been chiefly of interest in recent discussions of
mechanisms in biology (and we surmise in EM, SM, and Mechanism1). The latter is the
concept Kuorikoski sees at work in discussions of mechanisms in the social science literature.
The former kind of mechanisms are the sorts of mechanisms amenable to decomposition
and localization. The latter refer to mechanisms where these strategies fail, but which can
be characterized abstractly in terms of certain kinds of organization. He cites as examples
100
selection mechanisms in evolution, various kinds of market mechanisms, crowding out, and
self-fulfilling prophecies. (As Kuorokoski notes, abstract form of interaction (AFI) mecha-
nisms are not themselves abstract, but the explanations of such mechanisms are.)
We find that there is much that is informative in these kinds of attempts to taxonomize
mechanisms and Mechanisms, but there are clear limitations to these approaches—limitations
that are mitigated by more carefully attending to the varieties of mechanisms. These attempts
all try to identify a few kinds of mechanisms, and to illustrate how these different kinds of
mechanisms show up in different scientific domains or are appropriate for different projects.
For instance, we notice a frequent distinction between a “broad” and a “narrow” conception of
mechanism, where the narrow conception is “machine-like” and the broad one is not.
We find this approach problematic because it is unclear how narrow a narrow conception
should be, and any narrow conception can be broadened in many different directions. In par-
ticular, these accounts suppose a clear univocal conception of what a machine is, but machines
themselves are massive in their variety. A calculator is a machine, and so is a Watt governor, and
how different are their properties!
We also doubt that we can identify some kinds of mechanisms that are important for
ontological and metaphysical projects, and others that are the kinds that scientists talk about.
As we note above, Levy and Andersen suggest that Mechanism2/Causal Mechanism requires
a broad conception of mechanism; while this is true so far as it goes, we think that the dis-
tinctions between varieties of mechanisms we make here will be central to creating a broad
account of mechanism that still recognizes important variations. For instance, if we do not
clearly distinguish between constitutive and non-constitutive phenomena, we will not be
able to sort out the relationships between part-whole and causal forms of dependence, which
are often important. At the same time, there are other interesting projects that, in spite of
being metaphysical, might still focus on rather narrow forms of mechanisms. For instance,
we might imagine that distinctions along the dimensions we have offered would be of assis-
tance in sorting out various forms of emergence. Similarly, we can observe that very broad
conceptions of mechanisms, like Kuorikoski’s AFIs, can be methodologically significant.
Illari (2011) also argues for the need for a relatively broad understanding of mechanisms to
characterize an important role for evidence of mechanism in causal inference in medicine
(see also Clarke et al. 2014).
We can moreover do much to explicate some of the proposed taxonomies of mechanisms
by reference to our account of mechanistic varieties. Take, for instance, Kuorikoski’s notion
of mechanisms as abstract forms of interaction. From our perspective, the sorts of mechanisms
Kuorikoski identifies are mechanism varieties that are identified by the kinds of organization
they have. Different kinds of markets (e.g., perfect markets, monopolies and oligopolies, and
markets exhibiting various kinds of market failures) will be the kind of markets they are in virtue
of how they are organized. Monopolies, for instance, have only one seller, and we can make
predictions about how that market will behave in light of that fact.
In a recent unpublished paper, Levy and Bechtel have argued that it is time to move
to “Mechanism 2.0.” They describe structures and processes that biologists are seeking to
understand and that appear to be mechanisms (in our sense of minimal mechanism) but for
which the tools of discovery, representation, and explanation described by new mechanists
of the last 20 years do not quite work. They suggest that it is best to stop worrying about
whether these structures and processes are or are not mechanisms by some definition, and
instead to explore the various ways in which these biological mechanisms break the narrow
mold. We certainly concur, and we hope that our approach provides some tools that will
help in that exploration.
101
Notes
1 As we noted in Chapter 1, the word “mechanism” is used both to refer to worldly mechanisms and to
philosophical or methodological theories or approaches that are mechanistic. We will continue to follow
the convention we adopted there of referring to the former with lower-case “m” and the latter with
upper-case “M.”
2 Much of the terminology used in this taxonomy is introduced and elaborated in more detail in
Chapter 5 of Glennan (forthcoming).
3 In biology, the word “evolution” is typically used to refer to the change in populations and their charac-
ters over time, and contrasted with development, which is responsible for changes in individual organisms
from conception to maturity. But both evolutionary and developmental mechanisms are kinds of etiolog-
ical mechanisms, and the mechanisms that result from both of these processes have evolved (as opposed
to designed and built) etiologies.
4 Andersen seems to us to move between big “M” and small “m” mechanisms, presumably because she thinks
the philosophical projects presuppose particular conceptions of mechanisms. She provides three additional
conceptions of mechanism, which are mechanisms “that are ontologically ‘flat,’ or at least not explicitly
hierarchical in character: equations in structural equation models of causation, causal-physical processes, and
information-theoretic constraints on states available to systems” (Andersen 2014, p. 284). These are more
clearly small “m” mechanisms (though obviously structural equations are not in the world).
References
Andersen, H. (2014) ‘A Field Guide to Mechanisms: Part II’. Philosophy Compass, 9(4): 284–93.
Bechtel, W., & Richardson, R. C. (1993) Discovering Complexity: Decomposition and Localization as Strategies
in Scientific Research. Princeton, NJ: Princeton University Press.
Bogen, J. (2004) ‘Analysing Causality: The Opposite of Counterfactual is Factual’. International Studies in
the Philosophy of Science, 18(1): 3–26.
——. (2008) ‘Causally Productive Activities’. Studies in History and Philosophy of Science Part A, 39(1):
112–23.
Campaner, R. (2013) ‘Mechanistic and Neo-Mechanistic Accounts of Causation: How Salmon Already
Got (Much of) It Right’. Metateoria, 3(February): 81–98.
Clarke, B., Gillies, D., Illari, P., Russo, F., & Williamson, J. (2014) ‘Mechanisms and the Evidence
Hierarchy’. Topoi, 33(2): 339–60. Springer. DOI: 10.1007/s11245-013-9220-9.
Craver, C. F. (2007) Explaining the Brain. Oxford: Oxford University Press.
Craver, C. F., & Darden, L. (2013) In Search of Mechanisms: Discovery across the Life Sciences. Chicago:
Dupré, J. (2012) Processes of Life: Essays in the Philosophy of Biology. Oxford: Oxford University Press.
Gillett, C., & Aizawa, K. (2016) Scientific Composition and Metaphysical Ground. (K. Aizawa & C. Gillett,
Eds). London: Palgrave Macmillan UK. DOI: 10.1057/978-1-137-56216-6.
Glennan, S. S. (2010) ‘Ephemeral Mechanisms and Historical Explanation’. Erkenntnis, 72(2): 251–66.
DOI: 10.1007/s10670-009-9203-9.
——. (forthcoming) The New Mechanical Philosophy. Oxford: Oxford University Press.
Illari, P. M. (2011) ‘Mechanistic Evidence: Disambiguating the Russo–Williamson Thesis’. International
Studies in the Philosophy of Science, 25(2): 139–57. DOI: 10.1080/02698595.2011.574856.
Illari, P. M., & Williamson, J. (2012) ‘What Is a Mechanism? Thinking about Mechanisms across the
Sciences’. European Journal for Philosophy of Science, 2(1): 119. DOI: 10.1007/s13194-011-0038-2.
Kincaid, H. (2012) ‘Mechanisms, Causal Modelling, and the Limitations of Traditional Multiple
Regression’. (H. Kincaid, Ed.). The Oxford Handbook of Philosophy of Social Science. Oxford: Oxford
University Press, 46–64.
Kuorikoski, J. (2010) ‘Two Concepts of Mechanism: Componential Causal System and Abstract Form of
Interaction’. International Studies in the Philosophy of Science, 23(2): 1–19.
Leuridan, B. (2012) ‘Three Problems for the Mutual Manipulability Account of Constitutive Relevance
in Mechanisms’. British Journal for the Philosophy of Science, 63(2): 399–427. DOI: 10.1093/bjps/axr036.
Levy, A. (2013) ‘Three Kinds of New Mechanism’. Biology & Philosophy, 28: 99–114.
Machamer, P., Darden, L., & Craver, C. F. (2000) ‘Thinking about Mechanisms’. Philosophy of Science,
67(1): 1–25.
102
Mitchell, S. D. (2009) Unsimple Truths: Science, Complexity and Policy. Chicago: University of Chicago
Press.
Nicholson, D. J. (2012) ‘The Concept of Mechanism in Biology’. Studies in History and Philosophy of Biol
& Biomed Sci, 43(1): 152–63. Elsevier Ltd. DOI: 10.1016/j.shpsc.2011.05.014.
Salmon, W. C. (1984) Scientific Explanation and the Causal Structure of the World. Princeton, NJ: Princeton
University Press.
Simon, H. A. (1996) The Sciences of the Artificial, 3rd ed. Cambridge, MA: MIT Press.
Stinson, C. (2016) ‘Mechanisms in Psychology: Ripping Nature at Its Seams’. Synthese, 193: 1585–614.
Springer. DOI: 10.1007/s11229-015-0871-5.
Wimsatt, W. C. (1994) ‘The Ontology of Complex Systems: Levels of Organization, Perspectives, and
Causal Thickets’. Canadian Journal of Philosophy, Supplement: 207–74.
Winther, R. (2009) ‘Part-Whole Science’. Synthese, 178(3): 397–427.
103
8
MECHANISMS, PHENOMENA,
AND FUNCTIONS1
Justin Garson
1. Introduction
A mechanism is always a mechanism for a phenomenon. The phenomenon that a mechanism
serves is not somehow incidental to that mechanism, but constitutive of it: mechanisms are
identified, and individuated, by the phenomena they produce. Thus, we can ask about mecha-
nisms for blood clots, or demand-pull inflation, or social cohesion in naked mole rats. But it
does not make sense to ask, “how many mechanisms are in the human body?” or even, “is the
universe a mechanism?” These latter two questions are nonsensical because they do not specify
a relevant phenomenon. That each mechanism has a phenomenon has become a platitude in
the new mechanism literature in the philosophy of science (e.g., Glennan 1996, 52; Bechtel and
Abrahamsen 2005, 423; Craver 2007, 122; Darden 2008, 960; Craver and Darden 2013, 52)—
though as it turns out, there are different ways in which a mechanism “has” its phenomenon.
Sometimes, however, philosophers describe mechanisms not as having phenomena but as
serving functions. “Mechanisms are identified and individuated by the activities and entities that
constitute them, by their start and finish conditions, and by their functional roles” (Machamer,
Darden, and Craver 2000, 14). “A mechanism is a structure performing a function in virtue of
its component parts, component operations, and their organization” (Bechtel and Abrahamsen
2005, 423). “Mechanisms are systems that produce some phenomenon, behavior or function”
(Glennan 2010, 256). “The entities and activities that are part of the mechanism are those that
are relevant to that function or to the end state, the final product that the mechanism, by its
very nature, ultimately produces” (Craver 2013, 141). Sometimes, only parts of mechanisms are
described as having functions; sometimes the mechanism as a whole is described as serving a
function. In the following I will mainly focus on this latter use (though see section 3).
The fact that people sometimes describe mechanisms in terms of their functions, rather than
(or in addition to) their phenomena, raises an intriguing question: are these two ways of talk-
ing about mechanisms mere terminological variants of one another? Or, when we describe a
mechanism as having a function, are we saying something more than, or other than, that it has
a phenomenon? If so, what more, precisely, is being said? We can also frame the question from
an ontological perspective. Compare the set of all mechanisms and the set of all mechanisms that
serve functions. Are these two sets coextensional? Or is the set of mechanisms that have func-
tions a proper subset of the set of mechanisms that have phenomena? In other words, is it that
104
Mechanisms, phenomena, and functions
all mechanisms have phenomena, but for some of those mechanisms, those phenomena happen
to be their functions, too?
In the following, I urge the view that there are at least two broad senses of mechanism at
play in science and in the philosophy of science. (Technically, these should be thought of as two
families of senses of mechanism, since each family may encompass several definitional variants.)
According to the first sense of mechanism, which I will call, following Glennan (forthcoming),
“minimal mechanisms,” mechanisms are defined in terms of their phenomena, but there is no
additional implication that these phenomena are in any intuitive sense functions of those mecha-
nisms, where function is aligned with design, purposiveness, goal-directedness, or utility. Using
Glennan’s example, in this minimal sense, my car is a mechanism for locomotion, but it is also
a mechanism for melting chocolate bars, even though melting chocolate bars is not one of my
car’s functions. Its function is just to get me from place to place. (Perhaps it has other legitimate
functions, for example to serve as a status symbol.) In this minimal sense, the aggregation of mis-
folded proteins is part of the mechanism for Alzheimer’s disease, even though misfolded proteins
do not have the function of producing Alzheimer’s disease.
The second sense of mechanism is what I call, following Garson (2013), the functional sense
of mechanism. Here, mechanisms are identified by the functions they serve, where “function”
is understood as having a connotation of teleology, purposiveness, or design—though as we will
see, working out precisely what this sense of “function” amounts to is beset with controversy.
Cast in an ontological vein, I urge that the class of functional mechanisms is an interesting
proper subset of the class of minimal mechanisms. In this sense of mechanism, a car typically
would not be described as a mechanism for melting chocolate bars, even though it does so fre-
quently enough and even though we can explain how that works. The heart would not usually
be considered a mechanism for causing hemorrhages; it is a mechanism for pumping blood,
even though it does both of those things and we understand how both processes work.
One virtue of recognizing a distinct, functional sense of mechanism is that it helps us think
about the normativity of mechanisms; that is, it helps us understand how mechanisms can break
(see section 4). A mechanism breaks when it cannot perform the function that defines it. In
other words, the class of functional mechanisms is coextensive with the class of mechanisms
that can break. It is not clear whether the minimal sense of mechanism has the resources to
explain how a mechanism can break. A second virtue is that it helps us think about the distinc-
tion between a mechanism’s target and its byproduct (e.g., Craver and Darden 2013, 69), for
example locomotion and melting chocolate bars in the case of the car. The target/byproduct
distinction maps neatly onto the traditional function/accident distinction. A mechanism’s target
is just its function; a byproduct is any effect that is incidental to its discharging its function (in the
functions literature, an “accident”). A third virtue of this functional sense of mechanism is that it
helps us organize biomedical knowledge well (see section 4; also see Chapter 24).
In the following, I will do four things. First, I will provide an overview of the different ways
that a mechanism can have a phenomenon (section 2). Second, I will provide a survey of what
various philosophers have had to say about a mechanism’s function (section 3). Even when
philosophers agree that there is some intimate connection between mechanisms and functions,
they disagree about the nature of that relationship and about the underlying concept of func-
tion. (This is why I say that the functional sense of mechanism is actually a family of senses of
mechanism.) Third, I will discuss some benefits of recognizing a distinct, functional sense of
mechanism (section 4). In closing, I will return to the problem I started with: what is the rela-
tionship between these two ways of talking about mechanisms? The most economical solution is
to think of functional mechanisms as constituting an interesting subset of minimal mechanisms.
105
Justin Garson
2. How mechanisms have phenomena

There are two, and arguably only two, ways that a mechanism has a phenomenon. First, a
mechanism can have a phenomenon in the sense that it causes the phenomenon to occur. That
is, the phenomenon is the effect of the mechanism’s operation. Craver and Darden (2013, 65)
refer to this as the “producing” relation. Second, a mechanism can have a phenomenon in the
sense that the operation of the mechanism somehow constitutes, or realizes, the phenomenon.
Craver and Darden (ibid.) refer to this as the “underlying” relation (though there are open
questions about what, precisely, this constitution relation amounts to; see, e.g., Couch 2011;
Romero 2015; Kaiser and Krickel 2016). The fact that mechanisms can have a phenomenon in
one of these two ways gives rise to two different styles of explanation: etiological mechanistic
explanation and constitutive mechanistic explanation (see Craver 2001, 69). Kästner (in prep)
explicitly makes this connection between the distinction between producing/underlying, on
the one hand, and the distinction between etiological/constitutive explanations, on the other.
I will describe each of these relations in turn.
Consider the genetic mechanisms involved in cystic fibrosis (for now I will discuss cystic
fibrosis as “having” a mechanism or mechanisms, rather than arising from the breakdown of a
mechanism). Cystic fibrosis is characterized by debilitating and even fatal respiratory blockages,
among other problems. When we describe the genetic mechanism involved in the respiratory
blockade characteristic of this disease, we typically do so by describing a certain cause-and-effect
sequence. Although there are many biochemical pathways underlying cystic fibrosis, one very
common pathway begins with the deletion of three nucleotides in a certain segment of DNA
(the delta F-508 mutation in the CFTR gene). This causes the loss of an amino acid, pheny-
lalanine, in the corresponding protein sequence (the CFTR protein). This makes the protein
misfold, which, in turn, disrupts the normal passage of chloride across cell membranes. This
can lead to a buildup of sticky mucus in the lungs and elsewhere, though there is still some
uncertainty about the precise pathways involved. In this case, the delta F-508 mutation is part
of a mechanism for the disruption of chloride transportation in this first, “producing” sense.
Disruption of chloride transport is a late stage in a long and complex cause-and-effect sequence.
In contrast, consider the mechanism for the patellar (knee-jerk) reflex. A tap to the patel-
lar tendon causes a muscle spindle in the quadriceps muscle to stretch, which sends a sensory
signal to the ventral horn of the spinal cord. The sensory neuron synapses directly onto a motor
neuron, which sends a command back to the quadriceps to contract, causing the leg to kick.
(This reflex is an example of a “monosynaptic” reflex because it only involves a single synapse.)
Importantly, this mechanism does not cause the patellar reflex. The reflex is not the terminal
stage of a cause-and-effect sequence. Rather, the patellar reflex is constituted by the whole
sequence of events. When one explains how this mechanism works, one offers a constitu-
tive etiological explanation. (In contrast, if one wants to describe the mechanism that causes
the kick, we are back to the producing relation, since kicking is the terminal stage of a causal
sequence that begins with tapping the patellar tendon.)
Craver and Darden (2013) consider a third potential way that a mechanism can have a
phenomenon, which they call the “maintaining” relationship. This is exemplified by the main-
tenance of homeostatic set points, like the water level of the blood. Consider the mechanism
for dopamine homeostasis. Mesolimbic dopamine neurons originate in the ventral tegmental
area of the midbrain and terminate in the nucleus accumbens. They play a role in mediat-
ing reward, although the precise function of these dopamine signals is unclear (Berridge and
O’Doherty 2013). These neurons use a homeostatic mechanism for maintaining a more-or-less
constant level of dopamine in the synapse, usually at concentrations of between 20 and 50 nM
106
(Grace 2000, 336). To do this, the dopamine neuron’s axon terminal has autoreceptors to
monitor the level of dopamine outside the cell. When the level rises significantly above this
“set point,” the autoreceptor sends a signal to decrease the synthesis of dopamine (by inhibiting
the enzyme tyrosine hydroxylase). When it drops significantly below this level, the autorecep-
tor activates that same enzyme and causes more dopamine to be synthesized inside the neuron.
Sometimes, scientists wish to know how such “set points” are maintained over the long run in
the face of fairly regular perturbations.
On reflection, however, it is not clear that we need to recognize a third, distinct, category for
how mechanisms “have” phenomena to make sense of homeostatic phenomena like dopamine
homeostasis. Rather, dopamine homeostasis is just another sort of phenomenon that we might
wish to give a mechanistic explanation for. When we do provide a mechanistic explanation, we
will either identify a producing relation, or an underlying relation, depending on the details of
our analysis (Kästner in prep). Plausibly, if you were to ask, “how does the dopamine neuron
maintain a fairly steady extracellular dopamine level despite frequent disruptions?” and I tell
you about the dopamine neuron’s autoreceptors and how they regulate tyrosine hydroxylase,
I am describing the mechanism that underlies dopamine homeostasis, and giving a constitutive
mechanistic explanation. If, instead, I were to give you a developmental account of how genetic
and early environmental factors interact to create dopamine homeostasis, I am providing an
etiological mechanistic explanation.
It would be easy to form the impression that when scientists pursue a mechanistic explana-
tion, they first fix the phenomenon clearly and then discover its mechanism. But things are rarely
that simple. Rather, there is often a back-and-forth movement between the way we describe a
phenomenon, and the way we describe its mechanism. This movement is important for docu-
menting the role of mechanisms in the process of scientific discovery (see Chapter 19). This
is what Bechtel and Richardson (1993, 173) call “reconstituting the phenomenon”; Glennan
(2005) and Craver and Darden (2013; Chapter 4) also discuss it. Sometimes in the history of
science, scientists have a certain conception of the phenomenon they are after, and they start
to identify the mechanisms underlying it. What they discover about those mechanisms changes
how they think about the phenomenon itself. Consider how the concept of memory has changed
over the last 50 years because of advances in psychology and neuroscience (Bechtel 2008). Or
consider how the notion of schizophrenia has changed over the last century. Garson (forthcom-
ing) describes how the study of recreational amphetamine use in the United States helped to
“reconstitute the phenomenon” of schizophrenia in the 1960s and 1970s. The Swiss psychiatrist
Eugen Bleuler described the fundamental phenomenon as a “loosening of associations”; today,
most Western psychiatrists would consider this “loosening of associations” as representative of, at
best, one subtype of schizophrenia, as a result of ongoing research.
3. Mechanisms and functions

Several philosophers have recognized an intimate connection between mechanisms and
functions, where “function” has the connotation of teleology, design, or purpose. Consider
a statement like, “the function of zebra stripes is to deter biting flies,” or “the function of
eyespots on butterfly wings is to deflect attack away from vital organs.” Here, “function”
connotes teleology or purpose: deflecting attack is in some loose sense what the eyespots
are “there for.” Philosophers of biology have developed many different theories about what
functions are (Garson 2016). The point I wish to make here is that some philosophers and
scientists have explicitly defined the notion of mechanism in terms of a rich, teleologi-
cal notion of function. These philosophers include Craver (2001, 2013), Piccinini (2010),
107
Justin Garson
Piccinini and Craver (2011), Moss (2012), Garson (2013), and Rosenberg (2015), though
they disagree about the precise relationship between mechanisms and functions and they
disagree about what functions themselves are.
Why would anyone want to tie functions and mechanisms together in this way? One way to
motivate this narrower, functional, sense of mechanism is by bald appeal to intuition. In most
everyday contexts, it seems natural to say that a car is a mechanism for locomotion. In contrast,
it seems strange to say that a car is a mechanism for melting chocolate bars. It seems equally
strange to say that a car is a mechanism for spewing carbon dioxide. But why does it seem
strange? All three of those phenomena (locomotion, melting chocolate bars, emitting carbon
dioxide) are very common, and they are completely explicable in terms of physics and engineer-
ing. In everyday contexts, however, only the first description seems natural or normal; the latter
two sound unusual or strained. Why?
The same point can be made in the biological context instead of the context of artifacts and
tools. In the minimal sense of mechanism, the heart is a mechanism for pumping blood, and it
is also a mechanism for making beating sounds that one can listen to through a stethoscope. But
we typically say that the heart is a mechanism for pumping blood. It sounds strange to say that
it is a mechanism for making beating sounds, even though it does both and both are entirely
explicable in terms of basic physics and biology. What is going on here?
One way of articulating this intuition (rather than explaining it) is in terms of the distinc-
tion between a mechanism’s target and its byproduct (Craver and Darden 2013, 69). In the
case of the car, it seems natural enough to identify locomotion as the mechanism’s target
and the emission of CO2 as a byproduct. In the case of the heart, it seems natural enough
to identify the mechanism’s target as pumping blood and the beating sounds as a byproduct.
The fact that these attributions are fairly stable in most everyday contexts raises the question
of why we speak this way.
This is where functions come in. An obvious difference between locomotion and melting
chocolate bars (or spewing CO2) is that locomotion is the car’s function. That is what cars
are designed to do. Emitting CO2 is a (rather unfortunate) byproduct of its function. That
is not why cars exist. A similar point can be made about the heart. The reason the heart is a
mechanism for pumping blood, rather than making beating sounds, is that pumping blood
is the heart’s function. That is what the heart is for. Perhaps there are some contexts where
people are willing to say that the heart is a mechanism for making heart sounds (for example,
in a diagnostic context). But to the extent that there is some context-dependency in the way
people talk about mechanisms, that might be explicable in terms of the context-dependency
in the way people talk about functions. Quite generally, according to this functional sense
of mechanism, for X to be (part of) a mechanism for Y, Y must be its function. As I will
show in section 4, this construal of mechanism also helps us understand the normativity of
mechanisms; that is, how mechanisms break.
Suppose we accept that, in some contexts, mechanisms are defined in terms of the function
they serve. That raises an obvious question: what are functions? Philosophers have explored
this question for the last 40 years with little agreement (see Garson 2016 for a recent over-
view). There are three mainstream theories of function on the market: the selected effects (SE)
theory, the causal role (CR) theory, and the biostatistical theory (BST), though this list is not
exhaustive. Roughly, according to the SE theory, the function of a trait is the reason it evolved
by natural selection. The function of the heart is to pump blood, rather than make beating
sounds, because it evolved by natural selection for pumping blood. BST characterizes a trait’s
function in terms of its current-day contribution to survival and fitness, rather than in terms
of selection history. Like SE, BST defines function in terms of evolutionary considerations
108
(that is, present-day fitness) but remains neutral about history. According to the CR theory,
the function of a system part consists in its contribution to some system capacity that an
investigator has picked out as especially worthy of attention. Unlike the other views, CR
emphasizes the contextual and perspective-dependent aspects of functions.
I am not interested, in this place, in characterizing these theories of function more rigorously
or discussing alternatives. Rather, the point I wish to make is that people might think about
mechanisms slightly differently depending on how they think about functions. As Piccinini
(2010, 286) put the point, “different notions of mechanism may be generated by employing
different notions of function.” That is why I say that the functional sense of mechanism is really
a family of different senses of mechanism. We can talk about the SE-functional sense of mecha-
nism, the CR-functional sense of mechanism, and so on.
For example, one fairly austere way of thinking about mechanisms is to define mechanism
in terms of function, and then define functions in terms of selected effects (that is, using the SE
theory of function). Some philosophers and scientists have done just that. This is a fairly narrow
way of thinking about mechanisms since it implies that for X to be a mechanism for Y, X must
have been shaped by natural selection for doing Y. We can refer to this as the SE-functional sense
of mechanism. In this sense, the heart is a mechanism for pumping blood, and not making beating
sounds, because it was shaped by natural selection for pumping blood.
In the 1960s, the evolutionary theorist G. C. Williams (1966, 9) recommended this
way of talking about mechanisms. As he put it, “the designation of something as a means
or mechanism for a certain goal or purpose will imply that the machinery involved was fash-
ioned by selection for the goal attributed to it.” Here, mechanisms are defined in terms
of function (“purpose”), which, in turn, is defined in terms of natural selection. Williams
thought there was something deeply counterintuitive about describing something as a
“mechanism” for an effect that it was not plausibly selected for. “Should we therefore
regard the paws of a fox as a mechanism for constructing paths through snow? Clearly we
should not” (13). By the same token, I suppose that Williams would also be loath to say
that, “a car is a mechanism for melting chocolate bars,” or that, “the heart is a mechanism
for causing hemorrhages.”
Some of the contemporary evolutionary psychologists have adopted Williams’ usage in the
way that they characterize psychological mechanisms (see Garson 2013, 322 for discussion).
Among philosophers, Rosenberg (2015) seems to endorse this SE-functional sense of mecha-
nism. In (Garson 2013) I endorse the functional sense of mechanism, but do not specify which
theory of function is correct. However, I emphasize (p. 319) that function should be defined
somewhat narrowly in terms of selection, fitness, or design.
A reservation one might have about the SE-functional sense of mechanism is that it might be
too narrow; that is, it might fail to account for the wide spectrum of mechanisms out there. For
example, sometimes neuroscientists talk about neural “mechanisms” for various activities, even
when they do not think those activities are their evolved functions (Craver 2013, 141). It also
would not apply to mechanisms outside of the contexts of biology and engineering, such as in
physics, chemistry, or geology. As Glennan (2005, 445) points out, one may wish to talk about
mechanisms underlying the eruption of geysers, even though geysers have no evolved functions.
However, there is a broader sense of function that can encompass these diverse contexts. This is
the CR theory of function, to be discussed shortly.
Craver (2001, 2013) explored the connection between mechanisms and functions in substan-
tial detail. He says that mechanisms are defined and individuated in terms of the functions they
serve. As he puts it, “the entities and activities that are part of the mechanism are those that are
relevant to that function or to the end state, the final product that the mechanism, by its very
109
Justin Garson
nature, ultimately produces” (2013, 141). Here, the “function” of a mechanism is equated with
its “end state” or “final product.” This way of thinking about mechanisms also makes an appear-
ance in Machamer, Darden, and Craver’s well-known paper, which states that,
to the extent that the activity of a mechanism as a whole contributes to something in

a context that is taken to be antecedently important, vital, or otherwise significant,
that activity too can be thought of as the (or a) function of the mechanism as a whole.
(2000, 6)
However, instead of embracing the SE theory of function, Craver accepts the CR theory of
function. We can label the whole package of ideas the CR-functional sense of mechanism. One
distinctive feature of CR functions is that they have an explicitly perspectival, or contextual,
character. For Craver, the function of the “uppermost” system in a mechanistic analysis is simply
some capacity that a researcher or research team has found especially worthy of attention. To say
the system has a function does not imply that it has an intrinsic goal or purpose. To the extent
that CR functions can be associated with teleology or purpose, they have a derived teleology and
derived purpose—derived, that is, from the interests and goals of the researchers who investigate
them. As Craver puts it,
This teleological feature of mechanistic description is also implicit in the fact that
mechanisms such as the NMDA receptor are bounded: a judgment has been made
about which entities, activities, and organizational features are in the mechanisms and
which are not.
(Ibid., 140)
Craver’s view about mechanisms raises complex questions about realism and antirealism about
mechanisms (see Chapter 9).
One of the benefits of Craver’s expansive way of thinking about mechanisms is that it allows
us to make sense of mechanism-talk outside of the contexts of biology and engineering (for
more on the topic of mechanisms in engineering, see Chapter 34). Nothing prevents us from
giving a CR-functional analysis of, say, El Niño phenomena, or demand-pull inflation, or even
the way that atoms aggregate into molecules.
Lenny Moss (2012) also thinks there is a tight connection between mechanisms and functions
(or “goals”). He points out that, often in the life sciences, biologists only attribute mechanisms to
systems that are in some sense goal-directed: “To count as a biological mechanism the phenom-
enon in question thus must be perceived as being an expression of the ostensible ‘purposiveness’
of the living cell or organism” (165). Note that to say that an organism or biological system
is “goal-directed” is different from saying that it has functions (see Garson 2016, chapter 2
for discussion). The point here is that Moss agrees with Craver (2013) and Garson (2013) that
mechanisms have a teleological dimension. Part of Moss’ evidence for this view is that, in many
cases, we would refrain from calling something a mechanism if it were merely an artifact of
laboratory procedure and not somehow expressive of an organism’s goal-directedness:
if cells stick to tissue culture plastic because of a chemical reaction with the plastic that
resulted in the happenstance chemical production of an epoxy resin this too would
be registered as an artifact and not as a “mechanism of adhesion” . . . . It is strictly the
teleological aspect which makes the difference.
(Ibid.)
110
There is a potential source of ambiguity that comes up when we discuss the relation between
mechanisms and functions. There is a difference between attributing a function to a mechanism,
qua whole, and attributing functions to the mechanism’s parts. Up until now, the philosophers
that I have discussed hold that, for X to be a mechanism for Y, Y must be X’s function (or at
least, X must be part of a goal-directed system, as in Moss’ view). In other words, they see the
mechanism as a whole as the thing that has a function, purpose, or goal. Other theorists tend to
attribute functions, first and foremost, to the parts of mechanisms, rather than (or in addition to)
the whole mechanism. This seems to be a fairly common stance. For example, Craver (2001)
emphasizes the way that parts of mechanisms have functions. He refers to these part-functions
as “mechanistic role functions.” At times, Bechtel and Richardson (1993) also use “function”
specifically when they are describing the activities of a mechanism’s parts, rather than a mecha-
nism’s phenomenon. For example, they define “mechanistic explanations” as explanations that
“account for the behavior of a system in terms of the functions performed by its parts and the
interactions between these parts” (17). Piccinini and Craver (2011) also describe the way that
mechanisms have functions, but if I understand them correctly, they typically focus on the way
that parts of mechanisms have functions, rather than the whole mechanism.
Craver (2001) makes an interesting distinction regarding these mechanistic role functions.
He observes that when we attribute a function to a component of a system, we can point to
the item’s activity without making any reference to the role that it plays in the broader system,
or we can emphasize its role, irrespective of the specific activity by which it performs this
role. A simple example will clarify the distinction. Consider the function of the heart. On the
one hand, we can describe the heart as having the function of beating. This just describes what
it does irrespective of the contribution it makes to the system (circulating blood). On the
other hand, we can describe the heart as having the function of helping to circulate blood. This
describes the role it plays in the circulatory system without telling us the specific activity it
performs (beating). Craver (2001, 65) refers to these two roles as the item’s “isolated activity”
and “contextual role” respectively. In his view, “a complete description of an item’s role
would describe each of these.”
There is one further source of ambiguity that arises when we think about functions. I suspect
that some writers in the new mechanism tradition use the term “function,” purely synony-
mously with “phenomenon.” That is, in some cases I suspect that “function” represents nothing
more than a terminological variation on “phenomenon” or “behavior,” and hence the class of
mechanisms that have “functions” in this weak sense is coextensional with the class of what
Glennan calls “minimal mechanisms.” For example, Glennan (2002, 127, fn. 6) notes that “it
is tempting to use ‘function’ in place of the term ‘behavior’.” Ultimately, he suggests there
that we resist that temptation, because “function” may carry inappropriate connotations of
design. Similarly, Bechtel and Abrahamsen (2005, 433) characterize function in terms of the
familiar function/structure distinction, which is devoid of connotations of teleology or purpose.
However, I see a potential problem here. As I will describe in the next section, we often want
to describe a mechanism as “broken,” and one good way to make sense of that is by appealing
to function, in some appropriately restrained sense of that term.
4. When mechanisms break

Biologists and biomedical researchers employ a colorful vocabulary to talk about how mech-
anisms break. A mechanism can have a “breakdown.” It can be “usurped,” “coopted,” or
“hijacked.” It can be “interfered with,” “disrupted,” “impaired,” or “disabled.” It can sim-
ply “fail.” For example, “drugs of abuse can hijack synaptic plasticity mechanisms in key brain
111
Justin Garson
circuits” (Kauer and Malenka 2007, 844). “Potentially irreversible impairments of synaptic
memory mechanisms in these brain regions are likely to precede neurodegenerative changes”
(Rowan et al. 2003, 821). The idea that mechanisms can break plays an important role in the
way that biomedical researchers organize knowledge. Many philosophers in the new mecha-
nism tradition have also emphasized how the idea of a mechanism’s breaking is crucial for
understanding causation, identifying components of mechanisms, and explaining disease
(e.g., Bechtel and Richardson 1993, 19; Craver 2001, 72; Glennan 2005, 448; Darden 2006, 259).
So, what is it for a mechanism to break? At first pass, it is tempting to say that a mechanism
breaks just when it cannot yield the phenomenon that defines it. My truck is a mechanism
for locomotion, but because of a rusty spark plug, it cannot start. Sophia’s immune system is
a mechanism for fighting off harmful pathogens, but because of a low white blood cell count
(leukopenia) due to a virus, it cannot do that (or not as effectively as usual). It is broken.
But this way of speaking raises a deeper question. Suppose my truck cannot start. Why
would we say that the truck is a broken mechanism for locomotion, rather than simply that it
is not currently (or no longer) a mechanism for locomotion? What license do we have (no pun
intended) to say that my truck is a mechanism for locomotion, that is, that locomotion is its
phenomenon, even when it cannot actually yield this phenomenon? Why not just say it is no
longer a mechanism for locomotion?
Indeed, some of the standard definitions of mechanism, read strictly, do not allow us to
speak of broken mechanisms. The set of entities and activities constituting my broken truck
are not “organized such that they are productive of regular changes [that is, those constituting
locomotion].” My busted truck is not a “complex system that produces the behavior [of
locomotion] by the interaction of a number of parts.” This suggests that we should find
some way of loosening up those definitions to allow my truck to be a broken mechanism for
locomotion, rather than simply not a mechanism for locomotion at all.
There are a few ways to respond to this situation. One way is to bite the bullet and say that,
strictly speaking, there is no such thing as a broken mechanism. For example, in Glennan’s min-
imal sense of mechanism, one would have to say that my broken truck is not a mechanism for
locomotion, since it cannot actually perform that activity (though it might still be a mechanism
for something else such as melting chocolate bars). This maneuver has the disadvantage that it
seems to run against much of biomedical usage, as indicated in the quotations that open this
section. Another way of accounting for the possibility of broken mechanisms is to explain it in
terms of a purely statistical or epistemic norm. For example, one might hold that, when we say
that the truck is a “broken” mechanism, all we mean is that it is acting in a way that is atypical
or unexpected. This is a step in the right direction, but we should not equate brokenness with
atypical or unexpected behavior. The medulla oblongata is a mechanism for helping us breathe,
and it is also a mechanism for initiating the gag reflex. The latter is atypical, and perhaps unex-
pected, but one would not want to say it is broken when it causes us to gag.
My suggestion here is that a mechanism breaks when it cannot perform the function that
defines it. The function of my car is locomotion, but it cannot perform that function because of
a rusty spark plug, so it is broken. The function of Sophia’s immune system is to fight pathogens,
but it cannot do so because of an unusually low white blood count, so it is broken. Perhaps
there are other ways of explaining what it is for a mechanism to break; my only claim is that
this is an obvious option. I also do not want to restrict functions to selected effects or design (as
in artifacts); as I indicated in the last section, one might construe function broadly in terms of
the CR theory of function. My point here is that when we say that a mechanism is broken, we
seem to imply that it has some function that it is unable to perform, even if we disagree about
precisely what functions are.
112
Why are functions uniquely suited to this role? Because functions, at least on standard analy-
ses, have a normative character. To say that functions are “normative” just means it is possible for
a token system to have a function that it cannot perform. Functions have the remarkable property
that they linger, as it were, even in the absence of the corresponding capacity. Standard analyses
of function seem to provide a reasonable explication of this normative character, though there is
ongoing debate about how exactly the different theories of function should account for norma-
tivity. For example, on the standard SE theory, the function of a trait is the reason it evolved by
natural selection. So, the function of a trait has to do with its history, rather than its current-day
capacities. As a consequence, it is easy to understand how a trait can possess a function (owing
to its history) that it is nonetheless unable to perform.
Mechanisms, too, at least in standard biological usage, have a normative character.
Mechanisms are the sorts of things that can break. Put simply, it is possible for X to be a
mechanism for Y even when X cannot do Y. In my view, the normativity of mechanisms
derives from the normativity of functions. On reflection, this is not such a radical philo-
sophical move. Philosophers of biology have often appealed to the normativity of functions
to make sense of the normative character of other biological categories. Dretske (1986)
and Neander (1995) appeal to the normativity of function to make sense of the normative
character of representations. Neander (1991) and Rosenberg and Neander (2009) appeal
to the normativity of function to make sense of the way that types of traits are individu-
ated. Appealing to function is a natural, even obvious, way to think about the normative
character of mechanisms.
The idea that mechanisms serve functions not only helps us think about how mechanisms
break; it is also useful for organizing biomedical knowledge. Suppose we accept that mecha-
nisms serve functions, in some rich, teleological sense of that term. Then, we (generally) would
be prohibited from talking about mechanisms for disease. For example, in the functional sense,
there is no mechanism for anencephaly. Rather, anencephaly is an explicable result of the break-
down of a mechanism for neurulation. Moghaddam-Taaheri (2011, 608–10) makes a similar
point in her discussion of medicine, though she does not situate this insight within a general the-
ory of mechanism. Neander (2017, chapter 3) argues that diseases are generally best explained
in terms of deviations from proper function, though she does not specifically relate this point to
the concept of mechanism.
Let me clarify the point by contrasting two different ways of talking about diseases. On the
first way of talking about disease, all diseases have mechanisms. There is a mechanism for spina
bifida, and anencephaly, and encephalocele, among others. Biomedical research would be the
organized attempt to explicate the mechanisms for these diverse diseases, along with their spa-
tial, temporal, and hierarchical features. On the second way of talking about disease, diseases
generally do not have mechanisms. Rather, there are only mechanisms for functions. Here,
there is only one functional mechanism involved, namely the mechanism for neural tube forma-
tion. Different diseases result when this mechanism breaks in various ways.
Biomedical researchers often adopt this latter way of speaking, though see Garson (2013)
for important qualifications. The reason is that the latter way of speaking is highly informative,
and it also provides a useful heuristic for discovery. First, when I describe anencephaly as the
result of a breakdown in a mechanism for neurulation, rather than as having its own mecha-
nism, I convey critical information about its etiology. I am guiding you to the root problem, as
it were, underlying anencephaly. Second, I set up a heuristic for future biomedical discoveries.
Anencephaly results from disrupting neural tube folding at the anterior neuropore. What happens
if folding is disrupted at the posterior neuropore instead? These are the kinds of questions that
come up when we frame diseases in terms of broken mechanisms.
113
Justin Garson
In conclusion, I accept a form of mechanism pluralism. There are different concepts of

mechanism at play in biology and in philosophy, and my goal is not to argue for the unique
correctness or superiority of a single one. In one sense (Glennan’s “minimal mechanisms”),
the notion of mechanism serves as a way of understanding causation and constitution quite
generally, even outside of biology and engineering. For example, it can help us to understand
mechanism-talk in areas such as physics, chemistry, and geology, where talk about functions
is generally out of place. In another sense (the “functional” sense), mechanisms are defined in
terms of the functions they serve, where function is thought of in a suitably rich teleological
sense. Mechanisms in this functional sense constitute an interesting subset of mechanisms in
this minimal sense. Perhaps there are other senses of mechanism as well (see Chapter 7). This
functional sense of mechanism can help us understand certain aspects of mechanisms, such as the
distinction between a mechanism’s target and byproduct, the normativity of mechanisms, and
the role of mechanisms in biomedical research.
Note
1 I am grateful to several people for discussions that helped me develop the ideas in this chapter. These
include Carl Craver, Lindley Darden, Lenny Moss, Karen Neander, Gualtiero Piccinini, Anya Plutynski,
and Sahotra Sarkar. I am also grateful to Phyllis Illari and Stuart Glennan for inviting me to contribute a
chapter to this volume and for their valuable editorial comments.
References
Bechtel, W. (2008) Mental Mechanisms: Philosophical Perspectives on Cognitive Neuroscience, New York:
Routledge.
Bechtel, W., and Abrahamsen, A. (2005) “Explanation: A Mechanist Alternative,” Studies in the History and
Philosophy of Biological and Biomedical Sciences 36: 412–441.
Bechtel, William, and Richardson, R. C. (1993) Discovering Complexity: Decomposition and Localization as
Strategies in Scientific Research, Princeton, NJ: Princeton University Press.
Berridge, K. C., and O’Doherty, J. P. (2013) “From Experienced Utility to Decision Utility,” in
P. W. Glimcher and E. Fehr (Eds.) Neuroeconomics: Decision Making and the Brain, 2nd ed. Amsterdam:
Elsevier, 335–351.
Couch, M. B. (2011) “Mechanisms and Constitutive Relevance,” Synthese 83: 375–388.
Craver, Carl F. (2001) “Role Functions, Mechanism, and Hierarchy,” Philosophy of Science 68: 53–74.
——. (2007) Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience, New York: Oxford
University Press.
——. (2013) “Functions and Mechanisms: A Perspectivalist View,” in P. Huneman (Ed.) Functions:
Selection and Mechanisms. Dordrecht: Springer, 133–158.
Craver, C. F., and Darden, L. (2013) In Search of Mechanisms: Discoveries Across the Life Sciences, Chicago:
Darden, Lindley. (2006) Reasoning in Biological Discoveries, Cambridge: Cambridge University Press.
——. (2008) “Thinking Again about Biological Mechanisms,” Philosophy of Science 75: 958–969.
Dretske, Fred. (1986) “Misrepresentation,” in R. Bogdan (Ed.) Belief: Form, Content, and Function. Oxford:
Clarendon Press, 17–36.
Garson, J. (2013) “The Functional Sense of Mechanism,” Philosophy of Science 80: 317–333.
——. (2016) A Critical Overview of Biological Functions, Dordrecht: Springer.
——. (forthcoming) “‘A Model Schizophrenia’: Amphetamine Psychosis and the Transformation of
American Psychiatry,” in S. Casper and D. Gavrus (Eds.) Technique in the History of the Brain and Mind
Sciences. Rochester, NY: University of Rochester Press.
Glennan, Stuart. (1996) “Mechanisms and the Nature of Causation,” Erkenntnis 44: 49–71.
——. (2002) “Contextual Unanimity and the Units of Selection Problem,” Philosophy of Science 69:
118–137.
——. (2005) “Modeling Mechanisms,” Studies in History and Philosophy of Biological and Biomedical Sciences
36: 443–464.
114
——. (2010) “Ephemeral Mechanisms and Historical Explanation,” Erkenntnis 72: 251–266.
——. (forthcoming) The New Mechanical Philosophy, Oxford: Oxford University Press.
Grace, A. A. (2000) “Gating of Information Flow within the Limbic System and the Pathophysiology of
Schizophrenia,” Brain Research Reviews 31: 330–341.
Kaiser, M. I., and Krickel, B. (2016) “The Metaphysics of Constitutive Mechanistic Phenomena,” British
Journal for the Philosophy of Science. doi: 10.1093/bjps/axv058.
Kästner, L. (in prep) “On the Curious Trinity of Mechanisms.”
Kauer, Julie A., and Malenka, Robert C. (2007) “Synaptic Plasticity and Addiction,” Nature Reviews
Neuroscience 8: 844–858.
Machamer, Peter, Darden, Lindley, and Craver, Carl F. (2000) “Thinking about Mechanisms,” Philosophy
of Science 67: 1–25.
Moghaddam-Taaheri, Sara. (2011) “Understanding Pathology in the Context of Physiological Mechanisms:
The Practicality of a Broken-Normal View,” Biology and Philosophy 26: 603–611.
Moss, Lenny. (2012) “Is the Philosophy of Mechanism Philosophy Enough?” Studies in History and
Neander, Karen. (1991) “Functions as Selected Effects: The Conceptual Analysts’ Defense,” Philosophy of
Science 58: 168–184.
——. (1995) “Misrepresenting and Malfunctioning,” Philosophical Studies 79: 109–141.
——. (2017) A Mark of the Mental: In Defense of Informational Teleosemantics Cambridge, MA: MIT Press.
Piccinini, Gualtiero. (2010) “The Mind as Neural Software? Understanding Functionalism, Computationalism,
and Functional Computationalism,” Philosophy and Phenomenological Research 81: 269–311.
Piccinini, G., and C. Craver. (2011) “Integrating Psychology and Neuroscience: Functional Analyses as
Mechanism Sketches,” Synthese 183: 283–311.
Romero, F. (2015) “Why There isn’t Inter-Level Causation in Mechanisms,” Synthese. doi: 10.1007/
s11229-015-0718-0.
Rosenberg, A. (2015) “Making Mechanisms Interesting,” Synthese. doi: 10.1007/s11229-015-0713-5.
Rosenberg, Alex, and Neander, Karen. (2009) “Are Homologies Selected Effect or Causal Role Function
Free?” Philosophy of Science 76: 307–334.
Rowan, M. J., Klyubin, I., Cullen, W. K., and Anwyl, R. (2003) “Synaptic Plasticity in Animal Models
of Early Alzheimer’s Disease,” Philosophical Transactions of the Royal Society of London B 358: 821–828.
Williams, George C. (1966) Adaptation and Natural Selection, Princeton, NJ: Princeton University Press.
115
9
THE COMPONENTS AND
BOUNDARIES OF
MECHANISMS
Marie I. Kaiser
Most new mechanists agree that mechanisms are made of two kinds of components (see Chapter 1).
The first kind of components are material objects that are variously called entities (Machamer,
Darden and Craver (MDC) 2000), parts (Glennan 1996, 2002), or component parts (Bechtel
2006, 2008). Examples include neurotransmitters, muscle fibers, and genes. The second kind
of components are variously called activities (MDC 2000), interactions (Glennan 1996, 2002),
or operations (Bechtel 2006, 2008). They are what entities do and what produces change
(MDC 2000). Examples include binding, contracting, and being transcribed. I will use the term
“component” to refer to these two kinds of things that make up or compose mechanisms.1
Since the origin of the new mechanical philosophy, metaphysical questions about the com-
ponents of mechanisms (hereinafter “mechanistic components”) have figured prominently in
the debate (e.g., Tabery 2004; Psillos 2004; Machamer 2004; Bogen 2008; Torres 2009; Illari
and Williamson 2013). MDC (2000) provoked this dispute by defending a dualistic ontol-
ogy of mechanistic components, according to which mechanisms are made up of components
belonging to two distinct ontological categories: entities (having certain properties) and activi-
ties. MDC contrast their dualism with Glennan’s (1996, 2002) allegedly monistic view that
mechanisms consist of entities (having certain properties) and of interactions, which are nothing
but occasions on which a property change of one entity brings about a change in properties of
another entity.
The dispute between dualists and monists involves two sets of metaphysical questions: First,
how can the two kinds of mechanistic components be further characterized? What are their
major features and what distinguishes them from each other? How are entities/parts and activi-
ties/interactions/operations related to more well-known ontological categories, such as material
objects, properties, processes, events, and dispositions? I address these questions in section 1.
Second, what is the relation between the two kinds of mechanistic components? Is one more
fundamental than the other? Can activities, for instance, be reduced to entities and their proper-
ties? In what sense are the two kinds of mechanistic components mutually dependent? I examine
these questions in section 2.
The second part of this chapter (sections 3 and 4) is concerned with the individuation of the
components of mechanisms and with the boundaries of mechanisms. Mechanisms are always
mechanisms for specific phenomena or behaviors (Glennan 1996, 2002). Accordingly, the com-
ponents of a mechanism are individuated with respect to the phenomenon the mechanism is
116
The components and boundaries of mechanisms
responsible for (see Chapters 1 and 8): Only those entities/parts and activities/interactions/
operations that are relevant to the phenomenon count as components of the mechanism. This
raises the question of how to spell out the conditions of relevance. What makes entities/parts
and activities/interactions/operations relevant to a specific phenomenon? How do we draw
the boundary between what belongs to a mechanism and what does not? In section 3, I pre-
sent different criteria for individuating the components of mechanisms—natural boundaries,
robustness/stability, strength of interactions, mutual manipulability, and INUS conditions—and
I discuss their merits and limitations.
The claim that the individuation of a mechanism (by individuating all of its components)
depends on the characterization of the respective phenomenon invites the question of how real
the boundaries of mechanisms are. Mechanistic antirealists, as I call them, argue that mechanisms
do not exist as well-delineated entities in nature because it is the scientist who imposes bounda-
ries on certain sets of components while pursuing specific explanatory interests and adopting
specific perspectives (Bechtel 2015; see also Wimsatt 1972, 2007). In section 4, I examine
whether such an antirealistic move is inevitable and which arguments can be offered to defend
the realistic view that mechanisms exist in nature and have real boundaries and a determined
set of components.
1. The ontological nature of mechanistic components

The goal of this section is to provide a general characterization of the two kinds of components
that make up mechanisms. I will refer to mechanistic components as entities and as activities
because entity-activity terminology has prevailed in the debate (even Glennan forthcoming
speaks about activities now) and because there are good reasons for this terminological choice:
First, the term “entity” is more suitable than the term “part” because, as I argue somewhere else
(Kaiser forthcoming), being a component of a biological mechanism is not the same as being a
biological part of a biological object in general. The components of a mechanism are individu-
ated with respect to a single behavior, whereas the parts of biological objects are individuated
with respect to all of their characteristic behaviors. Moreover, natural boundaries are crucial to
the individuation of biological parts but they do not constrain the individuation of mechanistic
components in the same way (more on this in section 3). Second, we can drop Bechtel’s (2006,
2008) term “operation” because he does not tie any metaphysical claim to this term. Third,
against the term “entity” one might object that it is commonly used as a placeholder term for
any ontological category (including activities) and that we should be more precise and speak
about material objects instead (Kaiser and Krickel 2016). But in the context of this handbook it
is advisable to stick to the term “entity” in this specific usage because it is an established term in
the debate. Fourth, I prefer the term “activity” over the term “interaction” because interactions
can plausibly be seen as special kinds of activities. Fifth, against the concept of an activity one
might object that it is too restrictive and should be replaced by the broader metaphysical con-
cept of an occurrence (which includes processes, events, and states; Kaiser and Krickel 2016).
For the present purposes, however, I stick to the familiar concept of an activity while critically
discussing what activities are.
I begin here by considering what entities are (as components of mechanisms). The mecha-
nism for muscle contraction is composed of entities such as actin filaments (i.e., strands of
actin molecules), calcium ions, and sarcoplasmic reticula (i.e., cell organelles involved in
the storage and release of calcium ions). What the new mechanists refer to as “entities” are
concrete material objects that are located in space and time. Typically, entities are spatially
extended, have certain shapes and sizes, and often are surrounded by characteristic spatial
117
Marie I. Kaiser
boundaries, such as the membrane comprising the sarcoplasmic reticulum. Entities exist in
time and are wholly present at different time points. Contrary to activities, entities are not,
themselves, temporally extended in the sense of having temporal parts.2 A calcium ion in my
muscle fiber now is wholly present; it does not consist of the calcium ion now, the calcium
ion two days ago, the calcium ion in one minute, etc. Entities can engage in processes, events,
or activities which are temporally extended (e.g., diffusing through a channel). But the cal-
cium ion as such is not temporally extended. In metaphysical terms, entities are continuants not
occurrents, as activities are (Kaiser and Krickel 2016).
Entities are bearers of properties, including dispositional properties (or “causal powers”; Glennan
forthcoming). Properties allow entities to engage in specific activities (MDC 2000, 3; Craver
and Darden 2013, 16). For example, a sarcoplasmic reticulum is surrounded by a membrane
that contains different kinds of ion channels and ATP-driven calcium pumps. These properties
enable the sarcoplasmic reticulum to engage in the activity of releasing calcium ions.
Psillos (2004, 312) draws our attention to the fact that entities can exist without actually
engaging in activities (e.g., if entities do not manifest their dispositions). I agree that entities, in
general, can exist without engaging in any activities. But the important point is that entities can
only be components of mechanisms if they engage in activities. Without doing something, an
entity cannot be relevant to the phenomenon the mechanism is responsible for. All conditions
of relevance require something to occur, something to be active (see section 3). Hence, com-
ponent entities are necessarily involved in activities; that is, they are “working entities” (Darden
2008, 961; Craver and Darden 2013, 18). But we need to be cautious at this point. Entities
that are components of mechanisms need not be active at all times. For the working of some
mechanisms it is important that entities engage in activities only at certain times, not at others.
For example, the sarcoplasmic reticulum releases calcium ions when an action potential arrives,
not during later stages of the mechanism for muscle contraction. To conclude, an entity can be
a component of a mechanism only if it is involved in an activity, at least once while the mecha-
nism proceeds from its beginning to its end.
Component entities, however, can be involved in activities in different ways. They can
have an active role and initiate or maintain the activity (i.e., be “actors”; Glennan forthcom-
ing), such as the ATP molecule that binds to the calcium pump. On the other hand, I think
entities can also be passively involved in an activity and allow for or undergo the activity,
such as the calcium pump to which the ATP molecule binds. This difference has not been
recognized in the debate so far. Furthermore, there can be activities in which two or more
entities with an active role are involved, such as the sliding of actin and myosin filaments
past each other. According to the strict reading of the word, such cases would be examples of
interactions. In the debate, however, the term “interaction” is used in a wider sense as refer-
ring to any activity in which more than one entity is involved—let it be passively or actively
(e.g., Tabery 2004; Glennan forthcoming).
Finally, it has been argued that, as components of mechanisms, entities must be “stable
clusters of properties” (Craver 2007a, 131) and must “have a kind of robustness and reality
apart from their place within that mechanism” (Glennan 1996, 53). For example, ATPases
(i.e., transmembrane proteins that transport ions to produce ATP) can be found in a variety
of different mechanisms in different organisms and they retain their properties also if studied
in isolation; that is, if studied in different contexts than in situ (Kaiser 2015, 221–35). Even
though stability and robustness may not succeed as criteria for identifying the components
of mechanisms (see section 3), they might still be important features that many component
entities share.
118
Main features of entities
(1) Entities are material objects (i.e., continuants).

(2) Entities are bearers of properties, which allow them to engage in specific activities.
(3) As components of mechanisms, entities necessarily engage in activities (at least once during the
mechanism).
(4) Entities can be actively or passively involved in activities.
(5) Entities are relatively stable and robust.
I move now to considering what activities are (as components of mechanisms). These are the
second kind of components that make up the mechanism for muscle contraction: activities such
as sliding, releasing, and binding. “Activities are the things that the entities do” (Craver and
Darden 2013, 16). As such, activities are necessarily temporally extended and possess characteristic
durations, rates, and phases. For example, the binding of ATP to the calcium pump starts with
the collision of the ATP molecule with a certain region of the pump, then certain chemical
bonds are formed. Because activities do not only exist in time—as entities do as well—but are
also extended in time and have temporal parts (i.e., stages), they belong to the metaphysical cat-
egory of occurrents, which encompasses processes, events, and states (Kaiser and Krickel 2016).
Activities are thought to be “active rather than passive” and to be the “happenings”
(Machamer 2004, 29) in a mechanism. It is this active nature of activities that seems to make
them indispensable if one wants to account for the alleged activeness of mechanisms (MDC
2000, 5), and it is their active nature that is said to distinguish activities from, for instance,
property instantiations or property changes (Glennan forthcoming rejects this; more on this in
section 2). As I show elsewhere, it follows from the active nature of activities and from their
ontological status as occurrents that activities must be actualized or manifest rather than merely
potential or dispositional (Kaiser and Krickel 2016). Only manifestations of dispositions can be
components of mechanisms; unmanifested dispositions cannot. It is the release of calcium ions
that is a component of the mechanism for muscle contraction, not the unmanifested disposition
of the sarcoplasmic reticulum to release calcium ions if a depolarization occurs.
Dualists claim that activities account not only for the active nature of mechanisms but also
for their productivity (MDC 2000; Tabery 2004; Bogen 2008; Darden 2008). In their view,
the order of activities in a mechanism exhibits a “productive continuity” (MDC 2000, 3): One
activity productively causes the next, which ensures that a mechanism runs in its typical way from
beginning to end (see Chapter 10). For example, the release of calcium ions into the cyto-
sol causes the binding of calcium ions to troponin molecules, which causes the tropomyosin
complex to move off the actin binding site, which causes the myosin head to bind to the actin
filament, and so on. Hence, activities are characterized as the “producers of change” (MDC
2000, 3) and as the “causal components in mechanisms” (Craver 2007a, 5).3
Another central feature of activities is that they require entities that engage in them. As MDC
put it, “there are no activities . . . that are not activities of entities” (2000, 5). Any activity requires
at least one entity that engages in it. I think we can add to this that for each activity there must
be at least one entity that is actively involved in it. That is, activities seem to require actors. For
example, a gene can be passively engaged in the activity being transcribed but there must be
another entity that is actively involved in the same activity, such as the DNA polymerase.
119
Marie I. Kaiser
Different types of activities involve different numbers of entities. In other words, activities
have “unrestricted arity” (Illari and Williamson 2013, 72; my emphasis). Binary activities, such as
the binding of ATP to the calcium pump or the sliding of actin and myosin filaments past each
other, are widespread; activities that involve many entities, such as transcription or osmosis,
are common in the living world. In my view, there are also clear examples of activities that
involve only one entity (so-called unary activities or “un-interactive activities”; Torres 2009,
243), such as the changing of the spatial conformation of a protein, the breaking of a chemi-
cal bond, and the closing of a stoma cell in a plant leaf. Other authors, however, question the
existence of un-interactive activities (Tabery 2004; Fagan 2012). They claim that for an activity
to be productive, it must involve no fewer than two entities and thus be an interaction (or a
“jointly acting complex”; Fagan 2012, 464). Tabery (2004) argues that seemingly un-interactive
activities, such as shifting conformation of a protein, still involve interactions on a lower level.
This argument is not convincing because the claim was that there is no other entity (besides the
protein) involved in the conformation shift. This is compatible with there being interactions
among the parts of the protein. Furthermore, Fagan (2012) fails to provide an argument for
why all component entities must form complexes that jointly act. I conclude that there are cases
of unary activities and that interactions thus constitute a mere subset of the set of all activities.
One might want to add the sixth feature that activities do not only produce changes (i.e.,
act as causes) but also, themselves, involve changes in the properties of the engaged entities (i.e.,
are processes). An example that supports this claim is the release of calcium ions from the sar-
coplasmic reticulum into the cytosol. This activity involves several changes, such as the calcium
ions changing their location and the sarcoplasmic reticulum changing its concentration of stored
calcium ions. Despite its initial plausibility, I think we should reject the claim that activities must
involve changes because it overlooks the variety of mechanistic components. For the work-
ing of some mechanisms, it is crucial that certain properties are not changed but maintained
(i.e., continuously instantiated) during a certain time span. For instance, it is essential for the
mechanism of the action potential that some ion channels remain open for a certain period of
time. Moreover, the mechanism of natural selection seems to consist also of “passive properties”
(Skipper and Millstein 2005, 341), such as being camouflaged or being present (see Chapter 23).
Examples like these show that not all component activities involve changes, which is why this
feature is not listed below.4
Main features of activities
(1) Activities are temporally extended (i.e., occurrents).

(2) Activities are actualized (rather than merely potential).
(3) Activities produce change (i.e., are types of causes).
(4) Activities require at least one actively involved entity.
(5) Activities have unrestricted arity (i.e., involve one to many entities).
2. The relation between mechanistic components

This section examines the relation between entities and activities as components of mechanisms.
Is the relation one of reduction? Can entities, for instance, be said to be more fundamental than
activities? Or are both kinds of components ontologically on a par, for instance because they
120
are mutually dependent? These questions trace back to the already introduced dispute between
dualists (MDC 2000; Machamer 2004) and monists (Glennan 1996, 2002) that has been promi-
nent in the new mechanical philosophy. In this section, I first clarify in what respects dualists
and monists have different views about the relation between mechanistic components. Then I
critically discuss the arguments that have been provided in favor of dualism.
To begin with clarifying the dualism versus monism debate, MDC argue that “[m]echanisms
are composed of both entities (with their properties) and activities” (2000, 3). I call the claim
that mechanisms consist of components of two kinds Duality thesis. This is quite a weak thesis
because monists can accept it as well. Glennan argues that mechanisms are “complex systems”
that consist of “parts” (1996, 52; 2002, 344) having relatively stable properties and interacting
with each other in certain ways. Hence, even though Glennan (1996, 2002) denies that mecha-
nisms are composed of entities and of activities, he also recognizes a second kind of components,
namely interactions. In his recent work, Glennan (forthcoming) agrees with even MDC’s ver-
sion of the Duality thesis because he refers to the components of mechanisms as entities (or parts)
and activities, and characterizes interactions as a special, important class of activities.
Contrary to dualists, however, monists reject the Irreducibility thesis. This is the claim that the
two kinds of mechanistic components belong to two distinct ontological kinds in the sense that
one cannot be reduced to the other. Glennan characterizes an interaction between parts as an
“occasion on which a change in a property of one part brings about a change in a property of
another part” (2002, 344). In other words, Glennan (1996, 2002) assumes that interactions are
reducible to the properties of entities and rejects the Irreducibility thesis. He is thus only commit-
ted to the existence of entities (i.e., material objects) and their properties, including dispositional
properties. By contrast, MDC (2000) postulate the existence of activities on top of the usual
commitments to entities and properties. Because of this additional ontological category, MDC’s
approach is referred to as dualistic and contrasted with Glennan’s monistic account.5
Dualism is characterized by two further assumptions. The first is the Interdependency thesis which
says that entities and activities necessarily exist together and determine each other. MDC state that
“no activities without entities, and entities do not do anything without activities” (2000, 8). This
is why the new mechanists use terms such as “acting entities” (Craver 2007a, 189) or “working
entities” (Darden 2008, 961). The Interdependency thesis does not reinforce the dualistic character
of MDC’s approach—one might even argue that it runs contrary to it—but it is central to their
approach. Second, MDC (2000) defend what I call the Parity thesis. It is the claim that there is no
priority of, for instance, entities over activities and that entities and activities are ontologically on
a par (Illari and Williamson 2013, 70).6 Strictly speaking, the Parity thesis is no independent thesis
because it follows from the dualists’ assumptions of irreducibility and interdependency.
In sum, MDC’s (2000) dualism can be distinguished from Glennan’s (1996, 2002) monism
by the following three claims:
Core theses of dualism
(1) Irreducibility thesis: Entities and activities belong to two distinct ontological kinds because activities
cannot be reduced to the properties of entities.
(2) Interdependency thesis: In mechanisms, entities and activities necessarily exist together and determine
each other.
(3) Parity thesis: Entities and activities are ontologically on a par.
121
Marie I. Kaiser
In his recent work, Glennan (forthcoming) seems to have partly converged to MDC’s (2000)
dualistic position. He speaks of entities (or parts) and activities as the components of mechanisms
and he agrees with MDC in that there cannot be activities without entities, or entities without
activities (which is the Interdependency thesis). Moreover, Glennan rejects the reproach that his
account is guilty of an entity-bias. That is, he seems to want to accept the Parity thesis. This
is, however, the point at which the remaining difference between MDC (2000) and Glennan
(forthcoming) becomes apparent. Glennan rejects the label of dualism and the claim that enti-
ties and activities are two distinct ontological categories because he thinks that the ontological
distinctness of entities and activities is incompatible with the interdependency of entities and
activities. In other words, he sees an incompatibility between the Irreducibility thesis (in particular,
the distinctness claim it contains) and the Interdependency thesis. On my view, this casts into doubt
Glennan’s approval of the Parity thesis because if entities and activities are not ontologically dis-
tinct and activities can be reduced to property changes of entities, it seems highly questionable
that entities and activities are ontologically on a par. To conclude, Glennan (forthcoming) has
converged to dualism by accepting the existence of activities and the Interdependency thesis but he
still rejects the distinctness claim that the Irreducibility thesis contains and thus fails to convincingly
defend the Parity thesis.
Having examined the nature of the disagreement, I will now go on to explore challenges
to dualism. Even though entity-activity talk dominates the debate, dualism and the ontological
category of activities are still disputed (e.g., Woodward 2002; Psillos 2004; Machamer 2004;
Tabery 2004; Bogen 2008; Torres 2009; Illari and Williamson 2013; Glennan forthcoming).
In this section, I introduce and critically discuss five major arguments that have been offered in
defense of dualism.
The new mechanists agree with other philosophers that philosophy of science should pay
close attention to actual scientific practice. They accept “descriptive adequacy” (MDC 2000, 8;
Kaiser 2015, 9) as an important criterion of adequacy not only for epistemic but also for meta-
physical claims (Kaiser and Krickel 2016). Dualists make use of this criterion and argue that
only dualism is descriptively adequate because it accounts for the fact that scientists describe
mechanisms in terms of entities and activities (i.e., by using nouns, such as “enzyme” and
“repressor,” and verbs, such as “inhibits”) and because successful scientific practice is a prac-
tice of studying and manipulating activities (Illari and Williamson 2013, 73–5; see also Craver
2007a, 144–52). This might be convincing but it should be noted that the claim that activities
play an important role in scientific practice does not represent a brute fact about science. It
is already a philosophical interpretation or critical reconstruction of specific parts of scientific
practice which can be contested (Kaiser 2015, chapter 2). In my view, a monist might have
equally good reasons for viewing scientific practice as a practice of studying objects, their prop-
erties, and how these properties change through time. Descriptive adequacy can thus also be
used as an argument against dualism.
A second argument stresses the “epistemic adequacy” (MDC 2000, 21) of dualism. Dualists
argue that thinking about mechanisms in terms of entities and activities renders scientific
phenomena intelligible because entity-activity language corresponds best to how we natu-
rally describe the world (Machamer 2004, 31; Illari and Williamson 2013, 79). For example,
we say that a neurotransmitter binds to a receptor (activity talk), not that a neurotransmit-
ter changes from being unbound to being bound to a receptor (property talk), or that a
neurotransmitter manifests its bonding capacity (disposition talk). Even if this claim raises
questions about the naturalness of descriptions and even if it may have exceptions, I share
the intuition that activity talk is more comprehensible than property or disposition talk and
I agree that activity talk corresponds well to how scientists describe and explain phenomena.
122
When compared to an entity-disposition ontology, dualism is claimed to be epistemically

more adequate because we can observe a disposition of an entity only through the disposi-
tion being manifested; that is, through the entity acting (MDC 2000, 4; Machamer 2004, 30;
Illari and Williamson 2013, 78). This is true but proponents of an entity-disposition ontol-
ogy can accept the conceptual dependency of dispositions on their manifestations without
being committed to the existence of activities in particular. In general, one might object that
all of these arguments refer to mere epistemic points that have no metaphysical relevance
(Psillos 2004, 313). But I agree with the dualists that metaphysicians of science should pay
attention to the epistemic practices in the sciences.
MDC also appeal to common metaphysical intuitions to support dualism. They claim that
dualism is advantageous because it captures “healthy philosophical intuitions underlying both
substantivalist and process ontologies” (2000, 4). The ontological category of entities accounts for
the fact that activities or processes are not free-floating but that there is always an entity involved
in an activity or process. The ontological category of activities, in turn, accounts for the fact that
mechanisms are dynamic and “do things” (MDC 2000, 5) and thus cannot consist of entities hav-
ing certain properties only (recall section 1; Kaiser and Krickel 2016). Dualism is able to account
for these motivating intuitions because it conceives of entities and activities as ontologically
distinct and on a par. The problem with this argument is that it remains unclear about to which
kind of intuitions the dualists refer. The argument would be plausible if they referred to a sort of
pre-theoretical intuitions that metaphysical theories must account for (even then, there might be
disagreement about which intuitions are legitimate). However, MDC speak about philosophical
intuitions and seem to have something more theoretical and metaphysically laden in mind.
Another argument in favor of dualism is that only activities sufficiently explain the activeness
and the productivity of mechanisms (MDC 2000; Tabery 2004; Bogen 2008). Only activities
specify how in a mechanism one property change brings about another property change (i.e.,
how one entity interacts with another). In Tabery’s words, the “bringing about” in Glennan’s
interactions are “black boxes” (2004, 10) that can only be rendered intelligible by activities.
For instance, the activity “releasing” is said to specify how the change of the sarcoplasmic
reticulum’s ion channel from closed to open brings about the shift in the location of some
calcium ions. Dualists claim that if you reject the Irreducibility thesis and argue that activities are
nothing but property changes, you will overlook the active and productive nature of mecha-
nisms. This concern seems to be confirmed, for example, by Glennan’s thesis that “mechanisms
are things (or objects)” (2002, 345), such as watches or cells. As I argue elsewhere, from a
dualistic perspective, it becomes clear that bare objects that are neither active nor productive
(i.e., that do not involve any occurrent), such as a stopped watch, cannot be mechanisms (Kaiser
and Krickel 2016). But the argument that only dualism accounts for the productive nature of
mechanisms also encounters serious problems. First, dualists fail to provide a clear and metaphys-
ically satisfying analysis of what productivity is (e.g., Woodward 2002; Psillos 2004). Second,
the “black box problem” seems to be not a problem for monists but rather “is one for scientists
to solve in the laboratory” (Torres 2009, 239) because scientists are the ones to discover how,
exactly, one particular property change brings about another.
A final argument in favor of dualism pertains to parsimony. Illari and Williamson state that
an entity-activity ontology is more parsimonious than an entity-disposition ontology (2013,
79–81). At first sight, this claim might be surprising because dualism accepts the existence of
activities on top of the usual ontological commitments and thus entails a rather unparsimonious
move. However, Illari and Williamson do not compare dualism with monism, but rather an
entity-activity ontology with an entity-disposition ontology. They argue that, because of the
restricted arity of dispositions (they attach to only one entity), an entity-disposition ontology
123
Marie I. Kaiser
results in a proliferation of indefinitely many dispositions, which is why an entity-activity

ontology is more parsimonious. This argument raises the question of which kind of parsimony
is most important. On which level of graininess should we evaluate parsimony? If the goal of
metaphysics is to reveal the fundamental structure of reality, I doubt that it is worse to have
fewer general ontological categories (entity, property) and more numerous ontological sub-
categories (different entities, many different properties).
3. Individuating mechanistic components

The new mechanists agree that mechanisms consist of all and only those components that are rel-
evant to the specific phenomenon that the mechanism is responsible for. The term “responsible
for” (Bechtel and Abrahamsen 2005, 422; Illari and Williamson 2012, 123) can be specified in
two ways: etiological mechanisms cause their phenomena, whereas constitutive mechanisms con-
stitute their phenomena (Kaiser and Krickel 2016). Figure 9.1 illustrates the triangular relationship
between mechanistic components, a mechanism, and a phenomenon.
Figure 9.1 raises the question, for instance, how we should spell out the relevance relation
between components and phenomena. This question has an epistemic and a metaphysi-
cal side. Epistemically, we are asking for a criterion of explanatory relevance: What are the
explanatorily relevant factors that must be represented in a mechanistic explanation of a given
phenomenon? How do we distinguish adequate mechanistic explanations from mere descrip-
tions?7 Metaphysically, the question of relevance is a question about where the boundary of a
particular mechanism runs: Which entities and activities belong to the mechanism for a specific
phenomenon? Which criteria tell apart genuine components of a mechanism from mere pieces,
correlates, redundant parts, or background conditions?
The two types of mechanisms allow us to distinguish two kinds of relevance relations. The
components of etiological mechanisms are causally relevant to their phenomena, whereas the
components of constitutive mechanisms are said to be constitutively relevant to their phenom-
ena. Since the notion of causal relevance has been subject to extensive philosophical discussion,
the new mechanists focus on the notion of “constitutive relevance” (Craver 2007a, 139–60,
2007b; see Chapter 14).8 I conclude this section by reviewing different possible criteria for
individuating the components of constitutive mechanisms.
One might first suggest identifying the boundaries of mechanisms with what I call “natural
boundaries” (Kaiser 2015, 176) of biological objects (e.g., Darden seems to suggest this; 2008,
960). Paradigmatic examples of natural boundaries are the cell membrane, the exoskeleton of
insects, the skin, the blood–brain barrier, and the chain of mountains that borders a particular
ecosystem. I agree that the existence of natural boundaries is essential to the working of many
mechanisms. For example, the mechanism for muscle contraction requires that the muscle fiber
is surrounded by a membrane that functions as a selective barrier and receives extracellular
phenomenon
responsible for
relevant to
(constitutes/causes)
components mechanism
make up
Figure 9.1 Relationships between components, mechanisms, and phenomena
124
signals. In most cases, however, these natural boundaries will not be the boundaries of the
mechanism but of the corresponding biological object (e.g., of the muscle fiber).
According to my account, the boundary of a mechanism differs in two respects from the
boundary of the corresponding object (what Craver refers to as “S” in his characterization of
the phenomenon “S’s ψ-ing”; Kaiser and Krickel 2016). First, mechanisms frequently transgress
natural boundaries (Craver 2007b, 9). For instance, the mechanism for muscle contraction is
composed also of entities, such as neurotransmitters, that are located outside the membrane of
the muscle fiber. Also the gecko adhesion mechanism includes entities that belong to the gecko’s
environment (Kaplan 2012, 552). Hence, natural boundaries seem to be crucial to identifying
the parts of biological objects in general (e.g., muscle fibers or geckos; Kaiser forthcoming) but
not to individuating the components of mechanisms (e.g., the mechanism for muscle contrac-
tion or the gecko adhesion mechanism).
Second, biological objects, such as cells or geckos, “do many things at once” (Glennan
1996, 52) and all of these characteristic behaviors are relevant to individuating the parts of
the biological object (Kaiser forthcoming). By contrast, mechanisms are responsible only for
a single phenomenon (e.g., muscle contraction or gecko adhesion) and mechanistic compo-
nents must be relevant to this phenomenon only. Hence, a mechanism typically includes many
fewer entities than are parts of the corresponding system. To conclude, natural boundaries fail
to provide us with adequate criteria for demarcating mechanisms because the condition of
being located inside a natural boundary is neither necessary nor sufficient for an entity to be a
component of a mechanism.
A second idea for how to identify the boundaries of mechanisms is using strength of interac-
tions. The idea that the intensity of interactions determines part–whole relations can be traced
back to Simon (1962). This claim is based on the observation that, at least in “nearly decom-
posable” (1962, 474) systems, interactions among the parts of an object are generally stronger
and more frequent than the interactions between an object’s parts and its environment. For
example, the forces holding together a molecule are much weaker than those holding together
atoms within the molecule. Other authors have picked up Simon’s idea and applied it to the
boundaries of mechanisms (e.g., Haugeland 1998; Wimsatt 1974, 2007).
This account, however, faces various problems. In general, it is quite difficult to assess and
compare the strength of interactions. We seem to need to specify a threshold that will depend
on pragmatic considerations of researchers. In addition, the strengths-of-interaction criterion
seems unable to exclude background conditions and sterile effects as components of mechanisms
(Craver 2007a, 143f.). Finally, when looking at actual biological practice we recognize that the
strength of interactions may play a crucial role in individuating biological objects such as popu-
lations or ecosystems (e.g., Huneman 2014; Kaiser forthcoming) but not in individuating the
components of biological mechanisms.
As a third possibility, we might turn to an account of constitutive relevance. Craver (2007a,
139–60, 2007b) has offered the most elaborated and widely discussed account of constitutive
relevance. He identifies two conditions under which an acting entity (X’s φ-ing) is constitu-
tively relevant to a phenomenon (S’s ψ-ing) and thus is a component of the mechanism for this
phenomenon: the parthood condition and the mutual manipulability condition.
The parthood condition requires that “X is a part of S” (Craver 2007a, 153) or that X is
“contained within S” (Craver and Tabery 2017). Hence, Craver seems to identify parthood with
spatial inclusion. According to my view, the assumption that any component entity X must be
spatially included in the corresponding object or system S encounters two major objections. First,
it overlooks that temporal relations between component activities and the phenomenon of inter-
est are important as well (Kaiser and Krickel 2016). Second, it conflicts with the fact that many
125
Marie I. Kaiser
mechanisms transgress natural boundaries (a fact that Craver, himself, recognizes; 2007a, 141) and
include also entities and activities that are spatially located outside the corresponding object or
system S (recall the discussion of natural boundaries in this section).
The second condition of Craver’s account of constitutive relevance relies on Woodward’s
interventionist theory of causation (2003) and requires mechanistic components and phe-
nomena to be mutually manipulable. This means that there must be an ideal intervention on
the putative component X’s φ-ing that changes the phenomenon S’s ψ-ing and there must
be an ideal intervention on S’s ψ-ing that changes X’s φ-ing. This part of Craver’s account
of constitutive relevance has recently attracted much philosophical attention (e.g., Harbecke
2010; Couch 2011; Leuridan 2012; Baumgartner and Gebharter 2016; Harinen forthcoming).
It has been objected that Craver’s account is merely epistemic and does not provide a satisfac-
tory metaphysical analysis of what constitutive relevance or constitution really is (Harbecke
2010; Couch 2011). It seems to me that this criticism is too premature. Even though Craver’s
analysis of the explanatory and investigative practices of neuroscience (2007a) is not intended
to be a metaphysical analysis, it can easily be used to draw metaphysical conclusions about the
relation of constitution (Kaiser and Krickel 2016). Another criticism is that Craver’s notion of
constitutive relevance turns out to be a subtype of causal relevance and thus requires rethink-
ing either Woodward’s notion of an ideal intervention or the mutual manipulability account
altogether (Leuridan 2012; Baumgartner and Gebharter 2016; Harinen forthcoming; see also
Woodward 2014).
Finally, I will consider what are known as “INUS conditions.” Regularity theories of consti-
tution (Harbecke 2010; Couch 2011) develop criteria for individuating mechanistic components
by making use of Mackie’s idea of a cause being an INUS condition (i.e., an Insufficient but
Non-redundant part of an Unnecessary but Sufficient condition for the effect; Mackie 1965).
The general idea is that a component is an insufficient but necessary part of an unnecessary but
sufficient mechanism for a phenomenon. This idea nicely captures the fact that a single compo-
nent is insufficient for a phenomenon. A component must always be a non-redundant member
of a “team” (Gillett 2013, 311) of components, which, together, are sufficient for the phenom-
enon. The criterion also accounts for the possibility that different particular mechanisms might
be sufficient for a certain type of phenomenon.
It is problematic that regularity theories of constitution either require mechanistic phenom-
ena to be capacities, which is an inadequate view of what constitutive mechanistic phenomena
are (Kaiser and Krickel 2016), or they conceive of constitution as a relation between types
only, which presupposes a realism about biological kinds that is heavily contested. Moreover,
the claim that a mechanistic component must be a necessary member of a team of components
raises the well-known problem of how to exclude mere correlates from being components (see
also Craver and Tabery 2017).
4. Are boundaries of mechanisms real?

A scientific realist might begin with the intuition that mechanisms exist out there in nature
independently of the scientists discovering and representing these mechanisms. Where one
mechanism ends and another starts seems to be an objective feature of reality, which depends
neither on our ability to individuate mechanistic components nor on our having an account of
constitutive relevance. Mechanisms seem to be real things with real boundaries.
Although many mechanists share these realist intuitions, a realist interpretation of the bound-
aries of mechanisms encounters serious challenges. A commonly held assumption is that how
126
one identifies the components of mechanisms depends upon the phenomenon one seeks to
understand (recall section 3). Characterizing the phenomenon differently will yield different
sets of relevant components and thus result in drawing the boundary of a mechanism differ-
ently. Since the choice of the phenomenon is relative to scientists’ explanatory interests, some
authors have argued that the identification and decomposition of mechanisms is an inherently
“perspectival” matter (Darden 2008, 960; see also Kauffman 1971; Wimsatt 1972, 2007). From
here it seems to be only a small step to the antirealistic claim that mechanisms do not exist as
well-delineated things in the world but that scientists impose boundaries on mechanisms rela-
tive to their explanatory purposes (Bechtel 2015). The aim of this section is to discuss whether
such an antirealistic move is inevitable and which arguments a realist might invoke to defend
the independent existence of boundaries of mechanisms.
Let us begin with the claim that characterizing a phenomenon guides and constrains the indi-
viduation of the components of a mechanism. According to the new mechanists, characterizing
the phenomenon confines the space of possible mechanisms in so far as it uses the language of
a given field and implicitly calls up a set of explanatory concepts and a “store of accepted enti-
ties, activities, and organizational structures that people in the field are licensed to use” (Craver
and Darden 2013, 67). The relation between mechanisms and phenomena is thought to be
reciprocal: not only does the nature of the phenomenon constrain what is to be included in the
mechanism, but findings about the mechanism can also force one to “reconstitute the phenom-
enon” (Bechtel and Richardson 2010, 173).
But does this imply that the boundaries of mechanisms are not real in the sense of being no
objective feature of reality? I think it does not. A realist can agree that for different token phe-
nomena there exist different token mechanisms with different boundaries and that the choice of
scientists of which phenomena to study may change over time and is dependent on pragmatic
factors such as the scientist’s perspective and explanatory interests. A realist would, however,
emphasize that as soon as the phenomenon of interest is sufficiently specified and fixed, the
boundaries of the mechanism for this phenomenon are fixed too, and do not depend on prag-
matic factors. This way a realist can agree that the choice of the phenomenon is relative to the
scientist’s interests and perspective while holding on to the idea that for each token phenom-
enon there exists a token mechanism out there in the world that has a well-delineated, real
boundary. Such a realistic position is taken up, for instance, by Glennan (forthcoming), and it
seems to be the position that Craver is also committed to (not least because of his ontic concep-
tion of explanation; 2014).
Other authors reject the realistic assumption that boundaries of mechanisms exist in nature
independently of the fact that scientists take specific perspectives and pursue certain explana-
tory goals when decomposing mechanisms. Wimsatt, for instance, argues that in the case of
“interactionally complex” systems, different “theoretical perspectives” give rise to different non-
isomorphic decompositions of a system into parts with “non-coincident boundaries” (Wimsatt
1974, 69–71). This sounds quite antirealistic, as if Wimsatt claimed that parts and boundaries
would exist only relative to theoretical perspectives. But at other places in his work, Wimsatt
explicitly defends realism—though he is keen to add that his realism is a realism for a “messy
world” and is “piecemeal and usually satisfied with a local rather than a global order” (2007, 5f.;
see also 1974, 672).
We can find antirealism about the boundaries of mechanisms more explicitly in the
recent work by Bechtel. He draws on research in neurobiology to show that “mechanisms
as bounded entities don’t exist” (2015, 84). Marom’s (2010) argument that the time-course
of many phenomena in the life sciences is scale-free, in Bechtel’s view, does not challenge
127
Marie I. Kaiser
the mechanistic account. It merely shows that the common view of mechanisms as temporally
and spatially well-delineated entities is literally false—even though it is an idealization that
facilitates developing scientific explanations. Bechtel argues that it is “the scientists who
impose boundaries around entities and activities in nature and impose a time scale on which
their functioning is characterized” (2015, 85). A realist may object that if boundaries of
mechanisms do not exist in the actual world, mechanisms do not exist either, because it is
indeterminate which entities and activities belong to which mechanism. Bechtel’s antirealism
seems to result in the implausible view that the world consists of loose entities and activities
only. Alternatively, Bechtel’s antirealism can be read as turning the mechanistic account into
a purely epistemic account that has been detached from almost all ontological commitments.
Notes
1 I do not use the term “constituent” because some mechanisms are said to constitute their phenomena
(Kaiser and Krickel 2016) and the relation between mechanisms and phenomena should be kept apart
from that between components and mechanisms.
2 This claim presupposes endurantism (Lewis 1986).
3 In the debate, there is no agreement on what exactly a cause is. Some hold that causes are difference-
makers (e.g., Craver 2007a), in the sense that if they had not occurred, the purported effect would not
have occurred. Others claim that causation is production (e.g., Bogen 2008; Glennan 2010), where causes
are actively producing their effects.
4 Alternatively, one might stick to the claim that activities must involve changes and argue that mecha-
nisms consist of more than just entities and activities—namely property instantiations or maintenances
(i.e., states).
5 Strictly speaking, these numbers are false because material objects and properties are already two onto-
logical categories.
6 Only Machamer (2004) deviates from the Parity thesis insofar as he sometimes suggests the priority of
activities.
7 According to the ontic conception of explanation, the question of explanatory relevance is not an
epistemic but also a metaphysical question because explanations are regarded as things in the world
(Craver 2014).
8 Whether Craver’s account of constitutive relevance succeeds in keeping apart constitutive and causal
relevance or whether it collapses into an account of causal relevance is, however, contested (e.g., Leuridan
2012; Baumgartner and Gebharter 2016; Harinen forthcoming).
References
Baumgartner, M., Gebharter, A. (2016): “Constitutive relevance, mutual manipulability, and fat-handedness”,
British Journal for the Philosophy of Science 67 (3), 731–756.
Bechtel, W. (2006): Discovering Cell Mechanisms: The Creation of Modern Cell Biology, Cambridge: Cambridge
University Press.
—— (2008): Mental Mechanisms: Philosophical Perspectives on Cognitive Neurosciences, New York: Routledge.
—— (2015): “Can mechanistic explanation be reconciled with scale-free constitution and dynamics?”,
Studies in History and Philosophy of Biological and Biomedical Sciences 53, 84–93.
Bechtel, W., Abrahamsen, A. (2005): “Explanation: a mechanist alternative”, Studies in History and
Philosophy of Biological and Biomedical Sciences 36, 421–441.
Bechtel, W., Richardson, R.C. (2010): Discovering Complexity: Decomposition and Localization as Strategies in
Scientific Research, Cambridge, MA: MIT Press.
Bogen, J. (2008): “Causally productive activities”, Studies in History and Philosophy of Science 39, 112–123.
Couch, M.B. (2011): “Mechanisms and constitutive relevance”, Synthese 183, 375–388.
Craver, C. (2007a): Explaining the Brain, Oxford: Clarendon Press.
—— (2007b): “Constitutive explanatory relevance”, Journal of Philosophical Research 32, 3–20.
128
—— (2014): “The ontic conception of scientific explanation”, in M.I. Kaiser, O. Scholz, D. Plenge,
A. Hüttemann (eds.), Explanation in the Special Sciences – The Case of Biology and History, Dordrecht:
Springer, 27–52.
Craver, C., Darden, L. (2013): In Search of Mechanisms, Chicago: University of Chicago Press.
Craver, C., Tabery, J. (2017): “Mechanisms in science”, in E.N. Zalta (ed.) The Stanford Encyclopedia
of Philosophy (Spring 2017 Edition). Available at: <https://plato.stanford.edu/archives/spr2017/
entries/science-mechanisms/>
Darden, L. (2006): Reasoning in Biological Discoveries: Essays on Mechanisms, Interfield Relations, and Anomaly
Resolution, Cambridge: Cambridge University Press.
—— (2008): “Thinking again about biological mechanisms”, Philosophy of Science 75, 958–969.
Fagan, M.B. (2012): “The joint account of mechanistic explanation”, Philosophy of Science 79, 448–472.
Gillett, C. (2013): “Constitution, and multiple constitution, in the sciences: using the neuron to construct
a starting framework”, Minds and Machines 23, 301–337.
Glennan, S. (1996): “Mechanisms and the nature of causation”, Erkenntnis 44, 49–71.
—— (2002): “Rethinking mechanistic explanation”, Philosophy of Science 69, 342–353.
—— (2010): “Mechanisms, causes, and the layered model of the world”, Philosophy and Phenomenological
Research 81, 362–381.
—— (forthcoming): The New Mechanical Philosophy.
Harbecke, J. (2010): “Mechanistic constitution in neurobiological explanations”, International Studies in the
Philosophy of Science 24, 267–285.
Harinen, T. (forthcoming): “Mutual manipulability and causal inbetweenness”, Synthese.
Haugeland, J. (1998): Having Thought, Cambridge, MA: Harvard University Press.
Huneman, P. (2014): “Individuality as a theoretical scheme. II. About the weak individuality of organisms
and ecosystems”, Biological Theory 9, 374–381.
Illari, P., Williamson, J. (2012): “What is a mechanism? Thinking about mechanisms across the sciences”,
European Journal for Philosophy of Science 2, 119–135.
—— (2013): “In defense of activities”, Journal for General Philosophy of Science 44, 69–83.
Kaiser, M.I. (2015): Reductive Explanation in the Biological Sciences, Cham: Springer.
—— (forthcoming): “Individuating part-whole relations in the biological world”, in O. Bueno,
R.-L. Chen, M.B. Fagan (eds.), Individuation, Process and Scientific Practices, Oxford: Oxford University
Press.
Kaiser, M.I., Krickel, B. (2016): “The metaphysics of constitutive mechanistic phenomena”, The British
Journal for the Philosophy of Science, DOI: 10.1093/bjps/axv058.
Kaplan, D.M. (2012): “How to demarcate the boundaries of cognition”, Biology and Philosophy 27,
545–570.
Kauffman, S.A. (1971): “Articulation of parts explanation in biology and the rational search for them”,
PSA: Proceedings of the Biennial Meeting of the Philosophy of Science Association 1970, 257–272.
Leuridan, B. (2012): “Three problems for the mutual manipulability account of constitutive relevance in
mechanisms”, British Journal for the Philosophy of Science 63, 399–427.
Lewis, D.K. (1986): On the Plurality of Worlds, Oxford: Blackwell.
Machamer, P. (2004): “Activities and causation: the metaphysics and epistemology of mechanisms”,
International Studies in the Philosophy of Science 18, 27–39.
Machamer, P., Darden, L., Craver, C.F. (2000): “Thinking about mechanisms”, Philosophy of Science 67, 1–25.
Mackie, J.L. (1965): “Causes and conditions”, American Philosophical Quarterly 24, 245–264.
Marom, S. (2010): “Neural timescales or lack of thereof”, Progress in Neurobiology 90, 16–28.
Psillos, S. (2004): “A glimpse of the secret connexion: harmonizing mechanisms with counterfactuals”,
Perspectives on Science 12, 288–319.
Simon, H.A. (1962): “The architecture of complexity”, Proceedings of the American Philosophical Society 106,
467–482.
Skipper, R.A., Millstein, R.L. (2005): “Thinking about evolutionary mechanisms: natural selection”,
Studies in History and Philosophy of Biological and Biomedical Sciences 36, 327–347.
Tabery, J.G. (2004): “Synthesizing activities and interactions in the concept of a mechanism”, Philosophy
of Science 71, 1–15.
Torres, P.J. (2009): “A modified conception of mechanisms”, Erkenntnis 71, 233–251.
Wimsatt, W.C. (1972): “Complexity and organization”, Proceedings of the Philosophy of Science Association
1972, 67–86.
129
Marie I. Kaiser
—— (1974): “Reductive explanation: a functional account”, PSA: Proceedings of the Biennial Meeting of the
Philosophy of Science Association 1974, 671–710.
—— (1994): “The ontology of complex systems: levels of organization, perspectives, and causal thickets”,
Canadian Journal of Philosophy 24, 207–274.
—— (2007): Re-Engineering Philosophy for Limited Beings: Piecewise Approximations to Reality, Cambridge:
Harvard University Press.
Woodward, J. (2002): “What is a mechanism? A counterfactual account”, Philosophy of Science 69, 366–377.
—— (2003): Making Things Happen: A Theory of Causal Explanation, New York: Oxford University Press.
—— (2014): “A functional account of causation; or, a defense of the legitimacy of causal thinking by
reference to the only standard that matters—usefulness (as opposed to metaphysics or agreement with
intuitive judgment)”, Philosophy of Science 81, 691–713.
130
10
MECHANISMS AND THE
METAPHYSICS OF CAUSATION1
1. Introduction
Whether this volume is your first exposure to the new mechanical philosophy or you have been
contributing to the field for decades, one thing that quickly becomes obvious about the litera-
ture is the great diversity of causal concepts. There is talk of “activities,” “interactions,” “causal
processes,” and “counterfactual difference makers,” to name but a few. These concepts are not
synonymous. They have come to represent different approaches to thinking about causation
in the philosophy of mechanisms, and debate has ensued as advocates of each point out their
preferred approach’s virtues alongside the limitations of the others.
This chapter has two, interrelated purposes. The first is to provide a brief history that traces
these approaches to understanding causation back to their sources. That there are different
approaches with different sources will not be a novel contribution (see, for example, Andersen
2014a, 2014b; Craver and Tabery 2015). However, the tracing will be. That is, we link the
activity approach back to Elizabeth Anscombe, the counterfactual difference-making approach
back to David Lewis, the causal process approach back to Wesley Salmon, and the mechanical
causation approach back to Stuart Glennan, and most philosophers will recognize the influ-
ence of these figures on these approaches. What has not received attention is the way in which
Anscombe, Lewis, Salmon, and Glennan all introduced their ideas about causation explicitly in
response to Hume’s problem of causation and the regularity view of causation that dominated
philosophy after Hume. Anscombe, Lewis, Salmon, and Glennan were all responding to Hume,
but they did so in quite different ways with quite different resulting desiderata. What we will
show is how the differences in those responses and desiderata shaped the subsequent approaches
to understanding causation that entered the new mechanical philosophy.
Our second purpose is to draw on this history to disentangle some of the debates that are
now playing out in the philosophy of mechanisms literature concerning the metaphysics of
causation. It has become common for advocates of these approaches to point out problems
with the others—to complain, for example, that Salmon’s approach doesn’t apply outside fun-
damental physics, or to criticize proponents of activities for not providing a general account of
causation. By tracing the various approaches back to their roles in Anscombe, Lewis, Salmon,
and Glennan’s original responses to Hume, we’ll diagnose how such criticisms often overlook
the different desiderata that shaped those approaches from their very beginning and so hold the
approaches to standards that they were never designed to meet.
131
2. Hume and the problem of causation

Hume, in the Abstract to A Treatise of Human Nature, famously asked his readers to consider a
billiard ball at rest on a table with another ball rolling toward it.
They strike; and the ball, which was formerly at rest, now acquires a motion. This is
as perfect an instance of the relation of cause and effect as any which we know, either
by sensation or reflection. Let us therefore examine it. ’Tis evident, that the two balls
touched one another before the motion was communicated, and that there was no
interval betwixt the shock and the motion. Contiguity in time and place is therefore
a requisite circumstance to the operation of all causes. ’Tis evident likewise, that the
motion, which was the cause, is prior to the motion, which was the effect. Priority
in time is therefore another requisite circumstance in every cause. But this is not all.
Let us try any other balls of the same kind in a like situation, and we shall always find,
that the impulse of the one produces motion in the other. Here therefore is a third
circumstance, viz. that of a constant conjunction betwixt the cause and effect. Every
object like the cause, produces always some object like the effect.
(Hume 2000 [1738–40], p. 409)
Hume was an empiricist, requiring that knowledge be based on experience. When it came to
causation of the sort “A caused B,” all experience told Hume was that A came before B (pri-
ority), that A led immediately to B (contiguity), and that A was always joined by B following
it (constant conjunction). And that’s it. “In whatever shape I turn this matter, and however
I examine it, I can find nothing farther” (ibid). Hume readily acknowledged that his readers
would be inclined to think there was something more to causation—that there must be some
necessary connection to the causal relationship, that B must follow A. Hume countered that such
a necessary connection could not be observed by the empiricist; rather, the apparent sense of
necessity just habitually built up by experiencing the priority, contiguity, and constant conjunc-
tion over and over and over again. This was Hume’s problem of causation—the challenge to
find a necessary connection to causation above and beyond priority, contiguity, and constant
conjunction. Hume thus defined a cause in two ways in the Treatise, one definition pertaining
to a relationship between objects in the world: “An object precedent and contiguous to another,
where all the objects resembling the former are plac’d in like relations of precedency and con-
tiguity to those objects that resemble the latter,” and one definition pertaining to a relationship
between objects in the world and human minds:
A cause is an object precedent and contiguous to another, and so united with it, that
the idea of the one determines the mind to form the idea of the other, and the impres-
sion of the one to form a more lively idea of the other.
(Hume 2000 [1738–40], p. 114)
Hume scholars have debated what Hume himself meant with his definitions of cause—about
the relationship between his two definitions in the Treatise, as well as the relationship between
those definitions and the one he provided in his subsequent An Enquiry Concerning Human
Understanding2 (see, for example, Read and Richman 2000). Hume exegesis, however, is not
the purpose of this chapter. It is sufficient simply to state that a particular view of causation
inspired by Hume dominated philosophy for the next 250 years—the regularity view of causa-
tion, wherein causation was deemed to be patterns of regular succession (patterns of priority,
132
Mechanisms and the metaphysics of causation
contiguity, and constant conjunction). By the mid-twentieth century, this regularity view of
causation faced substantial challenges. It struggled, for example, with cases of common cause
and cases where there was constant conjunction but no causal relation, phenomena recognized
to be common enough in nature that any legitimate view had to accommodate them. Now, to
be sure, philosophers worked hard to bolster the view to surmount the identified weaknesses
(for a review, see Psillos 2009). Still, by the 1970s, the regularity view was sufficiently com-
promised to encourage philosophers to abandon the project entirely and look elsewhere for a
philosophical analysis of causation.
3. Four responses to Hume

Elizabeth Anscombe admitted that Hume upended long-held assumptions about causation;
however, at the same time she thought he reinforced the idea that causation was equated
with necessity by way of his focus on constant conjunction (Anscombe 1993 [1971]).
Indeed, on Anscombe’s reading, the causation-necessitation link only grew post-Hume
as subsequent philosophers (ranging from Kant to Russell) took the Humean challenge to
be finding the necessary connection that was not observable. Anscombe challenged this
causation-necessitation link. Necessity assumed an outcome was predetermined, but the
outcomes of complex systems (such as the presence or absence of infection following expo-
sure to a disease) cannot be specified in advance. In part, Anscombe explained, this was a
feature of humans’ cognitive limitations; we simply cannot keep track of the many con-
tributors to a complex system’s outcome. Additionally, it was a feature of a world governed
in part by indeterministic physics; if the world simply was not predetermined, then no
account of causation that assumed it could be accurate. Anscombe admitted that philoso-
phers and scientists alike pointed to apparent predetermined necessity and made universal
generalizations regarding regularity. But to get necessity in a complex system, Anscombe
argued, measurements of the system had to be made at an abstract enough level to wash
out the underlying indeterminism. And to make universal generalizations work, Anscombe
countered, all sorts of qualifications had to be inserted concerning the system in question
behaving under “normal conditions” (ibid., p. 94).
Importantly, Anscombe’s abandonment of necessity (and, in turn, the search for Hume’s
necessary connection) was not simultaneously an abandonment of a philosophical analysis of
causation. It was just that she didn’t think that the analysis benefited from being tied up with
artificially constructed generalizations and abstracted necessitation. Rather, she advised taking
the “messy and mixed up conditions of life” at face value and building an analysis of causa-
tion off that reality (ibid., p. 95). This involved shifting attention from the development of a
general theory of causation to an understanding of and appreciation for how humans obtained
knowledge of specific causal concepts with specific effects in specific instances, like “scrape, push,
wet, carry, eat, burn, knock over, keep off, squash, make (e.g. noises, paper boats), hurt” (ibid., p. 93,
emphasis in original). All that could be generally said about causation, Anscombe summarized,
was that “causality consists in the derivativeness of an effect from its causes. This is the core, the
common feature, of causality in its various forms. Effects derive from, arise out of, come of, their
causes” (ibid., pp. 91–2). Even this derivativeness, however, was not to be interpreted as a gen-
eral conceptualization of causation; rather, it was a placeholder for specific causings, like pushes
and scrapes. Necessity and universality—the traditional target of philosophical focus—were just
add-ons to this derivativeness according to Anscombe.
David Lewis, just two years after Anscombe, also looked for an alternative to the Humean
regularity view of causation. But he suggested the substitute could actually be found with Hume
133
himself (Lewis 1973). As mentioned above, Hume defined cause twice in his Treatise and then
twice again in his Enquiry (see note 2). In the Enquiry, Hume started by defining the concept
similar to how he defined it in the Treatise: “we may define a cause to be an object, followed by
another, and where all the objects similar to the first, are followed by objects similar to the sec-
ond.” But then he continued with a new addition: “Or in other words, where, if the first object
had not been, the second never had existed.” It was this idea of counterfactually understanding
the difference-making quality of causation that Lewis worked to develop. “We think of a cause
as something that makes a difference, and the difference it makes must be a difference from what
would have happened without it” (Lewis 1973, p. 557).
Lewis drew on possible world semantics to ground a theory of counterfactual difference-
making, where the truth conditions of counterfactuals were based on similarity relations between
possible worlds. The essential idea was that causation could be understood via counterfactual
dependence. To say that “A caused B” was to say that “If A happens, then B happens” and
“Had A not happened, B would not have happened.” Importantly, Lewis thought the counter-
factual dependence understanding of causation could handle the cases that posed problems for
the regularity view. For example, if a common cause (c) was responsible for first one effect (e)
and then another effect (f), the regularity view would interpret the first effect (e) as the cause of
the second (f) because it was prior to, contiguous with, and in constant conjunction with that
effect, even though e and f were in fact both effects of c. Using Lewis’ counterfactual approach,
however, offered a solution to the problem. The common cause c was the cause of both the first
effect e and the second effect f because, had c not occurred, e and f would not have followed;
however, e was not the cause of f because it was possible for e to not have occurred (because of
some breakdown in the relation between c and e) but f to still have occurred.
Whereas Anscombe and Lewis looked for ways around Hume’s challenge to identify the
observable necessity of causation (either by abandoning necessity itself or by turning to the
counterfactual alternative that depended on non-observable possible worlds), Wesley Salmon
instead attempted to take Hume’s empiricist challenge head-on. Starting with a series of papers
in the 1970s and then culminating with his Scientific Explanation and the Causal Structure of the
World, Salmon developed a theory of causation that based causal processes on their ability to
transmit marks (Salmon 1975, 1977, 1984; on Salmon’s indebtedness to Reichenbach for this
focus on transmission, see Williamson 2011). Consider Salmon’s favorite example: a rotating
spotlight in the center of a large, circular room. The light from the spotlight to the wall was a
causal process, but the light circling the room was a pseudo-process; we know this because a red
filter experimentally inserted in the beam itself would alter or mark the transmission from the
spotlight to the wall everywhere, but a red filter experimentally added anywhere on the wall
would not alter or mark the signal except where the light passed over it. Salmon saw this mark-
transmission aspect of causal processes as the key to answering Hume’s challenge: “Ability to
transmit a mark can be viewed as a particularly important species of constant conjunction—the
sort of thing Hume recognized as observable and admissible. It is a matter of performing certain
kinds of experiments” (Salmon 1977, p. 220).
Salmon’s claim to have answered Hume’s challenge made it an exciting development
for philosophers but also a target for criticism. The most severe charge against the mark-
transmission solution was that it was simply a counterfactual theory in disguise because the
only way to distinguish the causal processes from the pseudo-processes was to counterfactu-
ally assess what would have happened under experimental intervention (Kitcher 1984; Dowe
1992). Salmon was repulsed by this idea; unobservable counterfactuals were not empiricist in
nature, and so Salmon could not rightly claim to have met Hume’s challenge on Hume’s terms
if he relied on counterfactuals. Fortunately, one of his critics also offered a solution. Phil Dowe
134
suggested switching from the transmission of marks to the transmission of conserved quantities
(Dowe 1992). Salmon was receptive to the shift. His mark-transmission reliance on experiments
revealed how to identify causal processes, but it could not explicate the very concept of a causal
process that he sought (Salmon 1994, p. 303). The updated Salmon-Dowe account of causa-
tion linked causal processes with the transmission of conserved quantities, such as momentum,
mass-energy, and electrical charge. Basing the account on actual, physical things—the conserved
quantities—led Salmon to believe that he had avoided counterfactuals and ultimately met Hume
on his empiricist terms:
the idea was to present a “process theory” of causality that could resolve the fun-
damental problem raised by Hume regarding causal connections. The main point is
that causal processes, as characterized by this theory, constitute precisely the objective
physical causal connections which Hume sought in vain.
(ibid., p. 297; see also Salmon 1997, 1998)
Stuart Glennan’s “Mechanisms and the Nature of Causation” (1996) is rightly recognized as one
of the earliest contributions to the new mechanical philosophy; indeed, it is in that paper that
we receive the first definition of a mechanism by a new mechanical philosopher. What has not
been emphasized, however, is the fact that Glennan introduced that definition and his discus-
sion of mechanisms more generally with a very specific purpose in mind—answering Hume’s
problem. Hume, recall, challenged philosophers to find the necessary connection between cause
and effect that was above and beyond repeated exposures to priority, contiguity, and constant
conjunction. In response, Glennan argued that two events are causally connected when there is
a mechanism that connects them (Glennan 1996, p. 64). To use his example: we do not think
turning the key in an automobile’s ignition causes the engine to turn over because we have been
exposed to that process over and over again; rather, we think turning the key causes the engine
to turn over because we believe that there is a mechanism in the automobile that brings about
the ignition of the engine.
It was because a mechanism was at the heart of causal connection that Glennan introduced
his definition of the concept: “A mechanism underlying a behavior is a complex system which
produces that behavior by of [sic] the interaction of a number of parts according to direct causal
laws” (ibid., p. 52). Phenomena ranging from blood flow in the human body to the regulation of
water level in a toilet, Glennan pointed out, can be causally understood as products of complex
systems whose interacting parts produce the phenomena in question. Now, Glennan admit-
ted that this process of understanding phenomena at one level by appeal to the mechanism(s)
responsible for that phenomena at a lower level eventually bottomed out at the fundamental
laws of physics, which seemed to defy further mechanical decomposition. But, Glennan coun-
tered, the vast majority of causal reasoning (both scientific and in everyday life) took place above
this fundamental level. So while Glennan was willing to grant that, “At the level of fundamental
physics, Hume’s problem still remains,” he also claimed that, with his mechanical theory of
causation, “Hume’s problem is not a universal one” (ibid., p. 68). At higher levels, where regu-
larities could be mechanically explicated, mechanisms offered the necessary connection.
Anscombe, Lewis, Salmon, and Glennan all shared dissatisfaction with the Humean regular-
ity view of causation and all offered up alternatives to that view. Those alternatives, however,
were quite unique from one another, and they resulted in different desiderata that judged
the success of an account of causation. Salmon and Glennan both demanded that an account
of causation address Hume’s problem of causation by identifying the necessary connection.
Salmon offered this by going all the way down to the level of fundamental physics with the
135
transmission of conserved quantities. Glennan, in contrast, stopped just short of fundamental

physics, arguing that mechanisms connected causes and effects at higher levels. Anscombe and
Lewis, each in their own way, dismissed Hume’s challenge. Anscombe replaced the search
for a general account of causation’s necessary connection with a singularist search for the way
that specific effects are brought about by specific causings. Lewis instead switched the focus to
Hume’s remark about the difference-making quality of causation, demanding of an account of
causation that it capture this feature (which he did with counterfactual dependence).
At the most basic level, the differences between the four approaches can be seen by returning
to Hume’s famous billiard balls. For Anscombe, the causal relation was one of derivativeness in
that instance; the first ball struck the second, and the second ball’s motion derived from or arose
out of the first ball’s striking. It was about counterfactual difference-making for Lewis; had the
first ball not struck the second, the second ball would not have moved. For Salmon it was about
the transmission of a conserved quantity; the first ball transmitted an actual, physical thing to the
second ball. And for Glennan the entire set-up could be understood as a complex system with
parts (e.g. billiard balls) and interactions (e.g. a collision) that are mechanically explicable, and
where the causal connection is a function of that mechanism.
4. Derivativeness, counterfactual difference-making, causal processes,

and mechanical causation in the philosophy of mechanisms
The new mechanical philosophy emerged around the turn of the twenty-first century, as philos-
ophers of science looked for an alternative framework to logical empiricism. Logical empiricists
addressed a range of philosophical issues by formally characterizing the logical and mathematical
structures constitutive of science, and by focusing on the abstract and epistemic features of sci-
ence. New mechanical philosophers instead focused on detailed case studies of actual scientific
practice, and drew attention to the material elements of the world as the object of scientific
inquiry. What emerged was a new framework for thinking about the classic issues in philosophy
of science (Craver and Tabery 2015). Fundamental to this shift was the idea that science is in
the business of elucidating the mechanisms that are causally responsible for phenomena under
investigation. As a result, considerations of causation have been embedded in the new mechani-
cal philosophy from its beginning; however, different contributors to the framework have made
sense of that causation in different ways.
Anscombe’s approach to causation first appeared in the new mechanical philosophy with
the publication of Peter Machamer, Lindley Darden, and Carl Craver’s “Thinking about
Mechanisms” (2000). “Mechanisms,” they said, “are entities and activities organized such that
they are productive of regular changes from start or set-up to finish or termination conditions”
(ibid., p. 3). Anscombe, recall, dismissed the search for a general account of causation and instead
encouraged turning to specific instances where effects derive from their causes—scrapes, pushes,
burns. These are Machamer, Darden, and Craver’s activities; they are types of causings (ibid., p. 6).
In the ensuing years, a number of philosophers worked to develop this Anscombian-inspired
concept of an activity. Machamer (2004) picked up on Anscombe’s reference to child develop-
ment of verb-acquisition, pointing to research that explored how children procure concepts
of things by way of interacting with and manipulating what those things do (i.e. the enti-
ties’ activities). Bogen (2005, 2008) alternatively emphasized Anscombe’s dismissal of necessity,
highlighting examples of research where the elucidation of mechanisms provided explanations
even if those mechanisms did not operate with generalizable regularity. Darden (2006) docu-
mented ways in which the study of activities figured into scientific discoveries, especially in
molecular biology and genetics.
136
Despite these differences in foci, what Machamer, Bogen, and Darden all shared was an
Anscombian abandonment of the search for a general theory of causation, in favor of atten-
tion to how the understanding of specific activities figured into specific causal mechanisms.3 As
Machamer put it:
The problem of causes is not to find a general and adequate ontological or stipulative
definition, but a problem of finding out, in any given case, what are the possible, plau-
sible, and actual causes at work in any given mechanism.
(Machamer 2004, pp. 27–8)
Or take Bogen: “Instead of suggesting a general, uniformly applicable, answer to the question of
what differentiates causes from non-causes, what I take to be Anscombe’s view calls for piecemeal
treatments of questions about how specific effects are produced” (Bogen 2008, p. 114). And
Darden, referencing the productive nature of activities, wrote, “Rather than seeking a general
definition of production, it is more insightful to consider specific kinds of activities and the means
for discovering them” (Darden 2013, p. 25).
A difference-making understanding of causation made its way into the philosophy of
mechanisms by way of James Woodward’s manipulationist understanding of causal relevance
(Woodward 2003). The basic idea for Woodward is that scientists causally explain when they
know how to manipulate. The manipulations are understood counterfactually and compara-
tively. If some particular variable is a cause of some outcome, then manipulating the value of
the variable would be a way of manipulating the outcome. These counterfactual experiments
formulate and then answer what-if-things-had-been-different questions; and, in so doing, they
establish a pattern of counterfactual dependence that determines what difference a cause makes.
Counterfactual dependence is understood with the closely related concepts of intervention and
invariance. An intervention consists of an idealized experimental manipulation of the value of
some variable, thereby determining if it results in a change in the value of the outcome. So
the counterfactuals are formulated in such a way that they show how the value of the out-
come would change under the interventions that change the value of a variable. Invariance,
then, is a characterization of the relationship between variables (or a variable and an outcome)
under interventions on Woodward’s account. When there is an invariant relationship between
a variable and an outcome, then that relationship is potentially exploitable for manipulation, and
because of this it is a causal relationship.
Woodward himself applied his manipulationist framework to the new mechanical phi-
losophy, building his invariance-under-intervention formulation into a characterization of
mechanistic causation where “the behavior of each component is described by a generali-
zation that is invariant under interventions” (Woodward 2002, S375). The move brought
an account of how the difference-making quality of causal claims could be understood in
sciences that elucidate mechanisms; moreover, the concept of invariance and degrees in it
was attractive to philosophers because it offered a resource for characterizing mechanis-
tic regularities that were robust but not law-like universal generalizations. Woodward’s
manipulationist framework has influenced a number of philosophers of mechanism; for
example, Craver took Woodward’s approach to understanding causal relevance in an etio-
logical system and updated it to account for constitutive relevance in a multi-level system,
where the idea was that coordinated top-down and bottom-up interventions could establish
constitutive relevance by way of mutual manipulability (Craver 2007).
Woodward’s manipulationist framework certainly introduced a counterfactual difference-
making understanding of causation into the philosophy of mechanisms. But it is not obviously
137
a metaphysical understanding. That is, Lewis’ development of possible worlds and modal realism
about those possible worlds was an explicit attempt to build a metaphysics of counterfactual cau-
sation. But Woodward’s manipulationist framework is not so easily categorized as such. His is an
account of causal relevance and the role of causal relevance in causal claims. As such, it attends
more to the semantics and epistemology of causation than it does to a deeper metaphysics.
Indeed, even Craver, in his embracing of Woodward’s manipulationist approach, was careful to
say that he was using Woodward’s understanding of the difference-making quality of causation
because of its epistemological implications for the role of causal relevance in scientific explana-
tions, not for its metaphysical solution to Hume’s problem (Craver 2007, pp. 105–6, 225–6).4
Salmon’s development of the concept of a causal process was part of his more general revolu-
tion in philosophy of science aimed at reorienting philosophers toward the causal structure of the
world. When it came to questions of causation and answering Hume’s problem, as we saw above,
he appealed to the transmission of conserved quantities. When it came to questions of explana-
tion, he encouraged abandoning the epistemic conception that appealed to laws of nature and
derivation of phenomena from those laws of nature, and replacing it with the ontic conception of
explanation that situated phenomena in the physical structure of the world (Salmon 1984, 1998).
Although Salmon never identified with the new mechanical philosophy, his push for a phi-
losophy of science that thought about causation and explanation in terms of being physically
situated in the causal structure of the world certainly shaped the subsequent philosophy of
mechanisms (Campaner 2013). For example, Salmon’s focus on causal processes placed a prior-
ity on being able to trace the causal chain from the causal structure of the world through to
the phenomenon of interest; this attention to tracing a causal chain can be seen in Machamer,
Darden, and Craver’s (2000) emphasis on productive continuity which renders intelligible the
various stages of a causal mechanism. Salmon’s ally—Phil Dowe—most explicitly applied their
causal process approach to mechanisms, generating an account of “causal process mechanisms”
(Dowe 2011). On this account, mechanistic causation is the transmission of conserved quanti-
ties within and by a mechanism, and so an ontic explanation is provided by the elucidation
of a causal process mechanism when the propagation of the conserved quantities through the
mechanism can be traced.
Glennan, in contrast to Anscombe, Lewis, and Salmon, is unique in that he both replied to
Hume and then personally took that reply and incorporated it into the philosophy of mecha-
nisms. Indeed, Glennan has wrestled with the metaphysics of causation perhaps more than
any other new mechanical philosopher (see, for example, Glennan 2002, 2009, 2010a, 2010b,
forthcoming; Kuhlmann and Glennan 2014). Certain aspects of his account have evolved over
the years. For example, as described earlier, in his 1996 definition of a mechanism he based the
interaction of parts on “direct causal laws.” But then, in his 2002, those interactions were based
on “direct, invariant, change-relating generalizations.” This, however, was more of a termino-
logical switch than a deep philosophical concession; Glennan used “laws” loosely in his 1996
to describe a range of generalizations, both the mechanically inexplicable ones of fundamental
physics and the mechanically explicable ones of higher levels. And so, while Glennan dropped
the philosophically loaded term “laws” in 2002, he was still committed to the basic idea that
many generalizations above the level of fundamental physics were mechanically explicable ones.
And, most important, he stayed true to the idea that causes and effects were causally connected
when a mechanism links the cause to the production of the effect.
In more recent years, Glennan turned to considering the relationship between accounts of
causation that focused on causal production (like his, Anscombe’s, and Salmon’s) and accounts of
causation that focused on causal relevance (like Lewis’ and Woodward’s). Mechanisms remained
at the heart of the account; causal production was brought about by way of a mechanism that
138
linked a cause with an effect. But the mechanism also provided the foundation for claims about
causal relevance; relevant properties embedded in the mechanism determined in what sense
the causal production was difference-making (Glennan 2009, 2010a, forthcoming). “On the
account I am offering,” Glennan summarized, “productivity and relevance are not compet-
ing approaches to analyzing the nature of causal relationships, but rather are complementary
concepts which refer to different features of the causal structure of the world” (Glennan forth-
coming; see also Chapter 7 of this volume).
5. Assessing the contemporary debates

Tracing the current approaches to understanding causation in the mechanistic literature back to
the unique responses to Hume by Anscombe, Lewis, Salmon, and Glennan helps evaluate recent
criticisms of those approaches. It becomes evident that mechanistic accounts of causation are
often held to desiderata that were never intended to be met. For example, some new mechani-
cal philosophers have demanded that an account of causation be non-reductionistic. Not to be
confused with non-reductive (which we’ll discuss later in this section), a non-reductionistic account
of causation explicitly avoids reducing talk of “causes” across the sciences to the lowest levels of
analysis. Causal process approaches to causation in the mechanisms literature, that is, have been
criticized on the grounds that they fail to account for how causation is understood across all the
sciences—especially those beyond the level of fundamental physics. This non-reductionistic desid-
eratum becomes most evident in light of how the new mechanists have responded to Salmon’s
causal process approach. A number of new mechanists, for example, have complained that the
focus on the transmission of conserved quantities cannot explain causal language outside the
domain of fundamental physics. Machamer, Darden, and Craver (2000), for example, argued
that, although Salmon’s analysis captures causal interactions in fundamental physics, it “is silent as
to the character of the productivity in the activities investigated by many other sciences” (p. 7).
On a similar point, Glennan (2002) argued that, while Salmon’s account “has the advantage of
characterizing interactions in terms of physical theory rather than the semantics of counterfac-
tuals, it has the disadvantage that it obscures similarities between kinds of interactions among
higher-level entities” (S346). Glennan challenged readers to explain commonplace causal inter-
actions by appeal to the exchange of conserved quantities alone. Informed by investigative and
explanatory practices in modern neuroscience, Craver (2007) argued that when “one begins to
talk about causal relations that arise when parts are organized into mechanisms, the transmission
view loses traction; its austere descriptive vocabulary no longer applies” (p. 76).
On this criticism of Salmon’s approach that reveals the non-reductionistic desideratum, perhaps
Jon Williamson (2011) said it best:
conserved quantities are too low-level, far removed from most of our causal claims.
Consider a claim like eliminating the 10% tax band caused an increase in inequality:
not only does such a claim not explicitly mention conserved quantities, it is very hard
to see how it could be about conserved quantities at all. Most of our causal claims are
high-level claims like this, and it appears that the process theory has some work to do
before it can provide a convincing interpretation of such claims.
(Williamson 2011, p. 427)
Fair enough. But remember that Salmon’s goal was to develop a notion of causal process that
provided the necessary connection Hume said he could not observe. The goal wasn’t to develop
a notion of causal process that accounts for causal mechanical explanations across the sciences.
139
So this criticism holds Salmon’s account of causation to a specific desideratum—non-reductionistic—

that he never intended to meet. One could certainly say that Dowe’s causal process mechanism
restricts talk of mechanistic causation to cases where conserved quantities can be traced, thus leaving
out mechanistic causation in a higher-level science like macro-economics. But then Dowe could
simply respond that macro-economics doesn’t provide any understanding of mechanistic causation.
Alternatively, consider criticisms of the activity-based approach to causation in the mecha-
nisms literature. A number of new mechanists have criticized the activity-based approach on
the grounds that all it does is list a series of examples such as “diffuse,” “push,” and “bond,” or
vaguely say only that activities are the “producers of change.” Tabery (2004), for example, wor-
ried that introduction of the concept of activity fails to inform a general view of how activities
actually produce changes in the world (beyond listing examples of change-production). Psillos
(2004) made a similar point, highlighting that because production itself is an activity, then “we
are not given an illuminating account of that which some things share in common, in virtue of
which they are activities” (pp. 311–12). These criticisms of activity-based accounts of causation
hint at a specific desideratum regarding what one might expect out of a promising metaphysical
account of causation: conceptual generality. The thought is that any account of causation should
be expected to say what causation as a general concept is, to say what distinguishes all causings
from all non-causings.
Remember, however, that the activity-based approach to understanding causation is a direct
descendant of Anscombe’s response to Hume, and fundamental to that response was the aban-
donment of any attempt to generate a general account of causation beyond specific instantiations
of it (like diffusion, pushing, and bonding). In effect, abandonment of Hume’s call to provide
a general account of causation is precisely the Anscombian move. So, again, holding activity-
advocates to the conceptual generality desideratum that they’ve never acknowledged charges them
with breaking a rule in a game they’ve explicitly refused to play.
Counterfactual difference-making approaches to causation in the mechanisms literature
have also been held to an unacknowledged desideratum. From Salmon to contemporary new
mechanical philosophers, a number of authors have argued that a fault with the counterfactual
approach to causation is its understanding of causation in terms of what could have been, rather
than in terms of what actually is or was (Salmon 1984, 1998). For example Machamer (2004)
argued that “only ontological (or even ontic) principles or descriptions belong in a ‘real’ defini-
tion of causality” (p. 28). In other words, a satisfactory understanding of causation should be
grounded in terms of what actually exists in the world, as opposed to what could, would, or
might pertain to the actual world. Bogen (2004, 2005, 2008) criticized counterfactual depend-
ence on similar grounds, arguing that while it may be the case that counterfactual reasoning
influences experimental theorizing in scientific practice, it does not follow that causal claims are
actually grounded in counterfactuals. Ultimately, the truth of causal claims in scientific practice
is grounded in what actually happens—not what would have happened:
But it does not follow from the fact that regularities are of great epistemic, theoretical
and practical interest that actual or counterfactual regularities are constitutive of the
ontological or conceptual difference between the causes of an effect and the non-
causal items that accompany its production.
(Bogen 2008, p. 122)
Further, and in a very similar vein, Waskan (2011) defended an explicitly non-counterfactualist
mechanistic (what he called an “actualist-mechanist”) account of causation in mechanisms. What
is actual, the thought went, is more fundamental than what counterfactually could have been, and
140
so an approach to understanding causation that settles for counterfactual dependence is missing

something essential to causation.
We will follow Waskan in naming this the actualism desideratum of causation. Machamer,
Bogen, and Waskan (as well as Salmon before them) all demanded that a metaphysical account
of causation be based on actual causal connections between actual causes and actual effects.
Here the truth-makers for counterfactual causes are actual features of the world. The prob-
lem with this criticism, however, is that it charges counterfactualists with missing something
essential to causation, when in fact counterfactualists simply have a different idea about what
is essential to causation. Ever since Lewis, employers of the counterfactual approach take cau-
sation to be essentially about difference-making, and so the various formulations of it (from
possible world semantics to manipulationist interventions) have been designed to make sense
of this difference-making quality of causation. On this particular disagreement it is important
to remember that the actualism desideratum only works against counterfactual approaches to
causation that hold the truth-makers of counterfactual claims to be features of other/possible/
non-actual worlds; alternatively, if the truth-makers are based on features of the actual world,
then the actualism desideratum still holds (see the discussion of causal pluralism in Glennan
forthcoming for an example).
Lastly, consider a common criticism of Glennan’s mechanical theory of causation. Most
notably, his view that causal connections exist in virtue of mechanisms has been criticized on
the grounds that it is uninformative because it fails to offer an analysis of causation that does not
ultimately rely on the concept of “cause” itself. Kistler (2009), for example, argues that,
On closer inspection, it appears that the concept of mechanism presupposes that of

causation, far from being reducible to it. . . . The crucial point is that each step of the
analysis of a mechanism makes essential use of the notion of cause, and thus presup-
poses it. . . . It follows that the concept of mechanism cannot be used to analyse the
concept of causation and that, quite on the contrary, the concept of causation is among
the irreducible conceptual instruments of mechanistic analysis.
(pp. 599–600)
This criticism of Glennan’s mechanistic analysis of causation appeals to the reductive desideratum
(not to be confused with reductionistic). The concern, that is, is that the mechanistic account fails
to reduce our understanding of causation to something other than causation.
True, the mechanistic account is non-reductive—but it is not clear that Glennan is moved by
this desideratum. Glennan (forthcoming, chapter 6) responds to objections of this ilk by openly
endorsing a non-reductive analysis of causation, arguing that more important than being reduc-
tive is being informative. While it may be the case that the concept of causation is not reduced
by Glennan’s mechanistic approach, that is, it does not follow that it fails to further develop
our philosophical understanding of the causal structure of the world. On this view, the intricate
mechanical story that connects causes and effects—the orchestration of organized parts, engaged
in causally productive activities via mechanisms—is explanatory, despite the fact that it is non-
reductive. Thus there is a sense in which even Glennan’s mechanistic account of causation is
criticized for failing to meet a desideratum that he never intended to meet.
6. Conclusion
The philosophy of mechanisms presents a new framework for thinking about classic issues in the
philosophy of science. Under that general framework, however, there are a number of debates.
141
One of those debates pertains to how mechanistic causation should be understood, and because
causation is fundamental to the philosophy of mechanisms, this is a debate that has repercussions
throughout the framework.
We have showed how different approaches to understanding causation in the phi-
losophy of mechanisms trace back to four separate responses to Hume’s problem of
causation: Anscombe’s derivativeness approach, Lewis’ counterfactual difference-making
approach, Salmon’s causal process approach, and Glennan’s mechanical causation approach.
Recognizing the relationship between current approaches in the philosophy of mechanisms
and their historical ancestors best reveals the differences between those approaches and
where they came from. It also allows for diagnosing when criticisms of a certain approach
mistakenly hold that approach to a standard it was never designed to meet. As the debate
over the metaphysics of mechanistic causation continues, be it in the direction of advocat-
ing for one approach in the face of the others or in the direction of finding common ground
among the approaches, it is essential to keep this history in mind, as it offers a guide to
where the fundamental differences between the approaches lay.
Notes
1 We are grateful to Jim Bogen, Stuart Glennan, and Phyllis Illari for providing us with valuable feedback
on an earlier draft of this chapter.
2 In the Enquiry, Hume defines cause as such: “Similar objects are always conjoined with similar. Of this we
have experience. Suitably to this experience, therefore, we may define a cause to be an object, followed
by another, and where all the objects similar to the first, are followed by objects similar to the second. Or
in other words, where, if the first object had not been, the second never had existed. The appearance of
a cause always conveys the mind, by a customary transition, to the idea of the effect. Of this also we have
experience. We may, therefore, suitably to this experience, form another definition of cause; and call it,
an object followed by another, and whose appearance always conveys the thought to that other” (Hume
1993 [1748], p. 51).
3 Anscombe’s influence reached out beyond the philosophical debates about causation. Nancy Cartwright’s
well-known book How the Laws of Physics Lie was very much a development of Anscombe’s point about
generalizations working only on artificial, idealized systems (Cartwright 1983). Because this chapter is
about causation and not about laws of nature, we leave these other legacies of Anscombe’s influence aside.
4 We are particularly grateful to Stuart Glennan for driving this point home.
References
Andersen, Holly (2014a), “A Field Guide to Mechanisms: Part I”, Philosophy Compass 9: 274–283.
—— (2014b), “A Field Guide to Mechanisms: Part II”, Philosophy Compass 9: 284–293.
Anscombe, G. E. M. (1993 [1971]), “Causality and Determination”, in Ernest Sosa and Michael Tooley
(eds.), Causation. Oxford: Oxford University Press, pp. 88–104.
Bogen, Jim (2004), “Analysing Causality: The Opposite of Counterfactual is Factual”, International Studies
in the Philosophy of Science 18: 1–26.
—— (2005), “Regularities and Causality; Generalizations and Causal Explanations”, Studies in the History
and Philosophy of the Biological and Biomedical Sciences 36: 397–420.
—— (2008), “Causally Productive Activities”, Studies in the History and Philosophy of Science 39: 112–123.
Campaner, Raffaella (2013), “Mechanistic and Neo-mechanistic Accounts of Causation: How Salmon
Already Got (Much of) It Right”, Metatheoria 3: 81–98.
Cartwright, Nancy (1983), How the Laws of Physics Lie. Oxford: Oxford University Press.
Craver, Carl F. (2007), Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience. Oxford: Oxford
University Press.
Craver, Carl and James Tabery, (2015) “Mechanisms in Science”, The Stanford Encyclopedia of Philosophy
(Winter Edition), Edward N. Zalta (ed.), URL = <http://plato.stanford.edu/archives/win2015/
entries/science-mechanisms/>.
142
Darden, Lindley (2006), Reasoning in Biological Discoveries: Essays on Mechanisms, Interfield Relations, and
Anomaly Resolution. Cambridge: Cambridge University Press.
—— (2013), “Mechanisms Versus Causes in Biology and Medicine”, in Hsiang-Ke Chao, Szu-Ting Chen,
and Roberta Millstein (eds.), Mechanisms and Causality in Biology and Economics. Dordrecht: Springer,
pp. 19–34.
Dowe, Phil (1992), “Wesley Salmon’s Process Theory of Causality and the Conserved Quantity Theory”,
Philosophy of Science 59: 195–216.
—— (2011), “The Causal-Process-Model Theory of Mechanisms”, in Phyllis McKay Illari, Federica
Russo, and Jon Williams (eds.), Causality in the Sciences. Oxford: Oxford University Press, pp. 865–879.
Glennan, Stuart (1996), “Mechanisms and the Nature of Causation”, Erkenntnis 44: 49–71.
—— (2002), “Rethinking Mechanistic Explanation”, Philosophy of Science 69: S342–S353.
—— (2009), “Productivity, Relevance, and Natural Selection”, Biology & Philosophy 24: 325–339.
—— (2010a), “Mechanisms, Causes, and the Layered Model of the World”, Philosophy and Phenomenological
Research 81: 362–381.
—— (2010b), “Ephemeral Mechanisms and Historical Explanation”, Erkenntnis 72: 251–266.
—— (Forthcoming), The New Mechanical Philosophy. Oxford: Oxford University Press.
Hall, Ned (2004), “Two Concepts of Causation”, in John Collins, Ned Hall, and L. A. Paul (eds.),
Causation and Counterfactuals. Cambridge: The MIT Press, pp. 225–276.
Hume, David (2000 [1738–40]), A Treatise of Human Nature: Being an Attempt to Introduce the Experimental
Method of Reasoning into Moral Subjects. Oxford: Oxford University Press.
—— (1993 [1748]), An Enquiry Concerning Human Understanding, second edition. Indianapolis: Hackett
Publishing Company.
Kistler, M. (2009), “Mechanisms and Downward Causation”, Philosophical Psychology 22: 595–609.
Kitcher, Philip (1984), “Explanatory Unification and the Causal Structure of the World”, in Philip Kitcher
and Wesley C. Salmon (eds.), Scientific Explanation. Minneapolis: University of Minnesota Press,
pp. 410–505.
Kuhlmann, Meinard and Stuart Glennan (2014), “On the Relation between Quantum Mechanical and
Neo-Mechanistic Ontologies and Explanatory Strategies”, European Journal for the Philosophy of Science
4: 337–359.
Lewis, David (1973), “Causation”, The Journal of Philosophy 70: 556–567.
Machamer, Peter (2004), “Activities and Causation: The Metaphysics and Epistemology of Mechanisms”,
International Studies in the Philosophy of Science 18: 27–39.
Machamer, Peter, Lindley Darden, and Carl Craver (2000), “Thinking about Mechanisms”, Philosophy of
Science 67: 1–25.
Psillos, Stathis (2004), “A Glimpse of the Secret Connexion: Harmonizing Mechanisms with
Counterfactuals”, Perspectives in Science 12: 288–319.
—— (2009), “Regularity Theories”, in Helen Beebee, Christopher Hitchcock, and Peter Menzies (eds.),
The Oxford Handbook of Causation. Oxford: Oxford University Press, pp. 131–157.
Read, Rupert and Kenneth A. Richman, eds. (2000), The New Hume Debate. London: Routledge.
Salmon, Wesley C. (1975), “Theoretical Explanation”, in Stephan Körner (ed.), Explanation. Oxford: Basil
Blackwell, pp. 118–145.
—— (1977), “An ‘At-At’ Theory of Causal Influence”, Philosophy of Science 44: 215–224.
—— (1984), Scientific Explanation and the Causal Structure of the World. Princeton, NJ: Princeton University
Press.
—— (1994), “Causality without Counterfactuals”, Philosophy of Science 61: 297–312.
—— (1997), “Causality and Explanation: A Reply to Two Critics”, Philosophy of Science 64: 461–477.
—— (1998), Causality and Explanation. Oxford: Oxford University Press.
Tabery, James (2004), “Synthesizing Activities and Interactions in the Concept of a Mechanism”, Philosophy
Waskan, Jonathan (2011), “Mechanistic Explanation at the Limit”, Synthese 183: 389–408.
Williamson, Jon (2011), “Mechanistic Theories of Causality: Part I”, Philosophy Compass 6: 421–432.
Woodward, James (2002), “What Is a Mechanism? A Counterfactual Account”, Philosophy of Science 69:
S366–S377.
—— (2003), Making Things Happen: A Theory of Causal Explanation. Oxford: Oxford University Press.
—— (2011), “Mechanisms Revisited”, Synthese 183: 409–427.
143
11
MECHANISMS,
COUNTERFACTUALS,
AND LAWS1
1. Introduction
There have been two traditions concerning how the “link” between cause and effect is best
understood (Hall 2004; Psillos 2004). According to the first tradition, which goes back to
Aristotle, there is a productive relation between cause and effect: the cause produces, generates,
or brings about the effect. This productive relation between cause and effect has been typically
understood in terms of powers, which in some sense ground the bringing about of the effect
by the cause. According to the second tradition, which goes back to Hume, the link is some
kind of robust relation of dependence between what are taken to be distinct events. On this
account, the chief characteristic of causes is that they are difference-makers: the occurrence
of the cause makes a difference to the occurrence of the effect (for Hume’s theory of causa-
tion, see Chapter 10). There are various ways to understand the notion of difference-making
(e.g. in terms of laws or probabilities); but arguably the core notion of difference-making is
counterfactual, i.e. based on contrary-to-fact hypotheticals. That is, a causal claim of the form
“A caused B” would be understood as implying: if A hadn’t happened, B wouldn’t have hap-
pened either. It is in this sense that A actually makes a difference for B.2
The currently most popular version of the production approach cashes out the link between
cause and effect by reference to mechanisms. The central thought behind a mechanistic account
of causal production is that two events are causally connected, if and only if there is a mecha-
nism that connects them. Hence, where there is causation, there is mechanism. As we shall see,
however, various mechanists tie together a power-based account of production and a mecha-
nistic account. In this chapter, we shall focus our attention on mechanisms and aim to compare
mechanistic accounts with counterfactual accounts of causation.
As we will see, some mechanists tend to refrain from using counterfactuals. For others,
counterfactuals are needed to ground the laws that characterize the interactions between the com-
ponents of a mechanism; counterfactuals may in turn be grounded in lower-level mechanisms. Yet
other mechanists try to dispense with both counterfactuals and laws, in favor of activities. Hence,
understanding the relation between mechanisms and counterfactuals requires clarifying the rela-
tion between mechanisms and laws (on the relationship between mechanisms and laws, see also
Chapter 12). Laws will thus be central in the argument of this chapter. The key question then,
144
Mechanisms, counterfactuals, and laws
for our purposes, is: can there be a conception of mechanism which does not ineliminably rely on
some non-mechanistic account of counterfactual dependence?
To be exact, we want to investigate whether a mechanistic theory of causation ultimately
relies on a counterfactual theory (and hence, whether it turns out to be a version of the
dependence approach); or whether it constitutes a genuine version of the production approach
(either because it altogether dispenses with the need to rely on counterfactuals, or because it
grounds counterfactuals in mechanisms). To be clear on these issues presupposes, as we will
argue in section 2, a more careful analysis of the notion of mechanism. Here is the central
line of argument we want to investigate: since a mechanism is composed of interacting compo-
nents, the notion of a mechanism should include a characterization of these interactions; but if
(i) these interactions are understood in terms of difference-making relations and (ii) these
difference-making relations are not in turn grounded in mechanisms, then there is a fundamen-
tal asymmetry between mechanistic causation and causation as difference-making; for to offer
an adequate account of the former presupposes an account of the latter.
As we will see, not all philosophers that stress the importance of mechanisms for thinking
about science are after an account of causation. For some of them, mechanisms are important in
understanding scientific explanation and theorizing, but it is not the case that causation itself is
mechanistic (see, for example, Craver 2007, 86). Yet even if these philosophers do not have to
provide a full-blown account of the ontology of mechanisms, they have to explain the modal
force of mechanisms; hence the issue of the relation between mechanisms and counterfactuals
is crucial. So, we can formulate our central question in a more comprehensive way, as follows:
given that a mechanism consists of components that interact in some manner, and thus cause
changes to one another, does an account of these interactions require a commitment to counter-
factuals? As we shall see, ultimately, the issue turns not around the need or not to posit relations
of counterfactual dependence but around what the suitable truth-makers for counterfactuals are.
2. Mechanisms-for vs mechanisms-of
Two traditions have tried to reclaim the notion of mechanism in the philosophical literature of
the twentieth century. In exploring the relation between mechanisms and laws/counterfactuals,
it is important to distinguish between these two very different senses of “mechanism.”
In the fairly recent literature, mechanisms are always understood as mechanisms for certain
behaviors (see Chapter 8). In other words, mechanisms are individuated in terms of what they
do. For example, there are mechanisms for DNA replication, or for mitosis. What the mechanism
does, what the mechanism is a mechanism for, determines the boundaries of the mechanism and
the identification of its components and operations (see Chapter 9 for more on components and
boundaries of mechanisms). Such mechanisms, which we call mechanisms-for (i.e. mechanisms
for certain behaviors/functions), are the mechanisms that, according to many authors, figure in
explanations in biomedical sciences and elsewhere, and are what many scientists aim to discover.
Mechanisms-for we find, among others, in Machamer, Darden, and Craver (2000), Bechtel and
Abrahamsen (2005), Craver (2007), and Bechtel (2008). This kind of conception of mechanisms
is, arguably, the dominant one in various philosophical studies of mechanisms and their role in
the various sciences.3
The second sense of mechanism is typically found in the context of mechanistic theories
of causation. These theories aim to characterize the causal link between two events (to fathom
Hume’s “secret connexion”) in terms of a “mechanism.” For this second sense, what the
145
mechanism does is not important; what is important is that it is actually there underlying or
constituting a certain kind of process. More precisely, what makes a process causal is the pres-
ence of a mechanism which mediates between cause and effect (or whose parts or moments
are the “cause” and the “effect”). We call such mechanisms mechanisms-of. Mechanisms-of are
the mechanisms discussed in, for example, theories that view causation as mark-transmission
(Salmon 1984), persistence, transference, or possession of a conserved quantity (Mackie 1974;
Salmon 1997; Dowe 2000).
Talk of “mechanisms” in relation to causation goes back to John Mackie (1974), who took
it that causation consists in a “causal mechanism”; that is, “some continuous process connecting
the antecedent in an observed . . . regularity with the consequent” (1974: 82). His preferred
account of a causal mechanism in terms of qualitative or structural continuity, or persistence,
exhibited by certain processes, faced significant problems which led Wesley Salmon (1984) to
argue for an account of causal mechanism that is based on the notion of structure-transference
(see Psillos 2002: section 4.1; also Chapter 10 of this book). Salmon kept the basic idea that
“[c]ausal processes, causal interactions, and causal laws provide the mechanisms by which the
world works; to understand why certain things happen, we need to see how they are produced
by these mechanisms” (1984: 132). But he claimed that the distinguishing characteristic of a
causal process (and hence of a mechanism) is that it is capable of transmitting its own structure
or modifications of its own structure. Generalizing Hans Reichenbach’s (1956) idea that causal
processes are those that are capable of transmitting a mark, Salmon noted that any process, be it
causal or not, exhibits “a certain structure.”
A causal process is then a process capable of transmitting its own structure. But, Salmon
added, “if a process—a causal process—is transmitting its own structure, then it will be capable
of transmitting certain modifications in the structure” (1984: 144). But as many critics noted,
the very idea of structure-transference (aka mark-transmission) cannot differentiate causal pro-
cesses from non-causal ones, since any process whatever can be such that some modification
of some feature of it gets transmitted after a single local interaction. A typical example was the
shadow of a car with a dent—this is a “dented” shadow, and the mark is transmitted with the
shadow for as long as the shadow is there. In response to this Salmon strengthened his account
of mark-transmission by requiring that for a process P to be causal, it is necessary that “the
process P would have continued to manifest the characteristic Q if the specific marking inter-
action had not occurred” (1984: 148). It should be clear, though, that this kind of modification
takes us back to persistence! In effect, the idea is that a process is causal if (i) a mark made on
it (a modification of some feature) gets transmitted after the point of interaction and (ii) in
the absence of this interaction, the relevant feature would have persisted, where the required
persistence is counterfactual.
Salmon did modify this view further by adopting Phil Dowe’s (2000) conserved quantity
theory, according to which “it is the possession of a conserved quantity, rather than the abil-
ity to transmit a mark, that makes a process a causal process” (2000: 89). On what has come
to be known as the Salmon-Dowe theory, a causal process is a world line of an object that pos-
sesses a conserved quantity. And a causal interaction is an intersection of world lines that involves
exchange of a conserved quantity.
Dowe fixes the characteristic that renders a process causal and, consequently, the characteris-
tic that renders something a mechanism. A conserved quantity is “any quantity that is governed
by a conservation law” (2000: 91), e.g. mass-energy, linear momentum, charge, and the like.
Apart from various issues that have to do with the question of whether this theory can avoid
counterfactuals (see Psillos 2002: chapter 4), the main practical concern is that this account of
mechanism is too narrow. For even if physical causation—and hence physical mechanism—was a
146
matter of the possession of a conserved quantity, it’s hard to see how this account of mechanism
can even start shedding any light on causal processes in domains outside of physics (biological,
geological, medical, social). These will have to be understood either in a reductive way or in
non-mechanistic (Dowe-Salmon) terms.
A rather liberating conception of causal mechanism was offered by Rom Harré in the early
1970s. Harré connected the traditional idea of power-based causation with the traditional idea
that causation involves a mechanism. What he called “generative mechanism” can be put thus:
generative mechanism = powers + mechanisms
As he (1972: 121) put it: “The generative view sees materials and individual things as having
causal powers which can be evoked in suitable circumstances.” And he added: “The causal
powers of a thing or material are related to what causal mechanisms it contains. These deter-
mine how it will react to stimuli” (1972: 137). For example, an explosion is caused both by the
detonation and the power of the explosive material, which it has in virtue of its chemical nature.
On this view of causation, the ascription of a power to a particular form has the following
form:
X has the power to A = if X is subject to stimuli or conditions of an appropriate kind,

then X will do A, in virtue of its intrinsic nature.
But this is not a simple conditional analysis of powers, since as Harré stressed, power-ascriptions
involve two analysans:
a specific conditional (which says what X will or can do under certain circumstances and
in the presence of a certain stimulus); and
an unspecific categorical claim about the nature of X.
The claim about the nature of X is unspecific, because the exact specification of the nature or
constitution of X in virtue of which it has the power to A is left open. (Discovering this is sup-
posed to be a matter of empirical investigation.)
It is a fair complaint that, as stated above, the ascription of powers is explanatorily incomplete
unless something specific is (or can be) said about the nature of the particular that has the power.
Otherwise, power-ascription merely states what needs to be explained, viz. that causes produce
their effects. This is where mechanisms come in. Specifying the generative mechanism is cash-
ing the promissory note. As Harré put it: “Giving a mechanism . . . is . . . partly to describe
the nature and constitution of the things involved which makes clear to us what mechanisms
have been brought into operation” (1970: 124). So the key idea in this mechanistic view of
causation is this: causes produce their effects because they have the power to do so, where this
power is grounded in the mechanism that connects the cause and the effect and the mechanism
is grounded in the nature of the thing that does the causing.
This, as one of us has noted elsewhere (Psillos 2011), is a broad and liberal conception of
causal mechanism. Generative mechanisms are taken to be the bearers of causal connections.
It is in virtue of them that the causes are supposed to produce the effects. But there is no spe-
cific description of a mechanism (let alone one that is couched in terms of physical quantities).
A generative mechanism is virtually any relatively stable arrangement of entities such that, by
engaging in certain interactions, a function is performed, or an effect is brought about. As
Harré explained, he did not “intend anything specifically mechanical by the word ‘mechanisms’.
147
Clockwork is a mechanism, Faraday’s strained space is a mechanism, electron quantum jumps

[are] . . . a mechanism, and so on” (Harré 1970: 36).
Though this was not quite perceived and acknowledged when Harré was putting forward
this conception, this liberal conception of mechanism pointed to a shift from thinking of mecha-
nisms exclusively as the vehicle of causation (mechanisms-of) to thinking of mechanisms as
whatever implements a certain behavior or performs a certain function (mechanisms-for). On
this broader view, a mechanism is a complex system that consists of some parts (its building
blocks) and a certain organization of these parts, which determines how the parts interact with
each other to produce a certain output. The parts of the mechanism should be stable and robust;
that is, their properties must remain stable, in the absence of interventions. The organization
should also be stable; that is, the complex system as a whole should have stable dispositions,
which produce the behavior of the mechanism. Thanks to the organization of the parts, a
mechanism is more than the sum of its parts: each of the parts contributes to the overall behavior
of the mechanism more than it would have achieved if it had acted on its own.
One natural question may arise at this point. Can a mechanism be both what we called a
mechanism-for and what we called a mechanism-of? That is, can it be the case that a mechanism
both underlies or constitutes a causal process and is a mechanism for a specific behavior? Though
Harré adopted this view, this position acquired new strength in the early 1990s when Stuart
Glennan developed his own mechanistic theory of causation. For him, mechanisms are both
what underlie or constitute causal connections between events and thus provide the missing
link between cause and effect (mechanisms-of) and at the same time complex systems that are
responsible for certain behaviors (mechanisms-for) and are thus individuated in terms of them.
But, mechanisms-for are not necessarily mechanisms-of. Conceptually this is obvious if we
think of a mechanism as a causal process with various characteristics (such as those discussed
above—e.g., they possess a conserved quantity or some kind of persisting structure). There is
no reason to think that this kind of mechanism (e.g., the process by means of which the sum of
kinetic and potential energy is conserved in some interaction) is a mechanism for any particular
behavior. Conversely, if we think of a mechanism as a complex system such that the interactions
of its parts bring about a certain behavior, there is no ipso facto reason to adopt a mechanistic
account of causation. In light of this, we arrive at a tripartite categorization of “mechanistic”
accounts present in the literature (or at three independent notions of what a mechanism is):
mechanisms can be mechanisms-for, or mechanisms-of, or both.4
With this map of the conceptual landscape of the philosophical literature on mechanisms in
mind, our task now is to examine each case in turn and investigate the relations between each
sense of “mechanism” and laws/counterfactuals.
3. Mechanisms-of
Let us first focus on mechanisms-of that are not at the same time mechanisms-for. As noted
already, the best known such causal mechanisms are those discussed by Salmon and Dowe.
Though these accounts of causation are presented as being compatible with singular causation,
it should be quite clear that they rely on counterfactuals. We noted already that in Salmon’s
account counterfactuals loom large. In fact, counterfactuals play a double role in his theory. On
the one hand, they secure that a process is causal by making it the case that the process does not
just possess an actual uniformity of structure, but also a counterfactual one. On the other hand,
they secure the conditions under which an interaction (the marking of a process) is causal: if
the marking would have occurred even in the absence of the supposed interaction between two
processes, then the interaction is not causal.
148
On Dowe’s account, the very idea of a possession of a conserved quantity for a process to
be causal implies that both laws and counterfactuals are in the vicinity. Conserved quantities are
individuated by reference to conservation laws and it is hard to think of a process being causal
without the conserved quantity that makes it causal being governed by a conservation law.
Counterfactuals are also necessary for Dowe’s account of causation. Not just because laws imply
counterfactuals, but also because an appeal to counterfactuals is necessary for claiming that the
process is causal. That is, it seems that without counterfactuals there is no way to ground the
difference between objects to which conserved quantities may be applied and objects to which
they may not (e.g. a single particle with zero momentum vs. a shadow with zero quantity of
charge; the particle, but not the shadow, is a causal process precisely because it could enter into
interactions, which could make its momentum non-zero (see Psillos 2002: 126)).
4. Mechanisms-for and mechanisms-of

Let us now turn to an account such as Glennan’s, i.e. to an account that takes mechanisms
to be both mechanisms-for and mechanisms-of. There are two parts in Glennan’s defini-
tion of mechanisms. First, a mechanism consists of components that interact—in this, it is
similar to Salmon’s account of a mechanism-of as causal process. However, for Glennan,
the mechanism itself is a complex system with a stable arrangement of components (see
his 1996, though in more recent work he drops the stability requirement for some kinds
of mechanisms—see section 5). So, in contrast to the view of mechanisms-of as processes
(which can in principle be singular causal chains of events), such mechanisms are “types of
systems that exhibit regular and repeatable behavior” (Glennan 2010: 259).
How should we understand the interactions among the components of such mechanisms?
Should they be understood in terms of counterfactuals or not? To answer this question, let us
briefly review various possible options.
The first general case we will consider is interactions with laws. Interactions can be governed
by laws, where laws are understood in some robust metaphysical sense. For example, accord-
ing to Dretske (1977), Tooley (1977), and Armstrong (1983), laws are necessitating relations
between universals. So, if there is a necessitating relation between universals A and B, there
will be a law between them and as a result of this law when A is instantiated, so will be B.
Suppose we transfer that to the components of a mechanism: when component X instanti-
ates A at some time t1, some other component Y will instantiate universal B, perhaps at a later
time. Or take the rival view (but metaphysically robust too) that laws are embodied in relations
between powers. If this is the preferred account of laws, interactions will be understood in
terms of powers. Powers are properties possessed by components of a mechanism, and produce
specific manifestations under specific stimuli. Whereas for Dretske, Tooley, and Armstrong the
interactions within the mechanism are grounded in the external relation of nomic necessitation,
in the powers view, interactions are grounded in the internal relations between the powers of
the components of the mechanism. Alternatively, interactions between the components of the
mechanism may be viewed as being governed by metaphysically thin laws; e.g., by (Humean)
regularities. Here, component A can be said to interact with component B, in virtue of the fact
that this interaction is an instance of a regularity.
If, for the time being, we bracket laws, can we understand the interactions among the
components of the mechanism differently? Perhaps counterfactuals can be of direct help here.
So Glennan (2002), following Woodward (2000, 2002, 2003), understands interactions in
terms of change-relating generalizations that are invariant under interventions. Such generali-
zations are change-relating in the sense that they relate changes in component A to changes
149
in component B. They involve counterfactual situations in that they concern what would
have happened to component B regarding the value of quantity Y possessed by it, if the value
of quality X possessed by component A had changed. These generalizations are invariant
under interventions, in that they are about relations between variables that remain invariant
under (actual or counterfactual) interventions. These change-relating generalizations, then,
are grounded in counterfactuals (called interventionist counterfactuals by Woodward—on
Woodward’s theory of causation, see also Chapter 10).
But if we are to understand interactions between components in terms of counterfactuals,
the next question is: what grounds these counterfactuals? In particular, in virtue of what are inter-
ventionist counterfactuals true? The answers here are well known (see Psillos 2004, 2007).
Counterfactuals can be grounded in laws or not. If they are grounded in laws, following what
we said in the previous paragraph, these laws can be either metaphysically robust laws of the sort
adopted either by Armstrong or power-based accounts of lawhood, or thin Humean regularities,
instances of which are particular token-interactions between components. If the counterfactuals
are not grounded in laws, then it’s likely that there are counterfactuals “all the way down”; that
is, that there are primitive modal facts in the world (see Lange 2009).
In any of these accounts of law-governed within-mechanism interactions, counterfactuals
have a central place: either by directly accounting for interactions (as in Woodward’s theory),
or by being part of an account of the nomological dependences that ground the interactions,
or as a primitive modal signature of the world.5 So, if laws regulate the interactions between
the components of the mechanism, we cannot do away with counterfactuals in grounding
within-mechanism interactions. Before, for completeness, we consider the prospects for a
non-law-governed account of interactions, let us discuss an attempt to have mechanisms themselves
ground counterfactuals. This suggestion is put forward by Glennan (1996). For him, although
interactions are understood in terms of interventionist counterfactuals, these counterfactuals are
in turn grounded in (lower-level) mechanisms.
Here is Glennan’s suggestion in more detail. Interactions among components of a mecha-
nism are governed by laws, which are understood in terms of interventionist counterfactuals;
these laws are “mechanically explicable,” i.e. there are other mechanisms that ground them;
but these (lower-level) mechanisms themselves contain parts, the interactions among which are
understood in terms of counterfactuals, and which are in turn grounded in yet other mecha-
nisms, until we finally reach a level where we run out of mechanisms to explain the laws that
govern the interactions among components, and thus to ground the relevant counterfactuals.
At this fundamental level, interactions among components are directly grounded in counterfactuals.
But notwithstanding these not mechanically explicable laws, Glennan insists that at all other
levels mechanisms can ground interactions. So, even if we need to introduce counterfactuals to
account for interactions, mechanisms have priority over counterfactuals, and thus the account
is supposed not to be a version of a difference-making theory of causation, but a genuinely
mechanical account.
However, given the existence of not mechanically explicable laws, it is not clear how mecha-
nisms can ground counterfactuals at any level. That is, given that the mechanisms at the lowest
level depend on counterfactuals, the mechanisms at a level exactly above the fundamental must
be equally dependent (albeit derivatively) on the fundamental counterfactuals, and so on for every
higher level. In other words, to ground counterfactuals at any level, we need the whole lower
hierarchy of mechanisms and counterfactuals, and since we ultimately arrive at a level where
there are either only counterfactuals, or only laws (or both), it seems that there is a fundamental
asymmetry between mechanisms and laws/counterfactuals. The only way to block the asym-
metry would be to argue that the whole hierarchy is not needed to ground the counterfactuals
150
at higher levels. Even if this were to be granted for purposes of explanation—that is, even if
explanation in terms of mechanisms at level n does not require citing lower-level mechanisms—
metaphysically the whole hierarchy constitutes the grounds for the mechanism.6
In sum, if laws are admitted in our notion of mechanism, a reliance on counterfactuals is
inevitable. But can we perhaps avoid counterfactuals if we account for within-mechanisms
interactions in some other way?
We move now to a second approach. We have reviewed various options to understand inter-
actions of components of mechanisms, where these interactions are viewed as law-governed.
The question now is: can we have interactions without some notion of law in the background
(either in terms of regularities, or in some more metaphysically robust sense)? If yes, then this
could be a way to have mechanisms-of as complex systems, without the need to put laws and
counterfactuals in the picture.
For some mechanists, the interactions of components have to be understood in terms of
activities. Activities are a new ontological category that, together with entities, are said to be
needed for an adequate ontological account of mechanisms (Machamer, Darden, and Craver
2000; Machamer 2004). Activities are meant to embody the causally productive relations
between components. Causation in terms of activities is viewed as a type of singular causal-
ity, where the causal relation is a local matter, i.e. it concerns what happens between the two
events that are causally connected, and not what happens at other places and at other times
in the universe (as is the case for the regularity theorist). Activities, thus, have been taken to
obviate the need for laws.
A key argument in favor of activities turns on the claim that causation is singular. But, does
singular causation imply that there are no laws? It would be too quick to infer from singular
causal claims that laws are not part of causation. By singular causation we may simply mean
that there exist genuine singular causal connections, i.e. causal connections between particular
event-tokens. But this is not enough to prove that there are no laws in the background. For it is
consistent with the existence of singular causal sequences that there are laws under which these
causal sequences fall. To use a quick example, on Armstrong’s account of laws, singular causation
is ipso facto nomological causation since the nomic necessitating relation that relates two uni-
versals relates the instances of the two universals too (Armstrong 1997). Interestingly, the same
is true if we take singular causation to be grounded in the powers possessed by objects; powers
are again wholly present in the complex event that constitutes the singular causal sequence. And
though there is no nomic relation that relates the two powers, the regular instantiation of the
two powers implies the presence of a regularity. So, what both these cases show is that even
singular causation can be nomological, i.e. subsumed under laws.7
Thus, singular causation does not, on its own, constitute an argument in favor of viewing
interactions among components of mechanisms as not being law-governed, or more generally,
as not depending on difference-making relations. So, friends of activities need to (i) give more
reasons to justify the introduction of this new ontological category, and (ii) explain why activi-
ties qua producers of change are themselves counterfactual-free. Although it’s conceivable that
singular causation just amounts to the local activation of powers which in turn ground activities,
powers being universals, it’s upon the friends of powers to show that we can understand this
co-instantiation without also assuming that there is a law present.8
5. Mechanisms-for
So far we have argued that mechanisms-of (mechanisms considered as underlying or constituting
causal processes) require laws, and thus difference-making relations must be included in the notion
151
of a mechanism-of. But what about mechanisms-for, mechanisms as complex systems responsible

for certain behaviors? What is the relation between mechanisms-for and laws/counterfactuals?
Recall that a mechanism-for is a mechanism as a complex system such that the interactions of its
parts bring about a certain behavior-function. A mechanism-for need not commit us to a mecha-
nistic (e.g. à la Salmon-Dowe) account of the causal interactions between its parts.
In light of what was said in the previous section, there is an argument as to why mechanisms-
for have to incorporate laws and/or counterfactuals: a mechanism-for involves components
that interact with one another; but laws and/or counterfactuals are needed to account for these
interactions; hence, mechanisms-for need to incorporate laws and/or counterfactuals. However,
Jim Bogen (2005) has taken the existence of mechanisms that function irregularly as an argument
against the view that laws and regular behavior have to characterize the function of mechanisms.
In this section we will deal with this argument from irregular mechanisms.
The first point that we want to stress is that irregular and unrepeatable mechanisms are not
as ubiquitous as some philosophers want us to believe. So, consider a claim made by Leuridan
(2010). He thinks that mechanisms as complex systems ontologically depend on stable regulari-
ties, since there can be no such mechanisms (i) without macrolevel regularities (i.e. the behavior
produced by the mechanism), and (ii) without microlevel ones (i.e. the behaviors of the mecha-
nism’s parts). Kaiser and Craver (2013) have replied to this that Leuridan’s first claim is “clearly
false” since “[o]ne-off mechanisms are mechanisms without a macrolevel regularity,” where
“one-off mechanisms” are the causal processes discussed by Salmon and others (mechanisms-of
in our terminology). Moreover, they point to examples where scientists seem to be interested
in exactly this kind of mechanism, i.e. when they try to explain how a particular event occurred
(for example, a particular speciation event).
This kind of reply confuses the two different senses of mechanism we have tried to disentan-
gle. It is not the case that “singular, unrepeated causal chains . . . are a special, limiting case of
[complex system] mechanisms, not something altogether different,” as Kaiser and Craver insist.
For it is not at all clear that such mechanisms-of are at the same time mechanisms-for, i.e. mech-
anisms for a certain behavior. Similarly, we remain unpersuaded by Glennan’s (2010) claim that
the mechanisms that produced various historical outcomes are mechanisms-for (he calls them
“ephemeral mechanisms”). In any case, it would be very implausible to insist that any arbitrary
causal chain is for a certain behavior (which is identified with the outcome of the causal chain
or, alternatively, with the (higher-level) event constituted by the causal chain). For instance,
the reflection of a light-ray on a surface is a clear case of a mechanism-of (since it constitutes
a causal process), but it is not clear at all that it is a mechanism for a certain behavior (unless of
course we follow Glennan (forthcoming) and equate “behavior” with “phenomenon”; that is,
with causal effect).9
So, it is not enough to point to singular causal chains to argue that there can be irregular
mechanisms, or one-off mechanisms (mechanisms that function only once) (see also DesAutels
2011; Andersen 2012). Still, one may wonder: can there be mechanisms-for without a corre-
sponding macrolevel regularity? Thus, the issue that must be clarified is: what are the conditions
for being a mechanism for a behavior? Is it merely to have a function (which is the mechanism’s
behavior), or should we, in addition, require that the behavior be regular?
This is not the place to discuss at any length the concept of function (on the relation
between functions and mechanisms, see Chapter 8); for the purposes of the current argument,
let us interpret this requirement in a wide sense, i.e. as not requiring that for something to
have a function it has to be the product of conscious design or the result of natural selection. In
other words, we are going to take a function in the sense of Cummins (1975), for whom func-
tions are certain kinds of dispositions (see Craver 2001 for such an approach to the functions
152
of mechanisms). In particular, what it is for a mechanism M to have a function F is to have a

disposition to F, which contributes to a disposition of a larger system that contains M. Such
functions need not be restricted to living systems or artifacts.10 Yet, not anything whatsoever
can be ascribed a Cummins function. In particular, unrepeated causal chains of events, which
might well be Salmon’s and Dowe’s mechanisms-of, need not have a function. We can follow
Cummins and say that talk of functions only makes sense when we can apply what Cummins
calls the analytical strategy, i.e. explain the disposition of a containing system in terms of the
contributions made by the simpler dispositions of its parts.
There can be systems with Cummins functions that exhibit the corresponding behavior only
once; so, there are many biological functions, the realization of which requires that the biologi-
cal entity that has the function cease to exist. An example is the mechanism for apoptosis, i.e.
programmed cell death. Here, the relevant mechanism has a Cummins function, i.e. it causally
contributes to the death of the cell; however, this is a function that, when successfully carried
out, can occur only once. But even in such cases, the behavior of a particular mechanism of
apoptosis is a token of a type of behavior that occurs countless times every second.
Can there be genuinely irregular mechanisms-for; that is, mechanisms-for without a cor-
responding macrolevel regularity? Bogen (2005) has offered the case of the mechanism
of neurotransmitter release as an example of a mechanism-for that behaves irregularly. As
this mechanism more often than not fails to carry out its function, there exists no corre-
sponding macrolevel regularity; but moreover, and more importantly, Bogen thinks that
within-mechanism interactions must themselves be irregular, and thus we must abandon the
regularity account of causation in favor of activities.
We do not think that this last conclusion follows from Bogen’s example. To see why this
is the case, it is useful to distinguish between three cases of what we may call “irregular”
mechanisms-for. The irregularity of mechanisms-for may be only contingent (irregular1),
stochastic (irregular2), or (let us assume) more radical (irregular3).
Irregular1 mechanisms-for are mechanisms that could function regularly, but they in fact
do not. A defective machine that only functions once in a while is a case in point. Such a
machine (i) is a mechanism for a behavior and (ii) functions irregularly. However, it is cer-
tainly not the case that a successful operation of the machine is not subject to laws (e.g. laws
of electromagnetism, gravity, or friction). (Nor is it the case that defective machines falsify the
regularity account of causation.)
Irregular2 mechanisms are like irregular1 mechanisms in that they operate in accordance with
laws, but in this case the laws are probabilistic. So, the existence of such mechanisms does not
show that within-mechanism interactions need not be law-governed (or even that the regularity
account of causation is false—regularities can be stochastic). What if the operation of a mecha-
nism is completely chancy (e.g. because it involves the radioactive decay of a single atom)? Even
if we do not have a law here (perhaps because the relevant law concerns a population of atoms
rather than a single one), it is not at all clear to us that such a chancy “mechanism” could be an
example of a mechanism-for.11
Finally, we can imagine an irregular3 mechanism-for; such a sui generis mechanism only
operates once, and its unrepeatability is supposed not to be a contingent matter, but this is
because the interactions among its components cannot in principle be repeated. We are not sure
that the notion of an irregular3 mechanism-for actually makes sense. But this is the only kind
of example we can imagine, where the irregularity or unrepeatability of a mechanism would
be a reason to think that its operation is not law-governed. If that’s where the friends of genu-
inely irregularly operating mechanisms can pin their hopes for a non-law-governed account of
mechanism, then so be it!
153
6. Conclusions
In this chapter, we have examined the relation between mechanisms and laws/counterfactuals
by revisiting the main notions of mechanism found in the literature. We distinguished between
two different conceptions of “mechanism.” What we have called mechanisms-of tally with the
general conceptions of mechanisms offered in discussions of causation. A “mechanism” in these
views is what underlies or constitutes a causal process or connects the cause with the effect.
What we have called mechanisms-for, on the other hand, are complex systems that function so as
to produce a certain behavior. According to some mechanists, a mechanism fulfils both of these
roles simultaneously.
We have argued that for both mechanisms-of and mechanisms-for, counterfactuals and laws
are central for understanding within-mechanism interactions. We have examined two main
arguments in more detail. Concerning mechanisms-of, we have seen that singular causation is
compatible with several quite different ways of understanding within-mechanism interactions, in
all of which laws and counterfactuals are central. Concerning mechanisms-for, we have argued
that the existence of irregular mechanisms is compatible with the view that mechanisms oper-
ate according to laws. Both of these arguments point to an asymmetrical dependence between
mechanisms and laws/counterfactuals: while some laws and counterfactuals must be taken as
primitive (non-mechanistic) facts of the world, all mechanisms depend on laws/counterfactuals.
Notes
1 We wish to thank Phyllis Illari and Stuart Glennan for valuable comments on an earlier draft of this
chapter.
2 On a nomological account of causal dependence (i.e. B depends on A if there is a law that connects
the two), counterfactuals are required to account for the modal strength of laws (for more on this see
Psillos 2002). So, even if it were to be admitted that the alternative notions of dependence are distinct,
counterfactuals play a key role in all versions of the dependence approach to causation.
3 We take it that to be a mechanism-for is tantamount to being a system performing a function; this
is not always made explicit in general accounts of mechanisms-for (but see Bechtel and Abrahamsen
(2005), Garson (2013), and Chapter 8). The minimal mechanism of Glennan and Illari (see Chapter 1),
which is defined as a mechanism for a phenomenon, can be understood as either mechanism-of or
mechanism-for in our terminology, according to how we understand “phenomenon,” as a function or
as causal process.
4 See Levy (2013) and Andersen (2014a, 2014b) for similar distinctions.
5 In Lange’s (2009) theory of laws it is a counterfactual notion of stability that determines which facts are
lawful and which are accidental. In other theories of lawhood, counterfactuals come “for free,” so to
speak, as they must be part of any metaphysically robust theory of laws (such as that of Dretske, Tooley,
and Armstrong): any such theory must show why laws support counterfactuals. It is not obvious how
exactly counterfactuals are part of a regularity view of laws. But note that this is a problem (if at all) for
the regularity theorist, and not for the view that interactions have to be understood in terms of laws/
counterfactuals. For an attempt to reconcile regularity theory with counterfactuals see Psillos (2014).
6 See Glennan (2011) for an attempt to respond to this argument, and Casini (2015) for a detailed criti-
cism; see also Campaner (2006).
7 There is debate among friends of powers whether such a powers-ontology yields an account of laws in
terms of powers (Bird 2007), or a lawless ontology (Mumford 2004). But this need not concern us here.
8 For more on activities see Chapters 9 and 10; see also Waskan (2011) for a mechanist account of the
contents of causal claims that is not based on counterfactuals and Woodward’s (2011) answer that
causation as difference-making is fundamental in understanding mechanisms; Menzies (2012) provides
an illuminating account of mechanisms in terms of the interventionist approach to causation within
a structural equations framework; lastly, Glennan (forthcoming: chapters 5 and 6) offers a detailed
treatment of mechanistic causation as a productive account of causation not reducible to difference-
making relations.
154
9 In many examples where scientists refer to “mechanisms,” it may not be clear whether it is the notion
of mechanism-of or the notion of mechanism-for (or both, or neither) that they have in mind. For
example, such uses as “mechanism of chemical reaction,” “mechanism of speciation,” “mechanism
of action (of a drug)” can be construed in various ways; to insist on a widening of the concept of
mechanism-for based on scientific practice (without further argument) seems too quick.
10 However, note that for some (e.g. Kitcher 1993) we cannot ascribe even Cummins functions to entities
that are not products of (either conscious or nonconscious) design, i.e. that are neither artefacts nor
living systems.
11 For a notion of “stochastic” mechanism see DesAutels (2011), as well as Andersen (2012).
References
Andersen, H. (2012) “The Case for Regularity in Mechanistic Causal Explanation,” Synthese 189:
415–432.
—— (2014a) “A Field Guide to Mechanisms : Part I,” Philosophy Compass 9: 274–83.
—— (2014b) “A Field Guide to Mechanisms : Part II,” Philosophy Compass 9: 284–93.
Armstrong, D. M. (1983) What Is a Law of Nature?, Cambridge: Cambridge University Press.
—— (1997) “Singular Causation and Laws of Nature,” in The Cosmos of Science, J. Earman and J. Norton
(eds), Pittsburgh, PA: Pittsburgh University Press, pp. 498–511.
Bechtel, W. (2008) Mental Mechanisms: Philosophical Perspectives on Cognitive Neuroscience, New York:
Routledge.
Bechtel, W. and A. Abrahamsen (2005) “Explanation: A Mechanistic Alternative,” Studies in History and
Bird, A. (2007) Nature’s Metaphysics: Laws and Properties, Oxford: Oxford University Press.
Bogen, J. (2005) “Regularities and Causality; Generalizations and Causal Explanations,” Studies in History
and Philosophy of Biological and Biomedical Sciences 36: 397–420.
Campaner, R. (2006) “Mechanisms and Counterfactuals: A Different Glimpse of the (Secret?) Connexion,”
Philosophica 77: 15–44.
Casini, L. (2015) “Can Interventions Rescue Glennan’s Mechanistic Account of Causality?,” British Journal
for the Philosophy of Science, doi:10.1093/bjps/axv014.
Craver, C.F. (2001) “Role Functions, Mechanisms and Hierarchy,” Philosophy of Science 68: 31– 55.
—— (2007) Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience, Oxford: Oxford University
Press.
Cummins, R. (1975) “Functional Analysis,” Journal of Philosophy 72: 741–64.
Davidson, D. (1967) “Causal Relations,” Journal of Philosophy 64: 691–703.
DesAutels, L. (2011) “Against Regular and Irregular Characterizations of Mechanisms,” Philosophy of
Science 78: 914–25.
Dowe, P. (2000) Physical Causation, Cambridge: Cambridge University Press.
Dretske, F. I. (1977) “Laws of Nature,” Philosophy of Science 44: 248–68.
Garson, J. (2013) “The Functional Sense of Mechanism,” Philosophy of Science 80: 317–33.
Glennan, S. (1996) “Mechanisms and The Nature of Causation,” Erkenntnis, 44: 49–71.
—— (2002) “Rethinking Mechanistic Explanation,” Philosophy of Science, 69: S342–S353.
—— (2010) “Ephemeral Mechanisms and Historical Explanation,” Erkenntnis 72: 251–66.
—— (2011) “Singular and General Causal Relations: A Mechanist Perspective,” in P.M.K. Illari, F. Russo
and J. Williamson (eds) Causality in the Sciences, Oxford: Oxford University Press, pp. 789–817.
—— (forthcoming) The New Mechanical Philosophy, Oxford: Oxford University Press.
Hall, N. (2004) “Two Concepts of Causation” in J. Collins, N. Hall & L. Paul (eds) Causation and
Counterfactuals, Cambridge, MA: MIT Press, pp. 225–76.
Harré, R. (1970) The Principles of Scientific Thinking, London: Macmillan.
—— (1972) The Philosophies of Science: An Introductory Survey, Oxford: Oxford University Press.
Kaiser, M. and C.F. Craver (2013) “Mechanisms and Laws: Clarifying the Debate,” in H.K. Chao,
S.T. Chen and R.L. Millstein (eds), Mechanism and Causality in Biology and Economics, Dordrecht:
Springer, pp. 125–45.
Kitcher, P. (1993) “Function and Design,” Midwest Studies in Philosophy 18: 379–97.
Lange, M. (2009) Laws and Lawmakers: Science, Metaphysics, and the Laws of Nature, New York: Oxford
University Press.
155
Leuridan, B. (2010) “Can Mechanisms Really Replace Laws of Nature?,” Philosophy of Science 77: 317–40.
Levy, A. (2013) “Three Kinds of New Mechanism,” Biology & Philosophy 28: 99–114.
Machamer, P. (2004) “Activities and Causation: The Metaphysics and Epistemology of Mechanisms,”
International Studies in the Philosophy of Science 18: 27–39.
Machamer, P., L. Darden and C. F. Craver (2000) “Thinking about Mechanisms,” Philosophy of Science
67: 1–25.
Mackie, J.L. (1974) The Cement of the Universe, Oxford: Clarendon Press.
Menzies, P. (2012) “The Causal Structure of Mechanisms,” Studies in History and Philosophy of Biological and
Biomedical Sciences 43: 796–805.
Mumford, S. (2004) Laws in Nature, London: Routledge.
Psillos, S. (2002) Causation and Explanation, Montreal: Acumen & McGill-Queens University Press.
—— (2004) “A Glimpse of the Secret Connexion: Harmonising Mechanisms with Counterfactuals,”
Perspectives on Science 12: 288–319.
—— (2007) “Causal Explanation and Manipulation,” in J. Person and P. Ylikoski (eds) Rethinking
Explanation, Boston Studies in the Philosophy of Science, vol. 252, Dordrecht, the Netherlands:
Springer, pp. 97–112.
—— (2011) “The Idea of Mechanism,” in P.M.K. Illari, F. Russo and J. Williamson (eds) Causality in the
Sciences, Oxford: Oxford University Press, pp. 771–88.
—— (2014) “Regularities, Natural Patterns and Laws of Nature,” Theoria 79: 9–27.
Reichenbach, H. (1956) The Direction of Time, Berkeley: University of California Press.
Salmon, W. (1984) Scientific Explanation and the Causal Structure of the World, Princeton, NJ: Princeton
University Press.
—— (1997) “Causality and Explanation: A Reply to Two Critiques,” Philosophy of Science 64: 461–77.
Tooley, M. (1977) “The Nature of Laws”, Canadian Journal of Philosophy 7: 667–98.
Waskan, J. (2011) “Mechanistic Explanation at the Limit,” Synthese 183: 389–408.
Woodward, J. (2000) “Explanation and Invariance in the Special Sciences,” British Journal for the Philosophy
of Science 51: 197–254.
—— (2002) “What Is a Mechanism? A Counterfactual Account,” Philosophy of Science 69: S366–S377.
—— (2003) Making Things Happen: A Theory of Causal Explanation, New York: Oxford University Press.
—— (2011) “Mechanisms Revisited,” Synthese 183: 409–27.
156
12
WHAT WOULD HUME SAY?
Regularities, laws, and mechanisms
Holly Andersen
1. Introduction
Contemporary discussion of mechanisms sprang from many failures of laws to do what
philosophers of science wanted them to. Yet the expectations at that time for what laws should
be able to do were impossible to achieve. Is there still room for a better construal of laws in
understanding explanation and the nature of the world as studied by science, or do mechanisms
provide everything that is required for scientific explanations? How do genuine laws relate to
mechanisms in a mechanistic worldview?
After a survey of the discussion around laws and mechanisms, I offer an original argument
for two main ways in which laws and mechanisms can relate without either conceptually or
ontologically reducing to the other, using Hume to illustrate. The first connection between
laws and mechanisms involves the recognition of the “brute” character of laws, such that they
can explain but are not themselves further explicable. This is illustrated by Hume’s skeptical
realism, where the world’s hidden springs and secret principles are forever covered over from
our epistemic access. The second connection between laws and mechanisms involves a unique
role for laws via distinctively mathematical explanations, in which laws can provide a further
degree of necessity, via mathematics, than can any collection of mechanisms, no matter how
comprehensive. This is illustrated in terms of the evidence for such laws, which may include
the sort Hume labeled “relations of ideas,” rather than solely the type of evidence we have for
mechanisms that Hume labeled “matters of fact.”
2. A short recent history of laws and mechanisms

Mechanisms as a characteristic form of explanation were originally developed as an alterna-
tive to a then-dominant way of understanding explanation as necessarily involving laws. The
deductive-nomological model of explanation stated that explanations involve subsuming the
explanandum (that which is to be explained) under a general law. The explanatory work
was done by the law: event A led to event B because A was an F, B was a G, and all Fs are
Gs (e.g. Hempel 1962). The explanans derived its power from the nomologicity of the law
that figured in it. As such, laws needed to be general, or ideally, universal in that they apply
everywhere at all times. They also needed to be more than merely true descriptions of what
157
Holly Andersen
did or will happen. They had to involve a stronger degree of necessity, and convey what
must happen, given those laws. Nomologicity is thus a kind of necessity that is stronger than
mere actuality—it is more than what does happen; and it is weaker than what is sometimes
called metaphysical or logical necessity, which involves what must or could not happen
regardless of the laws.
In this potted mini-history, it is incredibly important to emphasize that the notion of laws
that served in deductive-nomological accounts of explanation were of a peculiar logical sort.
Traditionally, laws have been taken to be part of the world itself, and descriptions of those parts
of the world yielded statements that had a special status as descriptions. The term “law” would
be equally applied to both the features of the world being described, and to the description itself.
For instance, in the law F = ma, force, mass, and acceleration, as well as the necessary relation-
ship between them, are all part of the natural world, and are a special part such that learning that
such a relationship holds among those quantities allows us to make far more predictions than we
otherwise could with individual cases of acceleration, mass, and force. The written mathemati-
cal form describing the relationship among the quantities, “F = ma,” is also called a law. As such,
the term law applied to both the nomological features of the world and equally to the linguistic
devices such as mathematical formulas used to describe those features of the world.
Laws were taken to bear the mark of their necessity in their very syntactic structure, so that
only the schematic structure of the law, rather than its actual content or meaning, was required
to identify a genuine law from a merely true generalization. Much work in the twentieth
century went into trying to find ways to represent laws as sentences in logical notation such
that lawlike and non-lawlike statements were clearly differentiable by syntactic structure alone,
without reference to actual content of the laws (e.g. Ayer 1956). Their idea was that laws in
the world could be identified by their description. Genuine laws, and only laws, could be writ-
ten out logically such that only the form of the sentence matter (all Fs are Gs) was required to
identify such sentences as lawlike, without having to say anything about what the variables (the
F and the G) stood for. This project failed, gruesomely (Goodman 1983). There is no unique
logical structure of laws such that all and only laws, but no accidentally true generalizations such
as “all the coins in my pocket are silver,” have that structure.
This failure to find some logical structure unique to laws was fruitful in generating new ideas
about what makes laws lawlike, and about other forms of explanation that don’t require such
peculiarly structured laws or deductive relationships. Laws in physics, even under best-case sce-
narios, do not resemble the logical creatures required for deductive-nomological accounts (e.g.
Cartwright 1983). Unification of apparently disparate phenomena, for instance, is a historically
well-substantiated form of explanation that does not rely on the logical notion of laws (Kitcher
1981). Explanations in the so-called special sciences were noted to rarely, if ever, use laws of the
form presumed to be found in physics (Fodor 1974). Many causal explanations need not ever
involve such laws (e.g. Woodward 2005; Strevens 2006).
This opened the way for mechanisms to gain attention as a key form that explanations take in
certain sciences (e.g. Bechtel and Richardson 1993). Mechanisms differ from both the logically
distinctive but mythical creatures sought for the D-N model of explanation, and from the best
examples of actual laws found in physics and related disciplines. Mechanisms can be indefinitely local,
particular, and contingent. They can proliferate in number, rather than be eliminated via reduction.
They can serve explanatory roles for which laws, even if they were available, would be somewhat
unsatisfactory and brute (Andersen 2011). Some authors, such as Glennan (1996), even argued that
mechanisms could replace the characteristic kind of nomological necessity associated with laws.
It is important to recognize the heterogeneity of “mechanism” in current discussions for rea-
sons of clarity and philosophical precision. Some have argued that there is a growing consensus
158
What would Hume say?
on the character of mechanisms (see Chapter 1). Others have argued that the situation is more of
a muddle than a consensus, where very distinct senses of the term mechanism are used without
sufficient care as to the differences between them (Andersen 2014a, 2014b). Minimally, there
are relevant differences in emphasis between approaches to mechanisms. Some (e.g. Craver
and Darden 2013) concentrate on the details of scientific practice and the role that mechanisms
play in structuring discovery and investigation. Others (notably Glennan 2010) generalize the
notion to offer a mechanistic worldview strikingly similar to those of the early modern era,
which is more of an alternative to atomism or process ontology than it is an alternative to
specific explanatory practices in contemporary science. Some take mechanisms to involve the
notion of causation (such as Bogen 2005 and Machamer 2004) while others take mechanisms
to be causation (such as Glennan 1996, or for a different notion of mechanism, Salmon 1998).
Each of these conceptions of mechanism involves a different relationship to modal notions like
nomologicity, which in turn changes the details of how such mechanisms are related to laws.
To leave as little metaphysical space, as it were, for laws, I will take the broadest possible
construal of mechanisms (Glennan 2010) to push the question of what if any role for laws
remains in such a mechanistic worldview. In such a construal, explaining the often surprisingly
simple regularities in the world involves describing them with laws and then fleshing out laws
with supporting mechanisms that explain why such lawfully describable regularities hold. There
are two views about the relationship between laws and mechanisms that might follow from this
strongest construal of mechanism. On the first, laws are a useful step in describing regularities,
but cannot be the last step, since laws themselves require explanation in terms of the mechanisms
that support or sustain them. On the second view, mechanisms are a useful way to represent
higher-level structures in the natural world, but they cannot themselves be fundamental, since
if one were to go “all the way down” with mechanisms, there would still be lawful connections
between the elements or stages of a mechanism, such that the fundamental explanatory role
would require laws and their nomologicity, which would remain brute in the sense of providing
explanations but being themselves not further explicable.
If we focus on physics as providing the more fundamental account of the natural world,
laws are often taken to be required for mechanisms. If we focus on biological sciences, laws
and mechanisms appear more like distinctive stages in the process of discovery, investigation,
and explanation of phenomena of interest. It is a difficult and sometimes underappreciated task
to even locate genuine regularities and describe them in a sufficiently precise fashion that they
might qualify as a law in the first place (see also Mitchell 2000). Once a given regularity is identi-
fied and then described in terms of a law, though, that very law is itself a target for explanation.
Why does this law hold? What conditions give rise to or sustain the regularity? Why do the
relevant parameters have these value ranges rather than some other value ranges? Thus, a differ-
ence in initial emphasis between physics versus biology can yield a difference in how prominent
laws versus mechanisms appear in terms of explanation and fundamentality.
3. Contemporary discussions of regularities, mechanisms, and laws

The term “regularity” is often used to highlight a feature of the world, so that it picks out
regularly recurring patterns of behavior. Such regularities provide the grounds for individuating
phenomena for description and investigation. Using a historical example, there is a regularity in
the way in which heavy objects’ velocities change after being dropped, such that Galileo labels
it the Law of Odd Numbers (Cohen 1985). He found a surprising regularity that depended only
on how long the object had been falling, and not, for instance, on the weight of the object,
how large or small it was, and more. The regularity in question was formulated in terms of the
159
Holly Andersen
amount by which the velocity increased between each time interval—it went up by 1 unit, then
3, then 5, and so on. This simple progression in the rate at which the velocity of dropped objects
increased was surprising, generally observable for any object (for the right sorts of objects,
anyway), and not merely a coincidence. The Law of Odd Numbers describes a particular regu-
larity that we now know how to derive from the uniform acceleration of massive objects in a
gravitational field. Galileo’s version of the law is subsumed by Newton’s version as a particular
case of more general laws about mass, force, and acceleration. Newton’s laws thus describe a
different, more abstractly picked out, regularity, of which the regularity that Galileo described
is a proper subset.
This example highlights how some degree of generalizability is required to qualify as a regu-
larity. Regularities must be at least minimally regular, holding over some changes in a variety
of background circumstances and potentially relevant parameters. When scientists move from
identifying an intriguing regularity and calling the resulting description a law, there is a com-
mitment to the belief that the observed regularity does not hold merely for the circumstances
already observed, but will hold for unobserved and future circumstances. Importantly for the
connection with mechanisms, we can be very mathematically precise about how an object
described by the Law of Odd Numbers or Newton’s laws will accelerate by knowing only a few
parameters about the strength of the gravitational field, and without knowing anything about
the way in which a gravitational field brings about acceleration. Thus, laws describe regularities
in a way that commits to a certain scope of generalizability. Laws are regularities that must hold,
rather than ones that happen to hold. It is this necessity that makes them especially useful in
explanations and in understanding the underlying structure of the natural world.
However, generalizability does not necessarily require a law. There are many regularities
in fields such as biology, sometimes expressed in mathematical terms, that are often taken
to be generalizable and describable in mathematically precise ways. There are mathematical
relationships regarding the distribution of traits across generations that hold across species
and are taken to hold into the future as well. They must hold, with the necessity charac-
teristic of laws rather than of accidental generalizations. Yet these regularities don’t seem
to be part of the fundamental structure of the world; if we take laws to reveal the fabric of
the world, then these generalizations fail to be part of it. Beatty (1995), for instance, has
argued that if we ran the tape of evolution again, we would end up with a radically different
outcome. Such incredible contingency seems at odds with the idea of lawfulness. If Beatty
were right, then on the then-current understanding of laws, nothing in biology (or at least,
no products of evolution) could be explainable by recourse to nomological necessity. This
raised a pointed question for philosophers of biology about the status of generalizations in
biology versus physics.
Are there such regularities as would require laws in biology? And would the resulting laws be
of the same type as what goes under the name of law in physics? Cartwright (1983) has argued
persuasively that even physics doesn’t have the kinds of laws for which philosophers of science
went looking in the twentieth century. Rather than try to get biological regularities to look like
laws in physics, Mitchell (1997, 2000) argues for laws that apply to both physics and biology.
This involves rejecting the binary dichotomy between full laws and mere accidental generaliza-
tions. Pragmatic laws are generalizations that are stronger than mere accidental regularities, but
not as strong as universal, exceptionless laws.
The strength of pragmatic laws can be measured along multiple distinguishable dimensions.
The first dimension is stability: across what range of background conditions and parameter val-
ues does the generalization hold? A generalization that holds under a wider range of conditions
160
is more stable, with the limiting ideal of the universal law that holds everywhere at all times.
The stability of the conditions on which a law depends can vary, while nevertheless yielding
generalizations that bear the right degree of necessity under the right conditions. The second
dimension is strength: how strong is the necessity associated with the generalization, for the
conditions under which it holds? At one extreme is the exceptionless law, which is never vio-
lated in its domain. But there may be generalizations that hold almost always that are used in a
lawlike way, even though there are exceptions to them. A third dimension is degree of abstrac-
tion. Generalizations that are more abstract are those that ignore more of the details, to pull out
some broader pattern across a wider range of concrete examples. Woodward (2010) makes a
very similar point, but leaves laws behind in favor of causal generalizations.
This brings us to mechanisms. Mechanisms are often defined explicitly as providing explana-
tions for regularities in nature: why those regularities occur in the conditions that they do, why
they don’t occur under other conditions, how the underlying entities and activities are organ-
ized and causally connected such that they produce as a final stage or give rise to and sustain
the observed regularity for which an explanation is sought (e.g. Bechtel and Richardson 1993;
Machamer, Darden, and Craver 2000). A model of a mechanism for a given regularity may
provide the ability to make predictions about what would happen under new conditions.
This highlights one complementary role that laws and mechanisms can play in science. While
laws describe regularities, mechanisms may go further and explain them. It is a difficult, and
often underappreciated, task to precisely describe a regularity. Finding the right variables, the
right way to relate them, and the right way to delineate the conditions of application, in the
form of a law that is both accurate and can be used to make precise predictions, is itself a major
scientific achievement. To say that laws describe regularities is not to diminish the work that
laws can do. The task of describing via a law is not trivial; it is a huge breakthrough that is often
itself a prerequisite for further work on that phenomenon. Investigations into the mechanisms
responsible for a lawful regularity can yield the mechanisms that sustain, produce, or give rise
to the regularity. Mechanisms explain why that regularity holds as it does, and how it is that
instances of that regularity occur when and how they do.
It is vital to note that there need not be one mechanism = one law equivalence. Multiple
distinct kinds of mechanisms may give rise to one lawful regularity. It might be that a par-
ticular law, as the precise description of a recurrent regularity that involves some degree of
necessity, holds because of several distinct mechanisms operating under different circum-
stances. Conversely, there may be many different laws that turn out to involve the same
underlying mechanism in different contexts. Maxwell’s laws illustrate both directions of this.
According to Morrison (2007), the formalism Maxwell used allowed him to provide quite
general laws without specifying the variety of physical causes that might be involved in any
given instance to which the laws apply, allowing for multiple mechanisms underlying one
law. At the same time, Maxwell’s laws unified apparently disparate phenomena of electricity
and magnetism, such that several older laws could be subsumed in terms of unified mecha-
nisms of electromagnetism (e.g. p. 64).
Regularity can mean recurrent patterns of behavior that can be identified as phenom-
ena of interest. There are also further notions of regularity that appear in discussions of
mechanisms, focused on how tightly bound the stages of a mechanism are to count as a
mechanism. Bogen (2005) has argued that regularity should not be a requirement on count-
ing as a mechanism. Machamer, Darden, and Craver define mechanisms specifically in terms
of regularity. “Mechanisms are regular in that they work always or for the most part in
the same way under the same conditions” (2000, p. 3). Bogen (2005) argues that there are
161
Holly Andersen
mechanisms (or what ought to count as mechanisms) that do not meet the requirement of
“almost or for the most part.”
Bogen invokes the example of the release of neurotransmitter vesicles given the stimulation
of the neuron. Only about 10 percent of vesicles, in the right start-up conditions, actually release
neurotransmitter. According to the MDC definition, the case should not count as a mechanism,
since the “always or for the most part” condition is not met. Yet this release is still a key part in
the mechanism for transmitting a signal across the synapse. How tightly bound together must a
series of causal processes be to count as a mechanism, rather than a mere coincidental collection
of nearby interactions? Bogen’s aim is to push us toward a lower degree of regularity, such that
the 10 percent counts as regular enough under the circumstances.
It is helpful to distinguish two possible loci of regularity at issue here: first, there is the degree
to which a regularity exists in the world, as a phenomenon that could stand as the target of
explanation; second, there is the degree of regularity within the mechanisms that explains such
a phenomenon, where different organizational stages might have different probabilistic degrees
of connection (see Chapter 13). I have responded (2012) to Bogen’s challenge by arguing that
mechanisms must be considered as at least minimally regular, in the sense of not being one-off
causal chains, to be the target of scientific (rather than historical, for instance) explanation. If
mechanisms are to do explanatory work for individual occurrences of a mechanism, it must be
by dint of situating that individual instance as a member of a type of occurrence. A taxonomy of
different degrees of regularity, and locations of different organizational stages of the mechanism,
can convey a great deal of information about mechanism(s), and provide additional phenomena
requiring further explanation. For example, the 10 percent figure for vesicle neurotransmit-
ter release is part of a larger, embedding mechanism that essentially uses the 10 percent rate to
calibrate neurotransmitter levels to reduce noisy synapse firings. Thus, the purportedly irregular
10 percent release rate is a consistent regularity in the other sense, that of a recurrent pattern of
behavior requiring explanation for which a mechanism can be sought.
Notice how important the issue of generalization has been for construing the relationship
between mechanisms and regularities. The very idea of a regularity contains within it the notion
of recurrence: a singular event, that will only happen once, cannot be a regularity, and thus can-
not be explained as a regularity. Insofar as it is situated within a group of other possible instances,
even if they are only merely possible instances, it is already being treated as a kind of regularity.
Thus far, the relationship between laws and mechanisms has been more complementary than
competitive. There is one contemporary debate that does pitch mechanisms and laws as com-
petitors in the explanation business. A mechanism can explain why a law holds, and a law can
connect the stages within a mechanism. Which is more explanatorily fundamental, mechanisms
or laws? Suppose some regularity is identified and described as lawful. Further investigation
turns up a mechanism that explains how that lawful regularity holds. But this mechanism is
composed of organized entities and processes that are themselves lower-level lawful regularities.
The mechanism that explained the original law requires further laws for its own operation. And
each such law might be further explained by a mechanism, and the stages of that mechanism
must involve laws to connect them, and so on downwards.
Where does this end? There are two main options. It could terminate with laws of physics
at the most fundamental levels, such that there are no further mechanisms that could be posited
to underlie the laws. They would be brute, in that they could explain but not be themselves
explained; they would simply hold of the world. On the other option, even the laws of physics
could themselves be mechanistically explained, such that mechanisms would be ontologically
and explanatorily fundamental.
162
This framing of the question has several notable features, regardless of the eventual answer.
The first is that it emphasizes the ontological or even metaphysical aspects of the question. It
is not merely a question of what explains what. The explanatory consequences follow from
ontological priority. A second feature is that the framing presupposes that either laws, or mecha-
nisms, but not both, are ontologically primitive. This puts laws and mechanisms in a kind of
explanatory competition, where it must be the same explananda that both endeavor to explain,
but where there is only room for one genuine explanation, and only these two options on
the table. A third feature is that it lacks criteria by which this question would be adjudicated.
If string theory, for instance, turns out to provide the material for a grand unified theory, are
the core features of the world that it postulates mechanisms or laws? How much change to the
notion of mechanism would be required to accommodate such a heavily mathematical theory?
At some point in stretching the meaning of mechanism to fit mathematical models, it simply
can’t be the same kind of mechanisms as are posited to explain the firing of a synapse (although
a different conclusion is reached by Kuhlmann and Glennan 2014).
Finally, raising the question in terms of laws versus mechanisms elides the issue of causation
versus mechanisms, or even in terms of nonmechanistic causation versus mechanistic causation.
Mechanisms and counterfactuals are also taken to be opposing potentially ultimate categories
(see especially Psillos 2004; Bogen 2005; Machamer 2004; Woodward 2005; Glennan 2010;
Chapters 10 and 11 in this volume). The opposition is strikingly similar: counterfactuals govern
the relationships between elements or stages in a mechanism, but those counterfactuals can be
cashed out with yet lower-level mechanisms, and so on.
In these discussions of regularities, there is a kind of modal bump in the rug. This bump is the
“oomph” that has been associated with causation, or the nomologicity that has been associated
with laws, or the intricate architecture of a mechanism. The bump in the rug can be shifted,
depending on other philosophical considerations, to be located at laws, causation, or mecha-
nisms, but it has proven remarkably hard to just stomp it flat. When some regularity holds and
we have reason to think that it would hold if poked and prodded in various ways, we need a
way to capture the extra content that goes above and beyond merely describing what actually
happened. This is not a merely semantic point: if some particular event really did have to happen
a certain way, rather than merely happening to happen, leaving that out would be an incomplete
description. Insofar as science is in the business of working toward not merely accurate but also
complete descriptions of various parts of the natural world, it must be able to note these modal
characteristics in a way that accurately portrays the degree of connection.
Where Mitchell offered pragmatic laws to do this work in fields such as biology, others such
as Woodward (2010) offer very similar considerations, including scope, specificity, and stability,
for causal generalizations in biology. Cartwright (2002) notes how discussions of explanation
and scientific theories tried to eschew talk of causation by using talk of laws, and how that pen-
dulum has now swung back to causation. Lewis’ account of causation is based on regularities;
Salmon’s is based on mechanisms. Glennan (1996) also turns to mechanisms for causation.
There is an incredibly close link between laws and causation as the two leading candidates to
account for the degree of connection or necessity beyond mere accident that we find in many of
the most interesting generalizations in the sciences. We can attribute this necessity to causation,
and cash it out one way, or to laws, and cash it out a different way. In this regard, either laws or
causation, but not both, are required. This is not a tension per se, in that it needn’t mean that
only one of laws or causation are required (for instance, the idea of causal capacities or powers
involve both; see Chapter 10). But it tends to go along with relying on one or the other to do
the work of accounting for necessity and connection.
163
Holly Andersen
4. Two Humean roles for laws

We’ve now explored several subtle connections between laws and mechanisms as ways of
explaining regularities in the world that bear some degree of necessity. In this section, I will lay
out a schematic argument for two roles unique to laws that cannot be assimilated to mecha-
nisms. Taking the broadest possible construal of mechanisms, and a weak construal of laws,
consider: can mechanisms, if construed maximally broadly, do all the work that we wanted from
laws, or is there still a role left for laws no matter how broadly one construes mechanism? This
question is both perennial, in that it has arisen in a number of forms in philosophy since the early
modern period, and Humean, in that it often arises anew in discussions of well-loved passages
from Hume. I rely on Hume as a springboard for making the case for two new ways in which
laws and mechanisms could relate, since both appear in his work. This is not Hume exegesis: I
take it that no unambiguous answer can be given to the question of what Hume himself would
actually say. Instead, this is a riff on Humean-style answers as a way of illustrating the point.
As Beebee (2006) has persuasively argued, there are at least two viable ways of understand-
ing what Hume says about causation, and no further definitive answers about which is what
he “really” meant (if we are even willing to assume Hume had a single consistent view across
all his writings). One interpretation developed by Beebee is that Hume was a skeptical realist.
The secret connection, which might bind cause and effect or primary and secondary quality,
the connection that is tracked by the idea of force and purportedly conveys necessity along its
chain—to a skeptical realist, such a secret connection exists, but the world is such that we are in
principle barred from ever gaining genuine epistemic access to it. It exists, but we can’t know it.
It is confessed, that the utmost effort of human reason is to reduce the principles,
productive of natural phenomena, to a greater simplicity, and to resolve the many par-
ticular effects into a few general causes, by means of general reasonings from analogy,
experience, and observation. But as to the causes of these general causes, we should
in vain attempt their discovery; nor shall we ever be able to satisfy ourselves by any
particular explication of them. These ultimate springs and principles are totally shut up
from human curiousity and enquiry.
(Hume [1748] 2007, p. 17)
To the skeptical realist, laws may serve as the “the few general causes,” the most general expla-
nation that science may reach. These laws describe the phenomena that are produced by the
“ultimate springs and principles.” Yet the springs and ultimate principles themselves, the mecha-
nisms producing those laws, are “totally shut up” from our investigations, off limits to knowledge.
In other passages, Hume denies that any amount of knowledge of what we would now call
mechanisms is sufficient to discern what might happen with further instances prior to actual
observation.
Our senses inform us of the colour, weight, and consistence of bread; but neither sense
nor reason can ever inform us of those qualities which fit it for the nourishment and
support of a human body . . . . The bread, which I formerly eat, nourished me; that
is, a body of such sensible qualities was, at that time, endued with such secret powers:
but does it follow, that other bread must also nourish me at another time, and that like
sensible qualities must always be attended with like secret powers? The consequence
seems nowise necessary.
(Hume [1748] 2007, p. 33)
164
Students encountering this passage for the first time often wonder what Hume would say in
the face of modern science. Once we know the mechanisms by which chemicals interact with
the microscopic processes in our digestive tracts, wouldn’t we know whether a new piece of
bread, or even an entirely new foodstuff given straight to scientists before eating, would harm
us or nourish us? There is an intuitive appeal to the idea that this is a problem on which actual
progress has been made by science—the secret powers of food are not so secret anymore.
But Hume himself considers this question. Experience has led us to understand some
mechanisms about nourishment, certainly; but even those rest on further sensible qualities that
ultimately must be supported by the hidden powers and secret connections that Hume is chal-
lenging. Food science merely defers the problem. It does not and cannot solve it.
One needn’t even commit to the realist part of skeptical realism, leaving laws as sheerly
brute. If there are no such hidden springs, the result is still skepticism that precludes any
possible mechanistic explanation of laws. Thus, the first Humean role for laws is an espe-
cially poignant one. It allows for the possibility that there are mechanisms underneath the
fabric of the universe, and it is these mechanisms that give the modal oomph to those laws
by providing the secret connection. Yet they remain out of reach, if they even exist. The
laws that are the last description of regularity before those hidden springs will remain brute,
in that they can explain, but cannot be themselves explained. Even in the face of massive
amounts of contemporary human knowledge of mechanisms, these ultimate regularities
cannot be assimilated to them.
The second Humean role for laws that cannot be assimilated to mechanisms stems from the
evidence for relations of ideas versus for matters of fact. The first role for laws, just discussed, is
squarely within the realm of matters of fact: facts about regularities that are described lawfully,
resolved into mechanisms, but ultimately come up against the opacity of the hidden springs. The
second role for laws involves their status as mathematical relationships that fall at least partially
under the category of relations of ideas and thus have additional evidentiary support compared
to matters of fact.
Hume very famously distinguishes between knowledge in terms of the target of inquiry.
All the objects of human reason or enquiry may be naturally divided into two kinds, to
wit, Relations of Ideas, and Matters of Fact. Of the first kind are the sciences of Geometry,
Algebra, and Arithmetic; and in short, every affirmation which is either intuitively or
demonstratively certain. . . . Matters of fact, which are the second objects of human
reason, are not ascertained in the same way; nor is our evidence of their truth, how-
ever great, of a like nature with the foregoing. The contrary of every matter of fact is
still possible; because it can never imply a contradiction.
(Hume [1748] 2007, p. 25)
He considers the kind of evidence we could have for the truth or falsity of claims about mat-
ters of fact versus relations of ideas. Since the contrary of any factual claim is not contradictory,
we cannot use contradiction as a guide to truth and falsity for these claims, but must rely on
evidence of the senses. “All reasonings concerning matter of fact seem to be founded on the
relation of Cause and Effect. By means of that relation alone we can go beyond the evidence of
our memory and senses” (ibid., p. 70). And of course, reasoning based on cause and effect is
ultimately founded on sheer habit. It is in this gap between what we infer about matters of fact,
and that from which we infer it, that Humean skepticism arises.
In contrast, relations of ideas have an entirely different evidentiary status. They do not
ultimately rest on mere habit. They do not admit of the skepticism about knowledge to which
165
Holly Andersen
matters of fact are subject. Claims about relations of ideas can be known with certainty and
assurance, because their contraries are contradictions and therefore impossible. The claim I
am offering here is that there is at least the possibility for mathematical laws to have some
evidential support that is of the form of relations of ideas, and thus not reducible to mechanisms
(do note this is an original argument about how to apply this distinction to laws, not an existing
view in Humean scholarship).
Consider laws that are formulated mathematically (setting aside non-mathematical laws for
now). Laws that are supported by evidence that is at least partially mathematical in character
will have a different status than those based purely on matters of fact, even if the other part
of the evidential support is drawn from matters of fact. Laws can be used in ways that rely
on their mathematical features to draw conclusions that involve a markedly higher degree of
necessity than mere nomologicity can convey, even though those same laws, used some other
way, behave in a traditional way, conveying nomological but not mathematical necessity (for
instance, see Lange 2013 and Andersen 2016).
Consider what “relations of ideas” evidence might be available for a candidate mathemati-
cal law. Insofar as a law is derived from other mathematical laws, plus additional mathematical
machinery, some of the evidence for the new law taking the form that it does is drawn from
those mathematical relationships. Not all of the evidence for such a law is. Any law with genu-
ine physical content will require at least some evidence of matters of fact. The point I want to
make is that some of those laws may also have additional evidentiary support from relations of
ideas, and that such additional support is not even potentially available for mechanisms.
An intriguing potential example of this is leading versions of string theory, where this
second Humean role for laws can account for why string theory is even being pursued as a viable
physical theory despite the well-known lack of empirical confirmation. There is widespread
agreement that string theory is not supported by empirical evidence, since it is extraordinarily
difficult to even find ways to derive empirical predictions from it. In other words, string theory
lacks evidence of matters of fact. Why is it even being considered as a potential theory, much
less a fundamental one? The mathematical structures themselves, and the ways in which some
mathematical relationships emerge as lawfully governing any such structure in the world, pro-
vide the kind of evidence that physicists find sufficiently compelling to continue working on it.
This evidence is in the category of relations of ideas.
There is thus a philosophically pessimistic and a philosophically optimistic role for laws to
play that cannot be assimilated to mechanisms, no matter how broadly construed. Each of these
two roles have a distinctively Humean flavor. Pessimistically, laws might be brute and not fur-
ther explicable; we can never know if mechanisms behind those laws even exist, much less what
they are. Optimistically, laws can play a unique role by dint of mathematical relationships: this
yields mathematical necessity as part of the nomologicity of laws, and an additional potential
source of evidence, relations of ideas, that is not susceptible to inductive skepticism.
5. Conclusion
It is helpful in many contemporary discussions involving laws, mechanisms, regularities,
and even causation to consider the recent trajectory of these ideas since the mid-twentieth
century. The idea of a law was supposed to do an enormous amount of work in the devel-
opment and use of scientific theories. Peculiarities in views about the nature of language
painted philosophers into a corner. Laws were expected to shoulder the burden of explana-
tion and to clarify nomologicity in logical terms. This impossible task failed, in interesting
ways, not least of which was that it cleared the ground for mechanisms to surge as a locus
166
of research for philosophers of science trying to capture the investigatory and explanatory
practices of sciences like biology.
Yet alternative construals of laws, including but not limited to Mitchell’s pragmatic laws,
continue to offer something unique in terms of capturing the right degree of necessity associated
with many scientific claims. I have argued here for two new ways to think about the relation-
ship between mechanisms and laws. One is pessimistically Humean, where there may or may
not be ultimate mechanisms under the very fabric of the universe which give rise to the laws,
but which, if they even exist, are in principle locked away from us. The second is optimistically
Humean, where mathematically formulated laws can offer a way to evade inductive skepticism
at least partially, by relying on evidence drawn from mathematics, or from the relations of ideas
rather than matters of fact.
References
Andersen, H., 2011. Mechanisms, laws, and regularities. Philosophy of Science, 78(2), pp. 325–331.
Andersen, H., 2012. The case for regularity in mechanistic causal explanation. Synthese, 189(2), pp. 415–432.
Andersen, H., 2014a. A field guide to mechanisms: Part I. Philosophy Compass, 9(4), pp. 274–283.
Andersen, H., 2014b. A field guide to mechanisms: Part II. Philosophy Compass, 9(4), pp. 284–293.
Andersen, H., 2016. Complements, not competitors: Causal and mathematical explanations. The British
Journal for the Philosophy of Science, doi:10.1093/bjps/axw023.
Ayer, A.J., 1956. What is a law of nature?. Revue Internationale de Philosophie, 7, pp. 144–165.
Beatty, J., 1995. The evolutionary contingency thesis. In G. Wolters and J.G. Lennox (eds.), Concepts,
Theories, and Rationality in the Biological Sciences. Pittsburgh: University of Pittsburgh Press. pp. 45–81.
Bechtel, W. and Richardson, R.C., 1993. Discovering Complexity. Princeton, NJ: Princeton University
Press.
Beebee, H., 2006. Hume on Causation. London: Routledge.
Bogen, J., 2005. Regularities and causality: Generalizations and causal explanations. Studies in History and
Philosophy of Science Part C: Studies in History and Philosophy of Biological and Biomedical Sciences, 36(2),
pp. 397–420.
Cartwright, N., 1983. How the Laws of Physics Lie. Oxford: Oxford University Press.
Cartwright, N., 2002. In favor of laws that are not ceteris paribus after all. In Ceterus Paribus Laws.
Dordrecht, the Netherlands: Springer. pp. 149–163.
Cohen, I.B., 1985. The Birth of a New Physics. WW Norton & Company.
Craver, C.F. and Darden, L., 2013. In Search of Mechanisms: Discoveries across the Life Sciences. Chicago:
Fodor, J.A., 1974. Special sciences (or: the disunity of science as a working hypothesis). Synthese, 28(2),
pp. 97–115.
Glennan, S., 1996. Mechanisms and the nature of causation. Erkenntnis, 44(1), pp. 49–71.
Glennan, S., 2010. Mechanisms, causes, and the layered model of the world. Philosophy and Phenomenological
Research, 81(2), pp. 362–381.
Goodman, N., 1983. Fact, Fiction, and Forecast. Cambridge, MA: Harvard University Press.
Hempel, C.G., 1962. Deductive-nomological vs. statistical explanation. Minnesota Studies in the Philosophy
of Science, 3, pp. 98–169.
Hume, D. [1748] 2007. An Enquiry Concerning Human Understanding. Fq Classics.
Kitcher, P., 1981. Explanatory unification. Philosophy of Science, 48(4), pp. 507–531.
Kuhlmann, M. and Glennan, S., 2014. On the relation between quantum mechanical and neo-mechanistic
ontologies and explanatory strategies. European Journal for Philosophy of Science, 4(3), pp. 337–359.
Lange, M., 2013. What makes a scientific explanation distinctively mathematical?. The British Journal for the
Philosophy of Science, 64(3), pp. 485–511.
Machamer, P., 2004. Activities and causation: The metaphysics and epistemology of mechanisms.
International Studies in the Philosophy of Science, 18(1), pp. 27–39.
Machamer, P., Darden, L. and Craver, C.F., 2000. Thinking about mechanisms. Philosophy of Science, 67,
pp. 1–25.
Mitchell, S.D., 1997. Pragmatic laws. Philosophy of Science, 64, pp. S468–S479.
167
Holly Andersen
Mitchell, S.D., 2000. Dimensions of scientific law. Philosophy of Science, 67(2), pp. 242–265.
Morrison, M. 2007. Unifying Scientific Theories: Physical Concepts and Mathematical Structures. Cambridge:
Psillos, S., 2004. A glimpse of the secret connexion: Harmonizing mechanisms with counterfactuals.
Perspectives on Science, 12(3), pp. 288–319.
Salmon, W.C., 1998. Causality and Explanation. Oxford: Oxford University Press.
Strevens, M., 2006. Depth: An Account of Scientific Explanation. Cambridge, MA: Harvard University Press.
Woodward, J., 2005. Making Things Happen: A Theory of Causal Explanation. Oxford: Oxford University
Press.
Woodward, J., 2010. Causation in biology: Stability, specificity, and the choice of levels of explanation.
Biology & Philosophy, 25(3), pp. 287–318.
168
13
PROBABILITY AND CHANCE
IN MECHANISMS1
Marshall Abrams
Introduction
Though many authors recognize that mechanisms can involve stochasticity, there’s been little
discussion about the nature of the stochastic relationships in mechanisms. I try to elucidate some
issues that arise with stochastic mechanisms and propose new ways of thinking about them. I’ll
focus mainly but not exclusively on what I call recurrent mechanisms—token mechanisms that
operate in a similar manner at different times, or mechanism types that have many instances that
operate in a similar manner. Scientists’ uses of “mechanism” seem to refer primarily to these
mechanisms, which have also been the main focus of recent discussions of mechanistic explana-
tion by philosophers.
Probability-related terminology varies greatly. I’ll use “stochastic” for anything that seems at
least vaguely probabilistic, or chaotic, or erratic, or merely not guaranteeing specific outcomes,
etc. “Probability,” “probabilistic,” “chance,” and “random” will refer to things that involve a
particular probability distribution, whether known or not. Thus, random behavior is always
stochastic, as is erratic behavior (defined below); the converse doesn’t hold.
I begin in section 1 with a description of part of the mechanism of bacterial chemotaxis,
which provides an illustration I use throughout the chapter, and then make some general
flagellum motors
aspartate
methylation
sites CheY CheY
phosphate
Tar receptor
phosphate
methyl groups
Figure 13.1 Schematic diagram showing some elements (entities and activities) involved in E. coli
chemotaxis. Flagella are in their “pushy” state. See text
169
Marshall Abrams
remarks about roles for probability in mechanisms. The rest of the chapter is somewhat
arbitrarily divided into two sections (sections 2 and 3), focusing on metaphysical and episte-
mological issues. Under metaphysics I include a brief general discussion of interpretations of
probability, and a subsection in which I argue that those interpretations of probability relevant
to the functioning of mechanisms are usually what I call causal probability interpretations. I also
suggest that some stochasticity in mechanisms might not involve probabilities per se, but could
instead involve what are known as imprecise probabilities. The epistemology section discusses
various strategies for modeling probabilistic mechanisms, as well as evidence for probabilities in
mechanisms. I conclude in section 4 with a discussion of non-recurrent mechanisms.
1. Chemotaxis in E. coli
The following simplified sketch of Escherichia coli’s chemotaxis mechanism2 will help to clarify
ideas below and suggest their connection to scientific practice (see Figure 13.1). Unfamiliar
terms in this section can be treated as names for roles I’ll describe.
Each E. coli bacterium has several flagella that allow it to swim toward beneficial substances
and away from detrimental substances by sensing chemical gradients in water. In what I’ll
call pushy motion, flagella push in toward the body and soon wrap together into a bundle
at the rear of the bacterium, pushing it forward. In tumbly motion, flagella turn in the other
direction, pulling away from the cell membrane and causing the bacterium to tumble stochas-
tically. These two states alternate roughly once per second. By swimming for longer periods
when pointed in the direction of increasing aspartate, for example, and swimming for shorter
periods otherwise, the bacterium makes overall progress toward greater concentrations of
aspartate. This is a good strategy because bacteria are so small that Brownian motion repeat-
edly knocks them off course. The duration of pushy motion is controlled by several receptors
embedded in an E. coli’s cell membrane.
For example, the Tar receptor responds to gradients in aspartate concentrations. Tar recep-
tors in what’s called the “active” state facilitate the phosphorylation of CheY molecules.
Phosphorylated CheY molecules can bind to structures on a flagellar motor, increasing the
chance that the flagellum will switch from its default pushy rotation to tumbly rotation. An
active Tar receptor thus reduces the probability that directional movement will persist. The
probability that a Tar receptor is in the active state is in turn a function of (a) the number of
aspartate molecules bound to external sites on the receptor, and (b) the number of methyl
groups bound to methylation sites on the internal side of the receptor. When there is a balance
between the effects of bound aspartate molecules and effects of filled methylation sites, the
receptor is more likely to shift to its active, CheY-phosphorylating state. An excess of bound
aspartate molecules makes the receptor more likely to shift to its inactive state. Because bound
methyl groups accumulate somewhat slowly, a balance between the two kinds of bindings
means that there hasn’t been a recent increase in the number of bound aspartate molecules
on the outer side of the receptor. (I won’t discuss the process by which methyl groups are
removed from the receptor.) In sum, when there is an increase in the number of aspartate bind-
ings, this initially prevents the active state, allowing pushy rotation to persist. However, unless
there is a continuing increase in the number of aspartate bindings, methylation will gradually
counteract effects of the unchanging aspartate bindings, tending to cause the receptor to switch
to its active state, and increasing the chance of the flagella switching to tumbly rotation. All
of these processes seem to be stochastic: given a certain number of aspartate bindings, bound
methyl groups, etc., whether and when these processes occur varies from instance to instance,
even within a single bacterium.3
170
Probability and chance in mechanisms
Stochastic mechanisms
Machamer, Darden, and Craver (2000) (MDC) defined “mechanism” in this way:
Mechanisms are entities and activities organized such that they are productive of regular
changes from start or setup to finish or termination conditions.
(Machamer et al. 2000, p. 3)
Roughly, entities are things—Tar receptors, CheY molecules, etc.—and activities are
what they do—phosphorylation, movement of CheY molecules, etc. The activities associ-
ated with an entity can involve changes in that entity, as in the case of flagellar rotation,
or in other entities, as in the effect of a methyl group on the state of the Tar receptor to
which it is bound.
MDC also wrote that:
Mechanisms are regular in that they work always or for the most part in the same way
under the same conditions.
(Machamer et al. 2000, p. 3, emphasis added)
MDC’s discussion implies that activities whose setup conditions occur have (at least) a very high
probability of occurring. In the case of the E. coli chemotaxis mechanism, we might think of the
setup condition for the mechanism as an increase in the number of aspartate molecules binding
to it, resulting in the Tar receptor remaining in its “inactive” state for longer.
It’s now pretty clear that MDC-like views must be modified to allow for activities with
probabilities far from 1.4 The question of how probable activities must be was partially moti-
vated by questions about whether natural selection counted as a mechanism (see Chapter 22),
but some of the examples that have been central to philosophical discussion of mechanisms in
recent years turned out to be probabilistic (or at least stochastic). For example, Craver (2007,
p. 68), discussing a mechanism that leads to a certain kind of increase in neuronal efficiency,
notes that it seems to be successful no more than about 50 percent of the time. My discus-
sion of E. coli below provides another illustration. (I’ll emphasize causation as relevance, i.e.
difference-making, rather than production simply because probability is not only about what
actually happens, but about what might happen, and because, as we’ll see, scientific investigation
of probabilistic mechanisms routinely involves relevance.)
I assume that scientists have some flexibility in what they treat as entities and activities of
particular token mechanisms: Scientists model mechanisms, and different models can charac-
terize different components of the same token mechanisms, or do so with different degrees of
approximation (see Chapter 17). Claims about probabilities in a mechanism are then relative to a
model that picks out certain ways of decomposing it. Since my focus is primarily on probabilities
in the mechanism itself—i.e. in the world—I’ll treat models as specifying aspects of the world
about whose probabilities we then inquire.
Activity probabilities
Given a model of a mechanism that specifies its entities and activities, we can ask about
the probability of an activity producing certain changes in entities; I’ll call such changes
outcomes, and I’ll call this probability an activity outcome probability, or more simply an activ-
ity probability. A prima facie different question concerns the probability of one activity
171
Marshall Abrams
φ1 φ2
S X T
Figure 13.2 Three entities S, X, and T, and two possible activities ϕ1 and ϕ2 that might produce changes
in X and T when caused to occur as a result of changes in S and X, respectively
occurring rather than another. This probability of activity occurrence may be equivalent
to a probability of a change in an entity, or the probability of multiple entities interacting
in a particular way. For the sake of a simpler, more unified discussion below, if there are
probabilities for alternative activities ϕi to occur, I’ll treat the process that determines which
ϕi it is that occurs as itself an activity. Then the occurrence of one of the different ϕi’s is
this activity-determining activity’s outcome, and we can talk about its own activity prob-
ability. (Where there is only a single activity that may or may not occur, we can treat its
non-occurrence as a kind of null activity.)
I’ll restrict my focus to activity probabilities in the preceding sense, but some authors
discuss the probability of a mechanism as a whole exhibiting a particular behavior or phenom-
enon (e.g. DesAutels 2015; Andersen 2012; Krickel forthcoming). This is the probability of
exhibiting particular termination conditions given particular setup conditions. However, the
probability that a mechanism produces termination conditions given setup conditions reduces
to activity probabilities and facts about the mechanism’s structure. Consider, for example, a
simple mechanism with (a) startup conditions represented by variable S taking the value s,
(b) an entity X in one of two states x1 or x2, and (c) two possible termination conditions t1 and
t2, represented as values of a variable T. Suppose there are two activities ϕ1 and ϕ2 that may
or may not occur (see Figure 13.2). If we simplify part of the E. coli chemotaxis mechanism
described above, the result would be a mechanism of this form: Let S = s represent the presence
of an additional methyl group at one of a Tar receptor’s methylation sites, X = x1 represent
the Tar receptor remaining in its “active” conformation, and T = t1 represent a flagellar motor
with a bound CheY molecule. ϕ1 could then represent the processes by which the Tar recep-
tor changes from its active shape to inactive shape, and ϕ2 could represent the complex activity
consisting of a CheY molecule becoming phosphorylated, moving from the Tar receptor to a
flagellar motor, and binding to the motor.
If we assume that both activities will occur, the probabilities in this model are as follows:
1) For ϕ1: S being in state s results in X being in either state x1 or state x2 with probabilities
P(X = x1|S = s) and 1 - P(X = x1|S = s), respectively.
2) For ϕ2: X being in one of its states results in one of two termination states t1 or t2. The
probability of t1 is either P(T = t1|X = x1) or P(T = t1|X = x2), depending on whether
X is in state x1 or x2 (independent of S’s state). The corresponding probabilities of t2 are
1 - P(T = t1|X = x1) and 1 - P(T = t1|X = x2).
Then the probability P(T = t1|S = s) that T is in termination state t1, rather than t2, given that the
mechanism begins operation (i.e. S = s), is the sum of the probabilities of the internal pathways
that can lead to T being t1. For example, the probability that T ends up in t1 via X being in state
x1 is the product of the probability that X comes to be in x1 and the probability of t1 given x1:
P(T = t1|X = x1)P(X = x1|S = s). The probability of the mechanism as a whole exhibiting end
state T = t1 given startup condition S = s is:
172
P(T = t1|S = s) =
P(T = t1|X = x1)P(X = x1|S = s) + P(T = t1|X = x2)P(X = x2|S = s) (13.1)
This simply follows mathematically from the above assumptions. Analogous computations for
more complex mechanisms will, of course, be different, but the fact remains that probabilities of
termination conditions given setup conditions are a function of activity probabilities. Thus, focusing solely
on activity probabilities below will allow us to understand setup-to-termination probabilities
as well. Exceptions, mentioned below, concern cases in which an interpretation of probability
defines higher-level probabilities in ways that allow some variation in lower-level probabilities
(Abrams 2015), or when there are activities affecting the mechanism that are not considered to
be part of that mechanism.
I also discuss probabilities involved in constitutive relationships below.
2. Metaphysics
If it’s not necessary that mechanisms “are regular in that they work always or for the most part
in the same way under the same conditions” (Machamer et al. 2000, p. 3), then what is the dif-
ference between what are usually called mechanisms in science, and mere happenstance? I won’t
be able to do justice to several very valuable proposals. For example, Andersen (2012) proposes
that mechanism-hood may be a matter of degree, depending, for example, on the frequency or
probability of certain activity outcomes. Baetu (2013) argues that to be understood as a mecha-
nism, a system must at least be regular enough—in the sense that activities occur and succeed
with a certain relative frequency—that repeated interventions are possible. Krickel (forthcom-
ing) argues that the regularity of an overall mechanism must be such that the phenomenon it
explains occurs as a result of its operation more often than other outcomes, or that the phenom-
enon is brought about more often by that kind of mechanism than by others. I suspect that the
degree and kind of regularity needed for mechanism-hood may also depend on other aspects of
research contexts (cf. Abrams 2012a), even if it turns out that as a general rule of thumb, larger
probabilities explain better than smaller probabilities (Strevens 2000).
Nevertheless, I think that we can get some insight about the nature of stochastic mecha-
nisms by asking questions about the nature of activity probabilities. Are there constraints on the
character of activity probabilities in order for a system to count as a mechanism—even if some
probabilities represented in models of mechanisms need not correspond to reality?
Determinism and interpretations of probability

It will be helpful to begin with a few points about interpretations of probability. Mathematical
probability is defined by a set of axioms (e.g. Grimmett and Stirzacker 1992), and anything that
satisfies these axioms counts as probability. At a minimum, we need a set Ω of basic elements and
a set of subsets (“outcomes”) of Ω closed under unions and intersections. A probability measure
P is any function that assigns 1 to Ω, 0 to the empty set, and such that P(A ∪ B) = P(A) + P(B)
when A and B are disjoint subsets of Ω. An interpretation of probability is (supposed to be) a
specification of a way that a set of properties realized in a part of the world (a “chance setup”)
can satisfy probability axioms. Some of these properties must also determine the numerical
values of probabilities according to the interpretation. I’ll assume familiarity with several well-
known interpretations of probability: Bayesian credence, finite frequency, single-case propen-
sity, long-run propensity, and Best System analysis probabilities (e.g. Hájek 2012; Earman 1992;
Eagle 2004; Berkovitz 2015; Lewis 1980; Hoefer 2007).
173
Marshall Abrams
Figure 13.3 A croupier’s initial velocity distribution for a wheel of fortune. x-axis: angular velocity.
y-axis: frequency or probability of spins at that velocity. Black regions: overall frequency or
probability of spins resulting in the black outcome; gray regions represent the same thing for
the red outcome. This distribution is for a wheel with black wedges that are twice as large as
red wedges, which is why black regions are twice as wide as gray regions
It’s worth mentioning recently developed “complex causal structure” (CCS) interpretations
(e.g. Rosenthal 2010, 2012; Strevens 2011; Abrams 2012b; Beisbart 2016). While single-
case propensities are often viewed as requiring fundamental indeterminism, as is thought to
be present when quantum mechanical effects dominate a process, long-run propensities and
CCS probabilities are usually viewed as consistent with either determinism or indetermin-
ism. Indeed, one of the motivations for these interpretations is to develop a conception of
objective probability that can underpin references to probability in higher-level processes that
supervene on (nearly-)deterministic lower-level processes.5 CCS probabilities seem particu-
larly relevant to mechanisms such as those studied in molecular biology, because it’s plausible
that molecules interacting in fluids often realize the kind of causal structure on which CCS
interpretations are based. Ignoring important subtleties of different CCS interpretations, the
key idea is that some complex activities map input entity states to outcome states in such
a way that small changes in input states easily lead to large differences in outcomes.6 For
example, a wheel of fortune7 can be viewed as part of a mechanism in which a croupier’s
activity initiates an activity by which initial angular velocities are mapped to outcomes (red,
black) in such a way that a small difference in velocity would cause a red rather than a black
outcome. Figure 13.3 shows that unless a croupier were able to impart velocities signifi-
cantly more often in very narrow regions of the input space, the shallow slope of the input
distribution curve over any small contiguous region containing velocities leading to red and
black means that the relative frequencies of spins leading to red and black within that region
will be roughly equal to the relative widths of the intervals along the velocity (x) axis. Since
this is true throughout the input space, frequencies will be close to ratios between wedge
sizes. CCS interpretations use this insensitivity of outcome frequencies to variation in input
frequencies to define interpretations of probability. Complex interactions between numerous
swiftly moving molecules in fluid probably give rise to an analogous causal structure for many
outcomes (cf. Strevens 2003, 2013). (CCS interpretations are not without serious challenges,
but neither are the well-known interpretations mentioned above.)
174
Activity probabilities as causal probabilities

Despite the diversity of mechanisms in the world, their character and the way in which they are
studied generally places constraints on what interpretations of probability are suitable as analyses
of activity probabilities. Investigation of mechanisms typically involves interventions, manipu-
lating either entities or activities and measuring resulting changes (Craver 2007; DesAutels
2011; Baetu 2013). When activities are stochastic, what must be measured are the effects of
the intervention on frequencies of outcomes (Baetu 2013). This is true whether we are con-
cerned with a token mechanism examined over time or a class of similar mechanisms. If we
view these manipulable frequencies as depending on probabilities realized in the mechanism,
then any account of what activity probabilities are must allow that manipulating probabilities—
by manipulating properties that realize them—can also manipulate frequencies. In particular,
manipulating the probability of outcome A should usually produce frequencies of A that are
close to its probability. That manipulating probabilities manipulates frequencies in this way
is true only for some interpretations of probability; I call such interpretations causal probability
interpretations, and the probabilities they define causal probabilities (Abrams 2015). (This is a
difference-making sense of “causal.”)
Causal probability is not an interpretation of probability, but designates a property that
applies to probabilities defined by some interpretations but not others. Causal probability super-
venes on such interpretations. Causal probabilities can at least partially explain frequencies, in
that the probabilities, as realized in the chance setup, partially control the frequencies. However,
it’s not required that a causal probability of an outcome always be roughly equal to its frequency
in a large number of trials. That’s just what usually happens. Despite the need for philosophical
work on such claims, the vague idea that frequencies are usually roughly equal to probabilities,
and manipulable by the means for manipulating probabilities, is implicit in assumptions that are
widespread in successful science (Abrams 2015).
Which interpretations are causal probability interpretations? Notice, for example, that manipu-
lating a person’s credence that tosses of a coin will produce heads needn’t affect frequencies of
heads in tosses of that coin. Since manipulation of credences needn’t affect frequencies, credences
are not causal probabilities. Finite frequencies aren’t causal probabilities either; in this case, the
probabilities are the frequencies, so there is nothing further that can be affected by manipulating
probabilities. Single-case and long-run propensities are plausibly causal probabilities, however.
Consider the chance setup that consists of checking whether a particular Geiger counter clicks
during any given minute in a particular day, with the possible outcomes click, no click. Suppose
that the probabilities of these outcomes are single-case propensities. We can manipulate both this
propensity and the usual frequencies of click by changing the radioactive substances that are near or
in the Geiger counter, so such propensities are causal probabilities. CCS probabilities are also plau-
sibly causal probabilities. For example, by manipulating the sizes of wedges on a roulette wheel, we
can manipulate the probability of red as well as (usually) its relative frequency. On the other hand,
Best System Analysis probabilities may not be causal probabilities if manipulating probabilities in a
Best System just is manipulating frequencies in certain ways.
Causal probability helps to resolve a potential problem arising when different interpretations
of probability apply to distinct activities in the same mechanism. Consider the activity prob-
abilities involved in E. coli chemotaxis. These probabilities would usually be thought to be due
to the statistical mechanical process of Brownian motion, which causes molecules to move rap-
idly through the cytoplasm (intracellular liquid) in a stochastic fashion. This is true of transitions
between the active and inactive conformations of the Tar receptor as well (Hoffmann 2012, ch. 4).
Quantum mechanical effects in this transition are small enough that they are difficult to calculate
175
Marshall Abrams
(Kuriyan et al. 2013, ch. 6). Nevertheless, suppose it turned out that probabilities for active/inactive
state shifts were quantum mechanical (cf. Jeknić-Dugić 2009; Luo 2014), and that they should be
understood as propensities. Suppose that other chemotaxis activity probabilities, such as the proba-
bility of a CheY molecule becoming phosphorylated and subsequently binding to a f lagellar motor,
were CCS probabilities. We can capture this distinction by treating the probabilities on the right-
hand side of equation (13.1) as being of different kinds. What is the nature of the overall probability
P(T = t1|S = s) of a CheY molecule binding to the flagellar motor given the Tar receptor’s new
aspartate binding? It appears to be some kind of weird mixture of propensity and CSS probabilities.
Note that it won’t do to say that P(T = t1|S = s) is a single-case propensity simply because single-
case propensities are more fundamental if, as is often assumed, single-case propensities depend on all
factors that might affect what happens (Berkovitz 2015). This would seem to restrict the probability
of a CheY-bound flagellar motor to values near 0 or 1, since for any particular token bacterium at a
particular moment, the future movements of molecules would be largely determined by their cur-
rent spatial relationships and states of interaction. Yet our current best understandings of chemotaxis
treat these probabilities as non-extremal.
The concept of causal probability allows us to sidestep the preceding puzzle in a way that
captures what is crucial about probability in mechanisms. A complex chance setup consists of sev-
eral chance setups connected in such a way that whether a trial occurs on one or more setups
depends on the outcome of an earlier trial on another setup (Abrams 2015). This is what a
mechanism involving probabilistic activities is. For example, we can view the active and inactive
states of the Tar complex as defining two setups, so that P(T = t1|X = x1) and P(T = t1|X = x2)
are probabilities relative to two different setups, chosen by an outcome X = xi with probability
P(X = xi|S = s). Then the overall probability P(T = t1|S = s) of CheY-flagellar binding given
aspartate binding is the probability for an outcome of this complex chance setup.
I argued in (Abrams 2015) that probabilities of outcomes from complex chance setups are
causal probabilities if all of the probabilities from the component setups are causal probabilities.
Since each component chance setup allows one to manipulate probabilities and frequencies in a
generally coordinated way, and since frequencies of trials on later chance setups are controlled
by frequencies of outcomes of trials on certain earlier chance setups, manipulating any of these
component probabilities manipulates both probabilities and frequencies for the complex chance
setup. So if activity probabilities are typically causal probabilities, the probability of a mechanism
producing a phenomenon, for example, would typically be a causal probability as well, regardless
of what interpretations of probability apply to the individual component activity probabilities.
Baetu (2013) argues that frequencies of outcomes of activities must be high enough that
interventions can make an observable difference in frequencies, and Andersen (2012) and
Krickel (forthcoming) also suggest that whether something counts as a mechanism depends
on rough limits to the values of probabilities involved in its operation. I suggest that as long
as activity probabilities are causal probabilities—so that they can make a difference in repeated
tokenings of activities—a lower limit to probability values would depend on the nature of the
mechanism and the phenomenon to be understood. Low probabilities can be investigated using
larger samples, more money, advanced statistical methods, or convergent evidence.
Erraticity and imprecise probability

Objective probabilities, including causal probabilities, are only well-defined relative to
(a) a chance setup and (b) a set of outcomes. Must every chance setup and set of outcomes define
a set of causal probabilities? This seems unwarranted. Suppose I devise a way of classifying pieces
of paper into three mutually exclusive categories, low, medium, high, according to the percentage
176
of the paper’s mass that consists of ink. Is there some causal probability that the average ink level
among pieces of paper carried by the next ten people seen after reading this paragraph will be
high? I don’t see why there must be. (The question is not whether the probability of high is 0 or
1 for the nearly deterministic token process that you and your surroundings realize for the next
day or two. That process isn’t repeatedly realizable. Instead, think of the chance setup here as
analogous to a coin toss, realized every time someone reads this paragraph.)
This example suggests that not all chance setups must give outcomes causal probabilities of
some kind. Perhaps for some setups, outcomes are erratic (Hájek and Smithson 2012): Their
frequencies have none of the systematicity that probabilities usually produce. The processes
that produce them exhibit erraticity. For example, for all I know it may be that for a given strain
of E. coli, there is a determinate probability that the number of Tar receptors will lie within a
certain range, but the precise number of receptors within that range is erratic. This is not to
say that there would be no reason at all for erratically produced outcomes in particular tokenings
of the chance setup. In every single trial, these outcomes might be the result of a deterministic
course of events, or might have single-case propensities close to 0 or 1, for example. However,
where outcomes are erratically produced, the setup type and the outcomes are not such that
realizations of it would usually exhibit the kind of systematicity that’s typical of outcomes that
have causal probabilities.
We can also consider a kind of systematicity intermediate between causal probability and
pure erraticity. Causal imprecise probability exists, for example, when an erratically determined
outcome X = x in turn determines which of several different chance setups, with different
probability distributions, will give rise to a subsequent outcome Y = y. For this complex chance
setup, the imprecise probability of Y can be defined by the set of precise probabilities for Y that
the value X might determine. In (Abrams MS) I provide details of this proposal.8 Suppose, for
example, that the number of Tar receptors in bacteria of a strain of E. coli is erratic as above, but
that each number of Tar receptors defines a different probability distribution over frequencies
of phosphorylated CheY molecules for a given aspartate gradient. Then it would turn out that
for this E. coli strain, some of the activities that lead to the production of phosphorylated CheY
would involve imprecise probabilities, neither fully probabilistic nor fully erratic.
For a mechanism to generate characteristic phenomena, not all of its activities need to involve
(precise) probabilities. Biological and other mechanisms often involve entities and activities
that are robust to variation in those entities and activities that produce them (Wagner 2005;
Hermisson and Wagner 2004; cf. Wimsatt 1980, 2007). For example, average values of cer-
tain activities in the E. coli chemotaxis mechanism seem to be robust to some variations in the
chemical composition of a bacterium (Barkai and Leibler 1997; Alon et al. 1999; Yi et al. 2000).
Where this kind of robustness exists, a mechanism can function in a normal manner whether or
not causally prior activities involve probabilities, imprecise probabilities, or full-fledged erratic-
ity, as long as these activities’ outcome frequencies tend to remain within certain ranges. Note,
however, that imprecisely probabilistic activities can be modeled using probabilities when, for
each outcome of an activity, all of the precise probabilities in the set that makes up the out-
come’s imprecise probability have similar values.
3. Epistemology
Modeling probabilistic mechanisms

Once we start thinking of mechanistic activities as probabilistic, it’s natural to wonder whether
Bayesian network models can provide a general way of modeling the causal structure of
177
Marshall Abrams
mechanisms. These models are defined in terms of mathematical assumptions about probabilistic
relationships represented by arrows between nodes or variables (Pearl 2009; Spirtes et al.
2000). For example, Figure 13.2 and the assumptions about probabilities that I specified for it
made it a representation of a Bayesian network. However, many mechanisms involve causal
cycles (Bechtel and Abrahamsen 2005; Bechtel 2011). Bayesian networks can’t directly rep-
resent cycles, but the framework can be extended to model causal cycles, though in slightly
unnatural ways (Clarke et al. 2014; Gebharter and Kaiser 2014; Gebharter and Schurz 2016).
Further, cyclic processes (such as those involved in the E. coli chemotaxis mechanism) can
involve activities for which timing is important (Bechtel and Abrahamsen 2005; Bechtel
2011; Weber 2016), and current extensions of Bayesian network representations seem inad-
equate for modeling timing (Weber 2016; Gebharter and Kaiser 2014).
Bechtel (2011) notes that systems of differential equations are often appropriate for mod-
eling such cyclic systems, and suggests that other ideas from dynamical systems theory are also
useful. Differential equations can be used to model probabilistic systems, as Barkai and Leibler
(1997) did for chemotaxis, but Firth and Bray (Morton-Firth and Bray 1998; Firth and Bray
2001) argued that differential equation models such as Barkai and Leibler’s can make inaccurate
predictions for aspects of intracellular processes that involve small numbers of molecules. Firth
and Bray developed StochSim, an agent-based model in which “agents” are Tar receptors and
molecules of several kinds with multiple states (Morton-Firth and Bray 1998; Morton-Firth
et al. 1999; Firth and Bray 2001). Timing and outcomes of interactions between molecules in
StochSim are probabilistic, with probabilities represented explicitly in the model. This model
was able to reproduce several aspects of the empirical data on chemotaxis better than other
models such as Barkai and Leibler’s, suggesting further investigations.
Markov process theory also provides a wealth of mathematical tools for representing proba-
bilistic processes such as those found in some mechanisms (e.g. Grimmett and Stirzacker 1992;
Bharucha-Reid 1996 [1960]). For example, diffusion process models were originally developed
to model Brownian motion, which plays a significant role in E. coli chemotaxis. Perhaps some
Markov process models or computer simulations can capture some of the subtle complexity
that comes from the role of spatial relationships in some mechanisms, which is difficult to cap-
ture with Bayesian networks (Kaiser 2016). In any event, we needn’t think that there is some
one privileged or standard way of modeling mechanisms. Just as there can be different sorts of
diagrams that are useful for modeling mechanisms—even for modeling the same mechanism—
there are many different kinds of models that may be useful for modeling mechanisms.
Modeling constitutive relationships

Discussion of mechanistic explanation sometimes focuses on ways in which some entities and
activities constitute higher-level entities and activities (see Chapters 9 and 14). Craver (2007)
proposes that constitutive relationships be understood in terms of manipulating entities and
activities at one level and producing changes in entities at other levels. Since this view derives
from Woodward’s (Woodward 2003; Woodward and Hitchcock 2003) account of manipula-
tion, in which interventions can affect probabilities, we can ask whether the state of an entity
at one level could depend probabilistically on the state of entities that compose it, or vice versa.
There has in fact been recent research on ways to extend Bayesian networks to represent certain
probabilistic constitutive relationships in mechanisms (e.g. Casini et al. 2011; Gebharter 2014;
Gebharter and Kaiser 2014).
There is a constraint on such models, unrecognized by those investigating them as far as I can
see, which seems illuminating in its own right. The properties of an entity X at a higher level
178
must supervene on the lower-level entities Yi that constitute it and on the activities ψk involving
these lower-level entities. Thus, given a specific set of lower-level entities Yi in specific states yj
involved in specific activities ψk, the properties of the higher-level entity X that the Yi consti-
tute are determined. This means that the probability of X having the state constituted by those
lower-level states Yi = yj is equal to the probability of the conjunction of those states. Further,
that probability is a mathematical function of activity probabilities whose outcomes determine
whether Y1 = y5 and Y2 = y7, etc. Similarly, a probability that some lower-level entities Yi have
particular states given that a higher-level entity has a particular state X = x3 is just the probability
that lower-level activities cause the Yi to have certain states, where the lower-level processes are
restricted to those combinations of activities that can cause X to have state x3. These stated iden-
tities between probabilities are not merely mathematical; they are identities between objective
probabilities realized in the mechanism. Thus, probabilities concerning constitutive relationships are
really activity probabilities. Models that postulate probabilities concerning constitutive relationships
should be able to address this point.
Learning about activity probabilities

Background theory and similarity to related mechanisms may determine or suggest that certain
activities are likely to be probabilistic, and that manipulation of conditions alters probabilities
and frequencies in a coordinated fashion. For example, a transmembrane receptor in E. coli, the
Tsr receptor, is known to have many structural and functional similarities to the Tar receptor
(Parkinson et al. 2015), so it could be reasonable to assume that certain activity probabilities in
the Tsr mechanism are similar to corresponding probabilities in E. coli. However, background
theory or comparison to other mechanisms ultimately rests on earlier investigations, and there-
fore on other methods for learning probabilities. A few further details about E. coli chemotaxis
research provide illustrations.
A natural way to measure causal probabilities is to measure how frequencies are manipu-
lated by interventions, but this process is not always simple, and background theory can play
a role. Consider Firth and Bray’s (Morton-Firth and Bray 1998) use of Borkovich and col-
leagues’ results (Borkovich et al. 1989, 1992). The Borkovich teams measured relationships
between aspartate concentrations, numbers of methyl site bindings, and frequencies of CheY
phosphorylation. These researchers took the last two frequencies to be manipulable both
by varying aspartate concentrations and by varying methylation states of Tar receptors. The
researchers didn’t directly measure frequencies of active/inactive shifts in the Tar receptor,
but background theory on the relationship between the Tar receptor’s active and inactive
conformations and CheY phosphorylation allowed Firth and Bray to use the Borkovich
results to estimate probabilities of Tar receptor state changes in their computer simulation.
Firth and Bray (apparently) assumed that the manipulated frequencies reflected probabilities,
but used background theory to justify attributing probabilities to outcomes whose frequencies
had not been measured. The same example also illustrates a role that computer simulations
can play in learning about probabilities.
First note that in a computer simulation, probabilities are modeled using software imple-
menting a deterministic random number-generating algorithm, but it’s reasonable to assume in
such cases that there is some interpretation of probability (perhaps yet unknown) in terms of
which this software can be viewed as realizing a part of chance setup. The simulation doesn’t
just model probabilities; it realizes them as well (Glennan 1997; Abrams 2015).
When running their simulations Firth and Bray manipulated parameters so as to manipu-
late probabilities that were derived from the active/inactive switch probabilities. They then
179
Marshall Abrams
analyzed patterns of frequencies in results of simulation experiments. Because of the way in

which frequencies were manipulated by manipulating probabilities, it appears that probabilities
in the model are causal probabilities. Since these causal probabilities model probabilities in
real E. coli, the probabilities that Firth and Bray assumed to exist in E. coli appear to be causal
probabilities as well, with values like those in the simulations (cf. Abrams 2015).9 In this way
causal probabilities in real E. coli that had not been mathematically estimated from data were
estimated through inference from a combination of observed frequencies, background theory,
and computer simulation results.
Not all probabilities involved in a mechanism or its models must be causal probabilities.
Sometimes a frequency is just a frequency. For example, one causal factor in bacterial chemo-
taxis is the frequency of sites on a Tar receptor that are methylated at a particular time. This is
simply a frequency—nothing more. Further, some models incorporate probabilities assumed
for the sake of convenience or simplicity that are not thought to correspond to any prob-
abilities in the world. Morton-Firth and Bray’s (1998) chemotaxis simulation incorporated
“pseudo-molecule” software objects that could be used to model spontaneous changes in a
single molecule. The number of pseudo-molecules was chosen so that the probabilities of
single-molecule changes would not result in these changes happening at a significantly slower
rate than interactions between molecules. The probabilities of single-molecule changes in the
model were not motivated entirely by theoretical or empirical factors; they were adjusted to
accommodate the fact that the simulation modeled parallel, mostly independent activities as a
sequence of activities that took place one at a time.
4. Non-recurrent mechanisms
So far I’ve focused on recurrent mechanisms, in which the same processes are repeated and
realized either in a token mechanism or in many similar mechanisms. Some authors (Glennan
2009, forthcoming; Illari and Williamson 2012) advocate a broad conception of mechanism
that need not involve the likelihood of recurrence. So defined, entities and activities might
be organized in an entirely transient manner, and recurrence of such a mechanism could be
unlikely. For example, Glennan (2009) argues that there is a mechanistic explanation of the
literary critic Roland Barthes’ accidental death due to being hit by a truck on the way home in
Paris. The non-recurrent mechanism that produced Barthes’ death is what Glennan (2009) calls
an “ephemeral mechanism,” where “the configuration of parts may be the product of chance
or exogenous factors” and is “short-lived and non-stable, and is not an instance of a multiply-
realized type” (Glennan 2009, p. 260).
As Glennan (2009, forthcoming) suggests, even where a non-recurrent mechanism consists
of entities and activities in a short-lived configuration, each individual activity might neverthe-
less be of a kind that is quite systematic. The coming together of these entities and activities
might even be due to erraticity. Thus even if a mechanism as a whole is non-recurrent, it
could consist of activities such as certain ways of trucks hitting pedestrians that regularly par-
ticipate in mechanisms. For such activities, at least, we can apply arguments above that activity
probabilities are causal probabilities discoverable through manipulation and modeling. Even if
an activity in fact occurs only once, it may involve a causal probability, since that concept is
defined in terms of counterfactuals about what frequencies usually would be seen in repeated
trials. However, activities in non-recurrent mechanisms might also involve causal imprecise
probabilities in some cases. Or some activities in non-recurrent mechanisms could simply be
erratic—even so erratic that what happens just happened as it did for no reason other than the
idiosyncrasies of lower-level processes.
180
5. Conclusion
This chapter looked at stochastic mechanisms through a discussion of the nature of probabilistic
and other stochastic activities, primarily in recurrent mechanisms. I argued that activity probabili-
ties in such mechanisms are nearly always what I call “causal probabilities,” except, occasionally,
when they are bare relative frequencies. I also argued that some activities might not be probabil-
istic, but instead involve imprecise probabilities or constrained erraticity. I discussed methods for
modeling probabilistic mechanisms, and described several ways of learning about probabilities in
mechanisms. Among other things, I illustrated a role that computer simulations can play in this
process. I also suggested ways that the preceding ideas can be extended to ephemeral mechanisms
that are unlikely to recur.
Notes
1 I’m grateful for Stuart and Phyllis’s invitation to contribute this chapter, which drew me into fruitful
and rewarding investigations I wouldn’t have pursued otherwise, and for their feedback and patience
during its writing. I got helpful feedback on ideas presented here from Holly Andersen, Murat Aydede,
Paul Bartha, Sylvia Berryman, Scott Dixon, Kenny Easwaran, Chris French, Bruce Glymour, Alan Hajek,
Chris Hitchcock, Daniel Malinsky, Ron Mallon, David McElhoes, Scott Peck, Alirio Rosales, Beckett
Sterner, Johanna Thoma, and Jerry Tsui. Don Muccio and Scott Brande helped me understand the
(nearly nonexistent) role of quantum mechanics in cellular processes.
2 My sketch is based on Eisenbach 1996; Barkai and Leibler 1997; Morton-Firth and Bray 1998;
Morton-Firth et al. 1999; Firth and Bray 2001; Bray 2009; Parkinson Lab 2015.
3 E. coli’s so-called “exact” or “perfect” adaptation to chemical concentrations (e.g. Yi et al. 2000) refers
to robustness in its ability to return to roughly the same state, on average, after persistent exposure to
a concentration of a chemical such as aspartate. The states returned to exhibit significant stochastic
variation (Alon et al. 1999).
4 See e.g. Bogen 2005; Craver 2007; Illari and Williamson 2012; Andersen 2012; Barros 2008; Baetu 2013;
DesAutels 2011, 2015.
5 Rosenthal (2012) and I (2012b) have argued that our CCS interpretations don’t define single-case
probabilities, but Strevens (2011) argues that his does.
6 When I called my own CCS interpretation “mechanistic probability,” I meant “mechanistic” in a very
loose sense. CCS interpretations do always depend on processes that can be considered mechanisms
according to some conception of mechanism or other, but these mechanisms would generally operate
at a lower level than the activities that involve CCS probabilities. Because of this and because setup-to-
termination probabilities of a mechanism reduce to activity probabilities (see above), I’m skeptical of
DesAutels’ (2015) apparent suggestion that such setup-to-termination probabilities might be mechanistic
probabilities, unless component activity probabilities are CCS probabilities.
7 A wheel of fortune is like a roulette wheel, but with a fixed pointer rather than a ball indicating the
wedge representing the outcome.
8 Related ideas and relevant mathematical details can be found e.g. in Troffaes and de Cooman 2014;
Fierens et al. 2009; Bradley 2015, 2016; Dardashti et al. 2014.
9 I believe that analogous arguments can be given using mathematical models rather than simulations.
References
Abrams, M. (2012a), Measured, modeled, and causal conceptions of fitness, Frontiers in Genetics 3(196),
1–12.
Abrams, M. (2012b), Mechanistic probability, Synthese 187(2), 343–375.
Abrams, M. (2015), Probability and manipulation: Evolution and simulation in applied population
genetics, Erkenntnis 80(S3), 519–549.
Abrams, M. (MS), Imprecise probability and biological fitness. draft paper.
Alon, U., Surette, M. G., Barkai, N. and Leibler, S. (1999), Robustness in bacterial chemotaxis, Nature
397(6715), 168–171.
181
Marshall Abrams
Andersen, H. (2012), The case for regularity in mechanistic explanation, Synthese 189, 415–432.
Baetu, T. M. (2013), Chance, experimental reproducibility, and mechanistic regularity, International Studies
in the Philosophy of Science 27(3), 253–271.
Barkai, N. and Leibler, S. (1997), Robustness in simple biochemical networks, Nature 387(6636), 913.
Barros, D. B. (2008), Natural selection as a mechanism, Philosophy of Science 75(3), 306–322.
Bechtel, W. (2011), Mechanism and biological explanation, Philosophy of Science 78, 533–557.
Bechtel, W. and Abrahamsen, A. (2005), Explanation: A mechanist alternative, Studies in History and
Philosophy of Science Part C: Studies in History and Philosophy of Biological and Biomedical Sciences 36(2),
421–441.
Beisbart, C. (2016), A Humean guide to Spielraum probabilities, Journal of General Philosophy of Science
47(1), 189–216.
Berkovitz, J. (2015), The propensity interpretation of probability: A re-evaluation, Erkenntnis 80(S3),
629–711.
Bharucha-Reid, A. T. (1996 [1960]), Elements of the Theory of Markov Processes and their Applications, Dover.
Bogen, J. (2005), Regularities and causality: Generalizations and causal explanations, Studies in History
and Philosophy of Science Part C: Studies in History and Philosophy of Biological and Biomedical Sciences 36,
397–420.
Borkovich, K. A., Kaplan, N., Hess, J. F. and Simon, M. I. (1989), Transmembrane signal transduction in
bacterial chemotaxis involves ligand-dependent activation of phosphate group transfer, Proceedings of the
National Academy of Sciences 86(4), 1208–1212.
Borkovich, K. A., Alex, L. A. and Simon, M. I. (1992), Attenuation of sensory receptor signaling by cova-
lent modification, Proceedings of the National Academy of Sciences 89(15), 6756–6760.
Bradley, S. (2015), Imprecise probabilities, in E. N. Zalta (ed.), The Stanford Encyclopedia of Philosophy,
Summer 2015 edition.
Bradley, S. (2016), Vague chance?, Ergo 3(18–21), 524–538.
Bray, D. (2009), Wetware: A Computer in Every Living Cell, Yale University Press.
Casini, L., Illari, P. M., Russo, F. and Williamson, J. (2011), Models for prediction, explanation and
control: Recursive Bayesian networks, THEORIA. An International Journal for Theory, History and
Foundations of Science 26(1), 5–33.
Clarke, B., Leuridan, B. and Williamson, J. (2014), Modelling mechanisms with causal cycles, Synthese
191(8), 1651–1681.
Craver, C. F. (2007), Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience, Oxford University
Press.
Dardashti, R., Glynn, L., Thebault, K. and Frisch, M. (2014), Unsharp Humean chances in statistical
physics: A reply to Beisbart, in M. C. Galavotti, D. Dieks, W. J. Gonzalez, S. Hartmann, T. Uebel
and M. Weber (eds), New Directions in the Philosophy of Science, Springer, pp. 531–542.
DesAutels, L. (2011), Against regular and irregular characterizations of mechanisms, Philosophy of Science
78(5), 914–925.
DesAutels, L. (2015), Toward a propensity interpretation of stochastic mechanism for the life sciences,
Synthese, 192(9), 2921–2953.
Eagle, A. (2004), Twenty-one arguments against propensity analyses of probability, Erkenntnis 60(3),
371–416.
Earman, J. (1992), Bayes or Bust, MIT Press.
Eisenbach, M. (1996), Control of bacterial chemotaxis, Molecular Microbiology 20(5), 903–910.
Fierens, P. I., Rêgo, L. C. and Fine, T. L. (2009), A frequentist understanding of sets of measures, Journal
of Statistical Planning and Inference 139, 1879–1892.
Firth, C. A. J. M. and Bray, D. (2001), Stochastic simulation of cell signaling pathways, in J. M. Bower
and H. Bolouri (eds), Computational Modeling of Genetic and Biochemical Networks, MIT Press, chapter 9,
pp. 263–286.
Gebharter, A. (2014), A formal framework for representing mechanisms?, Philosophy of Science 81(1),
138–153.
Gebharter, A. and Kaiser, M. I. (2014), Causal graphs and biological mechanisms, in M. I. Kaiser,
O. Scholz, D. Plenge and A. Hüttemann (eds), Explanation in the Special Sciences: The Case of Biology and
History, Springer, pp. 55–86.
Gebharter, A. and Schurz, G. (2016), A modeling approach for mechanisms featuring causal cycles,
Philosophy of Science 83(5), 934–945.
182
Glennan, S. (1997), Probable causes and the distinction between subjective and objective chance, Noûs
31(4), 496–519.
Glennan, S. (2009), Ephemeral mechanisms and historical explanation, Erkenntnis 72(2), 251–266.
Glennan, S. (forthcoming), The New Mechanical Philosophy. Oxford University Press.
Grimmett, G. R. and Stirzacker, D. R. (1992), Probability and Random Processes, 2nd edition, Oxford
University Press.
Hájek, A. (2012), Interpretations of probability, in E. N. Zalta (ed.), The Stanford Encyclopedia of Philosophy,
Winter 2012 edition. http://plato.stanford.edu/archives/win2012/entries/probability-interpret/.
Hájek, A. and Smithson, M. (2012), Rationality and indeterminate probabilities, Synthese 187(1), 33–48.
Hermisson, J. and Wagner, G. P. (2004), The population genetic theory of hidden variation and genetic
robustness, Genetics 168(4), 2271–2284.
Hoefer, C. (2007), The third way on objective probability: A sceptic’s guide to objective chance, Mind
116(463), 449–596.
Hoffmann, P. M. (2012), Life’s Ratchet: How Molecular Machines Extract Order From Chaos, Basic Books.
Illari, P. M. and Williamson, J. (2012), What is a mechanism? Thinking about mechanisms across the sci-
ences, European Journal for Philosophy of Science 2(1), 119–135.
Jeknić-Dugić, J. (2009), The environment-induced-superselection model of the large-molecules confor-
mational stability and transitions, The European Physical Journal D 51(2), 193–204.
Kaiser, M. I. (2016), On the limits of causal modeling: Spatially-structurally complex phenomena,
Krickel, B. (forthcoming), A regularist approach to mechanistic type-level explanation, British Journal for
the Philosophy of Science.
Kuriyan, J., Konforti, B. and Wemmer, D. (2013), The Molecules of Life: Physical and Chemical Principles, 1st
edition, Garland Science.
Lewis, D. (1980), A subjectivist’s guide to objective chance, in R. C. Jeffrey (ed.), Studies in Inductive Logic
and Probability, Vol. II, University of California Press. Reprinted in (Lewis 1986).
Lewis, D. (1986), Philosophical Papers, volume II, Oxford University Press.
Luo, LiaoFu (2014), Quantum theory on protein folding, Science China Physics, Mechanics and Astronomy
57(3), 458–468.
Machamer, P., Darden, L. and Craver, C. F. (2000), Thinking about mechanisms, Philosophy of Science
67(I), 25.
Morton-Firth, C. J. and Bray, D. (1998), Predicting temporal fluctuations in an intracellular signalling
pathway, Journal of Theoretical Biology 192(1), 117–128.
Morton-Firth, C. J., Shimizu, T. S. and Bray, D. (1999), A free-energy-based stochastic simulation of the
Tar receptor complex, Journal of Molecular Biology 286(4), 1059–1074.
Parkinson, J. S., Hazelbauer, G. L. and Falke, J. J. (2015), Signaling and sensory adaptation in Escherichia
coli chemoreceptors 2015 update, Trends in Microbiology 23(5), 257–266.
Parkinson Lab (2015), An overview of E. coli chemotaxis, http://www.pdn.cam.ac.uk/groups/comp-cell/
StochSim.html. University of Utah Department of Biology, John Parkinson Lab. Downloaded October
17, 2015.
Pearl, J. (2009), Causality: Models, Reasoning, and Inference, 2nd edition, Cambridge University Press.
Rosenthal, J. (2010), The natural-range conception of probability, in G. Ernst and A. Hüttemann (eds),
Time, Chance, and Reduction: Philosophical Aspects of Statistical Mechanics, Cambridge University Press,
pp. 71–90.
Rosenthal, J. (2012), Probabilities as ratios of ranges in initial-state spaces, Journal of Logic, Language, and
Inference 21, 217–236.
Spirtes, P., Glymour, C. N. and Scheines, R. (2000), Causation, Prediction, and Search, 2nd edition, MIT
Press.
Strevens, M. (2000), Do large probabilities explain better?, Philosophy of Science 67(3), 366–390.
Strevens, M. (2003), Bigger Than Chaos: Understanding Complexity through Probability, Harvard University
Press.
Strevens, M. (2011), Probability out of determinism, in C. Beisbart and S. Hartmann (eds), Probabilities in
Physics, Oxford University Press, chapter 13, pp. 339–364.
Strevens, M. (2013), Tychomancy: Inferring Probability from Causal Structure, Harvard University Press.
Troffaes, M. C. M. and de Cooman, G. (2014), Lower Previsions, Wiley.
Wagner, A. (2005), Robustness and Evolvability in Living Systems, Princeton University Press.
183
Marshall Abrams
Weber, M. (2016), On the incompatibility of dynamical biological mechanisms and causal graphs,
Wimsatt, W. C. (1980), Randomness and perceived randomness in evolutionary biology, Synthese 43,
287–329.
Wimsatt, W. C. (2007), Re-Engineering Philosophy for Limited Beings: Piecewise Approximations to Reality,
Woodward, J. (2003), Making Things Happen, Oxford University Press.
Woodward, J. and Hitchcock, C. (2003), Explanatory generalizations, Part I: A counterfactual account,
Noûs 37(1), 1–24.
Yi, T.-M., Huang, Y., Simon, M. I. and Doyle, J. (2000), Robust perfect adaptation in bacterial chemo-
taxis through integral feedback control, Proceedings of the National Academy of Sciences 97(9), 4649–4653.
184
14
MECHANISTIC LEVELS,
REDUCTION, AND
EMERGENCE1
1. Why levels?
In The Sciences of the Artificial, Herbert Simon (1969) offers the parable of the watchmakers,
Tempus and Hora. Tempus builds watches holistically, holding each part in place until it forms a
stable whole. Hora builds watches modularly, first assembling stable components and then organ-
izing them into watches. Each is interrupted now and then. When Tempus is interrupted, she
loses all her work on the watch and has to start again from scratch. When Hora is interrupted, she
loses only her work on a single component. Hora thrives; Tempus struggles. The Moral: Evolved
systems are likely to have (nearly) decomposable architectures. Their working parts are likely to
be organizations of working parts, which are themselves organizations of parts, and so on. In
other words, evolved systems likely exhibit regular mechanistic organization at multiple levels.
One might object to this argument on the ground that watches do not reproduce or that bio-
logical evolution does not assemble components sequentially (Bechtel and Stufflebeam 2001).
As Simon points out, this would miss his point, which has now been repeated across many
domains of science and philosophy: viz., that nearly decomposable (or modular) systems are
more stable than holistic systems; they are to be expected when we find improbable stability in
the face of random challenges. Contemporary work in network analysis, for example, shows
that modular networks (those with high intra-modular connectivity and sparse inter-modular
connectivity) are more “robust” than networks that are less modular; e.g., their mean minimum
path length remains low in the face of random attacks. This is important, for example, in tele-
communication, air traffic control, and neural networks (Albert et al. 2000). Such “modularity”
is a widely accepted principle of both development and evolution (see Schlosser and Wagner
2004). Systems with causally independent parts are expected, in part, because nearly decompos-
able components can be independently assembled, regulated, damaged, and repaired without
disturbing the other components. In short, everywhere we look from abstract theoretical con-
siderations to concrete details of biology, we find reasons to expect dynamic but stable systems
to be arranged in nearly decomposable hierarchies. Where we find embedded decompositions
of working parts within working parts, we find nested mechanistic explanations.
Indeed, the assumption of near decomposability underlies the strategy of reverse engineer-
ing, of discovering how something works by learning how its parts interact. Kauffman (1970)
calls this practice “articulation of parts explanation”; Haugeland (1998) calls it “explanation by
185
system decomposition”; Cummins (1975) calls it “functional analysis”; Fodor (1968), Craver
(2007), and others (Machamer 2004; Glennan 2002; Menzies 2012) call it “mechanistic expla-
nation.” Mechanistic explanations explain a phenomenon by situating it in the causal structure
of the world. In constitutive mechanistic explanations (as opposed to etiological and contextual
mechanistic explanations; see Craver 2001), one looks to a lower level, within the phenom-
enon, to reveal its internal causal structure. In many such systems, one can explain the behavior
of the whole in terms of the organized behaviors of its parts, and one can explain the behaviors
of the parts in terms of the organized behaviors of their parts. Levels, on this view, are simply
embedded mechanistic explanations.
In this chapter, we explicate this mechanistic view of levels (section 2), contrast it with other
senses of “level” (section 3), and sketch its implications for emergence (section 4) and reduction
(section 5), for the ontological status of higher-level phenomena (section 6), and for thinking
about the lowest level(s) in such hierarchies (section 7).
2. Mechanistic levels
We use the term “mechanism” permissively to describe causal systems in which parts are organ-
ized such that they collectively give rise to the behavior or property of the whole in context
(cf. the notion of minimal mechanism in Chapter 1). As mentioned above, not all mechanisms
are modular at intermediate levels. In regular networks of simple nodes, for example, there is
no nearly decomposable structure between the behavior of the mechanism as a whole and the
activities of the parts. All mechanisms involve the organization of interacting parts, but not all
mechanisms have intermediate levels of organization.
Mechanisms so construed contrast with aggregates (see section 3) in that mechanisms are
organized. The parts of mechanisms have spatial (e.g., location, size, shape, and motion),
temporal (e.g., order, rate, and duration), and active or otherwise causal (e.g., feedback
or other motifs of organization; see Levy and Bechtel 2013) relations to one another such
that they work together. This is why mechanistic hierarchies are often called “levels of
organization.” Aggregates, in contrast, are the lower limit of mechanistic organization: No
between-component interactions are relevant to the true aggregate.
Figure 14.1 depicts three mechanistic levels. At the top is the activity of some mechanism
as a whole (S’s ψ-ing). S’s ψ-ing includes the mechanism as a whole (e.g., the protein synthesis
mechanism) doing what it does (e.g., synthesizing proteins). In this diagram, ψ is the topping-
off activity of the mechanism; it is the activity to which all of the lower-level components
are relevant. Beneath S’s ψ-ing are the activities (ϕi) and components (Xj) organized such that
together they ψ. Beneath that is an iteration of the levels relation, in which one of the X’s ϕ-ings
is decomposed into the organized ρ-ings (rho-ings) of Ps. As Simon suggested, this process of
decomposition might continue until we run out of either relevant or practically salient decom-
positions. Individually, the Xj and ϕi are organized to compose S’s ψ-ing. Collectively, the Xj
and ϕi exhaustively constitute (or, as is often said, “realize”) S’s ψ-ing in context.
X’s ϕ-ing is at a lower mechanistic level than S’s ψ-ing if and only if X and its ϕ-ing are
component parts and activities in S’s ψ-ing. A component is a part whose activities contribute to
the behavior of the mechanism as a whole. The components (Xj) are spatially contained within
S (because components are necessarily parts of the things they compose). Furthermore, each
component is relevant to S’s ψ-ing: X’s ϕ-ing should contribute to (or be operative in) S’s ψ-ing
(see Martin and Deutscher 1966).
There is some debate about how this interlevel relevance relationship should be understood.
Craver (2007) defines it in terms of the mutual manipulability of the whole and the part: X
186
Mechanistic levels, reduction, and emergence
SΨ-ing
X2φ2-ing
X1φ1-ing X4φ4-ing
X3φ3-ing
P2ρ2-ing
P1ρ1-ing P4ρ4-ing
P3ρ3-ing
Figure 14.1 A schematic of mechanistic levels
and its ϕ-ing are component parts and activities of S’s ψ-ing if one can manipulate S’s ψ-ing
by intervening on X’s ϕ-ing, and one can manipulate X’s ϕ-ing by manipulating S’s ψ-ing.
This is a sufficient, but not necessary, condition for componency, as there might be parts of
a mechanism that do not change during its operation (consider the walls surrounding pistons
in a car engine) or redundant parts that can be independently manipulated with no effect on
the behavior of the mechanism as a whole. Couch (2011) argues that componency should be
understood in terms of Mackie’s notion of an INUS condition. Components are Insufficient
and Non-redundant parts of an Unnecessary but Sufficient set of contributors to S’s ψ-ing. Still
others (such as Harinen 2014 and Romero 2015; see also Craver 2007) understand constitutive
relevance as causal betweenness: a component of a mechanism is causally intermediate between
the mechanism’s start conditions and its termination conditions (see Chapter 9 on parts and
boundaries of mechanisms).
Furthermore, Figure 14.1 is ambiguous: Two distinct relations are represented in both the
top and bottom (corresponding roughly to what Gillett (2002) calls dimensioned and flat realiza-
tion). The first is a part–whole relation between S’s ψ-ing and each of the X’s individual ϕ-ings.
We use the term “mechanistic level” exclusively for this relation. The second is a whole–whole
relation between two ways of describing the same thing. We can talk about S’s ψ-ing or we
can talk about the organized collection of ϕ-ing Xs. These are two ways of describing one and
the same object. Likewise, at the bottom of the figure, we can talk about a given X’s ϕ-ing or
we can talk about the organized collection of ρ-ing Ps in virtue of which X ϕs. Again, these are
two levels of description that apply to one and the same object. Marr’s levels of realization are
like this: the computational level, the algorithmic level, and the implementation level are three
ways of describing one and the same whole (Marr 1982; see also Chapter 29). We return to this
issue in section 6.
187
The above explication of mechanistic levels has several conceptual benefits. First, it accurately
describes the multilevel explanatory structures one finds in biology and other special sciences,
and makes explicit the kinds of evidence required to evaluate such explanations. Second, it
satisfies many of our pre-analytic intuitions about the levels of explanation. Things at higher
mechanistic levels are typically larger (and necessarily no smaller) than things at lower levels
because the latter are parts of the former. Things at lower levels often take less time than things
at higher levels because activities at lower levels compose activities at higher levels. And the idea
of mechanistic levels captures the common idea of “levels of organization” because higher-level
mechanisms are made up of organized parts and activities. Third, this understanding of levels
clarifies why interlevel causation is at least prima facie mysterious: causation between mechanistic
levels must involve parts interacting with their wholes (see section 4).
Finally, this understanding of levels helps to undermine the thought that nature can be use-
fully divided into monolithic levels of, e.g., atoms, molecules, cells, organs, organisms, and
societies (Oppenheim and Putnam 1958). For mechanistic levels, there is no unique answer
to the question of when two items are at the same mechanistic level. There is only a necessary
condition: X’s ϕ-ing and S’s ψ-ing are at the same level of mechanisms only if (i) X’s ϕ-ing and
S’s ψ-ing are components in the same mechanism, (ii) X’s ϕ-ing is not a component in S’s ψ-ing,
and (iii) S’s ψ-ing is not a component in X’s ϕ-ing. Unlike size levels or levels defined in terms
of the types of objects found at a given level, mechanistic levels are defined by the componency
relationship between things at higher and lower levels. If two things are not related as part to a
whole, they are not at different levels; if they are in the same mechanism, then they are in this
limited sense “at the same level.” But one might just as easily say on this basis that sameness of
level has no deep conceptual significance for mechanistic levels.
Note, for example, that the levels of nearly decomposable structure in stars do not cor-
respond to the levels in starfish. Each has parts and wholes, but the parts and wholes are of
different kinds. It is an empirical question, answered case by case, how many levels there are
in a system. It makes no sense to ask whether hippocampi are at a higher or lower mechanistic
level than horseshoes. They are not components in the same mechanism; they are not related
as parts to wholes. For some, this is tantamount to abandoning the idea of levels (e.g., Eronen
2015). We see our account rather as a distillate of the ordinary scientific concept, an extraction
of an explanatorily and metaphysically central idea, leaving behind as residue the problematic
commitments inherent in our inchoate, folk talk of “levels.”
3. Comparisons and contrasts

The ordinary concept of “level” is inchoate, in part, because the term is used promiscuously to
describe many distinct kinds of relata and relations. Here we distinguish mechanistic levels from
aggregates, size levels, causal levels, and Oppenheim and Putnam’s levels.
As noted above, mechanisms contrast with aggregates. In aggregates, the property of the
whole is literally a sum of the properties of its parts. The concentration of a fluid is an aggrega-
tion of particles; allelic frequency is a sum of individual alleles. Aggregate properties change
linearly with the addition and removal of parts, they don’t change when their parts are rear-
ranged, and they can be taken apart and reassembled without any special difficulty. This is
because in true aggregates, spatial, temporal, and causal organization are irrelevant (Wimsatt
1997). Mechanisms, in contrast, are literally more than the sums of their parts: they change non-
linearly with the addition and removal of parts, their behavior is disrupted if parts are switched
out, and this is because their spatial, temporal, and causal organization make a difference to how
the whole behaves.
188
Though distinct, mechanisms and aggregates are nonetheless species of a genus: each involves
a relationship between the properties or activities of wholes and the relevant properties or
activities of their parts. Many mechanisms (from steam engines to action potentials) rely on both
mechanistic and aggregative relations. As noted above, aggregation is the limit as mechanistic
organization goes to zero.
Mechanistic levels also contrast with size levels. The relata in size levels are space-involving
entities (like cells), and they are higher- or lower-level than one another because they are big-
ger or smaller, respectively. They are “at” a level when they have the same (or similar) sizes.
Mechanistic levels are also orderable by size, but the size relationship between mechanistic levels
follows from the more fundamental, compositional relationship. Like Wimsatt’s (1976) classic
image of levels as peaks of regularity and predictability, nearly decomposable mechanisms will
tend to carve most naturally at interfaces between components, thereby identifying isolable
pockets of regularity and predictability. Wimsatt describes but does not explain why these levels
of regularity and predictability cluster at different size scales. In mechanistic levels, the clustering
is explained by patterns of near decomposability in the mechanism’s causal organization.
Sometimes the term “level” is used to describe relations among the stages in a causal path-
way, as when one distinguishes “higher-level” and “lower-level” visual areas (e.g., area MT and
V1, respectively) or, arguably, when one speaks of “higher-level executive functions.” These
causal relations are clearly distinct from the compositional relation involved in mechanistic lev-
els. MT is not part of V1, but downstream in a causal process from V1; lower-level cognitive
functions are not parts of higher-level cognitive functions, but asymmetrically controlled or
dominated by higher-level cognitive functions (see Churchland and Sejnowski 1992).
The term “levels” is also associated with Oppenheim and Putnam (1958), who structure their
view of the unity of science around a monolithic conception of levels. They carve the world
into roughly six ontological strata (societies, organisms, cells, molecules, atoms, and elementary
particles). Each stratum corresponds to a distinct field of science, from economics at the top to
particle physics at the bottom. Each level has its distinctive theory. The unity of science consists
in the explanation of higher-level theories in terms of lower-level theories.
Oppenheim and Putnam could easily have embraced mechanistic levels as the ontic component
of their picture; things at higher levels are wholes made up of things at lower levels. As Wimsatt
(1976) notes, however, this minor amendment does violence to the simplicity of Oppenheim and
Putnam’s vision of scientific unity. If levels are compositional, and different mechanisms decom-
pose into different kinds of parts, then we should not expect a uniform decomposition of all things
into the same kinds of parts (compare stars to starfish and hippocampi to horseshoes).
Oppenheim and Putnam’s vision of scientific unity is descriptively inadequate, as they would
likely acknowledge. Both sciences and theories run rough-shod over levels. Models and theo-
ries often span levels of organization, linking phenomena studied by different fields of science
(Darden and Maull 1977; Craver 2007). Scientific fields are also increasingly transdisciplinary
(e.g. neuroscience, ecology). Things that are the same size can usefully be investigated by alto-
gether distinct fields of science: Cytologists, anatomists, and electrophysiologists all study cells
with different tools. The relation between fields of science, levels of theory, and ontological
levels is thus many to many to many.
The notion of mechanistic levels provides a compelling alternative to the Oppenheim
and Putnam model. Mechanisms span multiple levels of organization. Scientists approach
these structures and processes with diverse tools and from diverse theoretical vantage points.
The unity that results from this collaboration is more like a mosaic than a layer cake: it is
achieved when different scientists with different instruments and modeling tools use their
diverse expertise to understand the same mechanism (Craver 2007).
189
4. Levels, emergence, and interlevel causation

Do things at higher levels “emerge” from things at lower levels? And can things at different
levels causally interact? If we think of levels as mechanistic levels, these questions have clear
content, and it is clear what is at stake in answering them.
Mechanisms are not aggregates. A property or activity at a higher level of mechanistic
organization (e.g., S’s ψ-ing) is literally more than the sum of the properties of its parts. So, if
emergence is defined as the failure of aggregativity (call this organizational emergence; Wimsatt
1997), then things at higher mechanistic levels emerge organizationally from things at lower lev-
els. No ontological extravagance is required: two stacked toothpicks have the organizationally
emergent capacity to catapult raisins. Neither toothpick can catapult raisins alone: so the whole
has capacities the parts alone do not possess.
Often the term “emergence” is used in an epistemic sense to refer to the inability to predict the
properties or behaviors of wholes from properties and behaviors of the parts. Epistemic emer-
gence can result from our ignorance, such as failing to recognize a relevant variable, or from
failing to know how different variables interact in complex systems. It might also result from
limitations in human cognitive abilities or current-generation representational tools (Boogerd
et al. 2005; Richardson and Stephan 2007). The practical necessity of studying mechanisms by
decomposing them into component parts raises the epistemic challenge of conceptually putting
the parts back together so they work (Bechtel 2013). Epistemic emergence results from the lim-
its of our knowledge or of our representational capacities, not from discontinuity in the world’s
causal structure (see Chapter 27).
Ontic emergence is suspect or promising (depending on one’s perspective) precisely because it
involves such discontinuity: there are higher-level properties and capacities that have no suf-
ficient (ontic) explanation in terms of the parts, activities, and organizational features of the
system in the relevant conditions. Some say life, consciousness, or intentionality are emergent in
this sense. Mechanistic levels, in contrast, are levels of mechanistic dependence: they are defined
in terms of componency (Craver 2014). If that ontic relationship is severed, then the sense in
which emergent properties are at a “higher level” must be different than the sense in mecha-
nistic levels. Indeed, it is unclear why properties that ontically emerge should be thought of as
higher-level at all (rather than, e.g., effects of a cause). Ontic emergence, whatever its virtues, is
mysterious precisely because it is distinct from, and so gains no plausibility from verbal associa-
tion with, organizational or epistemic emergence. Organizational and epistemic emergence are
unmysterious both in scientific common sense and common sense proper (Van Gulick 1993;
Kim 1998). Appeal to ontic emergence, on the other hand, arouses suspicion because it is com-
mitted to the existence of phenomena that have no sufficient (ontic) mechanistic explanation.
Can things at different mechanistic levels properly be said to causally interact with one
another? Many common assumptions about causation appear to block this thought. For Hume
and Lewis, the relata in a causal relationship must be distinct,
and distinct not only in the sense of nonidentity but also in the sense of nonoverlap and
nonimplication. It won’t do to say that my speaking this sentence causes my speaking
this sentence; or that my speaking the whole of it causes my speaking the first half of
it; or that my speaking causes my speaking it loudly, or vice versa.
(Lewis 2000, p. 78)
Wholes and parts overlap; a token S’s ψ-ing as a whole includes every part of S’s ψ-ing, and that
includes all the Xj, ϕi, and organizing relations. On Salmon’s process view of causation, the relata
190
in causal relations must intersect in space-time, exchange conserved quantities, and maintain
those changes after the intersection ends. But parts and wholes are always everywhere together;
the whole has no additional conserved quantity to pass to its parts. The relationship between
LTP and the opening of NMDA receptors during LTP induction is directly analogous to the
relationship between speaking the whole of a sentence and speaking its first half. The induction
of LTP is partly constituted by the opening of the NMDA receptor. The would-be cause in this
top-down causal claim already contains the would-be effect; talk of causation across levels of
mechanism thus appears inappropriate (Craver and Bechtel 2007).
What about the bottom-up case? Here we must guard against an ambiguity. Sometimes
mechanists describe the “phenomenon” as an activity or process that starts with the mechanism’s
setup conditions and ends with its termination conditions (Machamer et al. 2000). For example,
Long-Term Potentiation (LTP) is described as a process beginning with rapid and repeated
stimulation of the presynaptic neuron and ending with enhanced synaptic transmission. Other
times, mechanists describe the phenomenon as the product of that process (as one of its termi-
nation conditions). For example, the mechanism of LTP produces a potentiated synapse. This
second phrasing leads us to seek the antecedent causes: the tetanus and the subsequent changes
in the NMDA receptor. But if we think about the phenomenon as a process, an input–output
relation starting with the tetanus and ending with a potentiated synapse, then we should not say
the NMDA receptor is a cause of that, if we wish to speak accurately about the ontic structure
of the situation at hand. The NMDA receptor is a part of that causal process; it causes neither
itself nor its antecedents in the mechanism.
5. Reduction
Like “level,” “reduction” is used many ways. Sometimes, it expresses a thesis about explanation,
about how one theory or law is explained by other theories or laws. Other times, it is a thesis
about scientific integration and unity, about the relationships among the diverse branches of sci-
ence. And still other times, it is a metaphysical thesis about the structure of the world (section 6).
The idea of mechanism offers considerable insight into explanation and scientific integration.
Although the mechanistic view of levels alone does not resolve the deepest ontological puz-
zles, it nonetheless offers a descriptively adequate image of the kind of ontological structure by
which the manifest phenomena of our world might be (and arguably are) related to its most
fundamental parts and activities.
According to the covering law (CL) model, reductive explanations involve deriving higher-
level theories from lower-level theories with the aid of bridge laws connecting their distinct
theoretical vocabularies (Nagel 1961; Schaffner 1993). This model of explanatory reduction
serves as a valuable ideal in the search of mechanisms. A description of the behavior of the
mechanism should be shown to follow from a description of the parts and their interaction, as
one understands them. For example, Hodgkin and Huxley showed that one could recover the
shape of the action potential from measured values of ionic concentrations and experimentally
determined ion-gating functions coupled with laws governing electrical circuitry. Yet this epis-
temic test of the sufficiency of an explanation should not be confused with the explanation itself.
Not all predictively adequate models are explanatory. We can, as Hempel (1965) argued,
derive Boyle’s law from the conjunction of Boyle’s law and Kepler’s laws, but not in a way
that explains it. We can predict the behavior of a part from the description of the behavior
of the whole and the behaviors of a few other parts, despite the fact that many expect such
micro-reductive explanations to work only from the bottom up. We can predict the behavior
of a system by tracking the right correlations in the system irrespective of their causal relations
191
to one another (as in diagnosis and prognosis) and would not thereby explain the system’s
behavior. Finally, the classical model of reduction is not equipped with an account of explan-
atory relevance, a way to determine which predictively relevant features are explanatorily
relevant and which are not. If one thinks of reduction not as primarily an epistemic, deductive
relationship between descriptions of the behavior of a mechanism and of the organized activi-
ties of its component parts but as a matter of learning and revealing the causal structure of a
mechanism, these problems for the classic theory of micro-reductive explanation are directly
addressed (Craver 2007).
Given the multilevel structures found in mechanistic explanations, it’s clear that scientific
integration will not conform to the layer cake pattern Oppenheim and Putnam describe across
science as a whole. Interlevel linkages will be of more local significance, relativized to particu-
lar explanandum phenomena. Different fields of science collaborate with one another both to
bridge across mechanistic levels and to bring their unique perspectives to bear on one and the
same thing. A kind of mosaic unity results from having the same mechanism as the target of
one’s scientific investigation; integration occurs when the findings from different fields and per-
spectives mutually constrain our understanding of how the mechanism works.
This mechanistic perspective provides an attractive successor to empiricist views of explana-
tion as a model of interfield integration. First, it is based on a more accurate and informative
view of levels. Second, it provides significantly more insight into the diverse evidential con-
straints by which interlevel bridges are evaluated and into the forces driving co-evolution
across different levels (Craver and Darden 2013). Constraints on the parts, their causal interac-
tions, and their spatial, temporal, and hierarchical organization all help to flesh out an interfield
theory. Finally, many mechanists argue that classical reduction models are myopically focused
on downward-looking explanations at the expense of a fuller account of the details by which
interlevel and interfield theories more generally are constructed and evaluated. This is especially
true when one must look up to higher levels to see how a part contributes to a mechanism
of which it is a component and in which two fields combine findings “at” rather than across
mechanistic levels.
6. The ontological status of higher-level phenomena

We turn now to the metaphysical questions central to many philosophical discussions of reduc-
tion: what is the relationship between higher-level phenomena and things at lower levels? We
disambiguated two questions implicit in Figure 14.1. The first is: Is the behavior of the mecha-
nism as a whole greater than the sum of the behaviors of its parts? We defend an affirmative
answer above. The second question is: Is the behavior of the mechanism as a whole greater
than the organized interaction of its parts in context (commonly called the realizer of the
mechanism’s behavior)?
Simple examples suggest that the relationship between the phenomenon and its realizer is
more intimate than correlational, causal, and nomologial relationships are. In a standard mouse-
trap, if the trigger is depressed, the catch slides, and the impact bar completes its arc (collectively,
M), then the trap necessarily fires (F). We are precluded from imagining that all the stages of
M have occurred but that the trap has not fired precisely because there is nothing more to be
done, no activity or property to add, to transform M in context, into an instance of F. We are
similarly precluded from imagining in a population of mousetraps composed of identical parts
organized and interacting identically with one another in the same conditions that only 30
percent of these mechanisms should fire and 70 percent should not. Correlational, causal, and
nomological relationships are not necessarily like this. The intimacy between a phenomenon
192
and its realizing mechanism thus appears to be stronger, more metaphysically necessary, than a
simple matter of regularity, causation, or law. The behavior of the whole contains the behaviors
of the parts, and the behaviors of the parts collectively and exhaustively constitute the behavior
of the whole. Dualism, parallelism, and emergentism do not share this commitment and allow
that there are properties and activities of wholes that are not exhaustively constituted by lower-
level mechanisms.
This metaphysical commitment functions as a constitutive ideal in the search for mechanistic
explanations (Haugeland 1998). If one knows all of the relevant entities, activities, and organi-
zational features, and knows all the relevant features of the mechanism’s context of operation,
and can in principle put it all together, then one must know how the mechanism will behave.
This was Hempel’s important insight about the epistemology of explanation. It is an epistemic
warning sign if features of the mechanism’s behavior cannot be accounted for in terms of our
understanding of its parts, activities, organization, and context. Mechanists thus operate with a
background assumption that the phenomenon is exhaustively explained (in an ontic sense) by
the organized activities of parts in context.
One is tempted to say that higher-level phenomena (in a given context) are identical to the
organized activities of the mechanism’s parts (in that context). By this one might assert a type
identity, according to which ψ-ing in general is identical to the organized ϕ-ing of X’s, or a
token identity, according to which each instance of ψ-ing is identical to some organized ϕ-ing of
X’s. Classical reductionists favor type identity, though this thesis is largely out of fashion (though
see Polger 2004). As Putnam (1967) and Fodor (1974) argued, such type-identity statements
are false when the phenomenon is or can be multiply realized in distinct mechanisms. There
are many ways to trap a mouse and many ways to produce an action potential. If so, the same
type of phenomenon might be instantiated on different occasions in different mechanisms; the
realized type is not the same as the realizer type.
Bechtel and Mundale (1999) counter that this argument rests on an illusory matching of
grains between the characterization of the phenomenon and the characterization of the mecha-
nism. If one characterizes the phenomenon abstractly, then many mechanisms can or could
give rise to it. If one characterizes the phenomenon in fine detail (including, for example, all its
excitatory, inhibitory, and modulatory conditions), then the class of actual and potential realizers
might shrink to one. If the phenomenon and the mechanism are characterized at comparable
degrees of abstraction, then the thesis of type identity begins to look more promising.
But perhaps a deeper problem for type identity is lurking in the background and not so
easily addressed: the question of whether there is, in general, an “appropriate” grain of analy-
sis that adequately captures the phenomenon and the mechanism at once. Even some token
things are multiply realizable: One and the same rook in a chess game might be instantiated
at different times by different figurines (imagine replacing a lost plastic castle midgame with
a Fig Newton; Haugeland 1998). One and the same kidney is made up of different parts over
a person’s lifetime. To answer whether ψ and the organized collection of ϕ-ers are the same
thing, we need to be able to answer the question “The same what?” with a non-dummy
sortal (Marcus 2006). The conditions under which something is the same ψ-er might be
different from the conditions under which it is the same organized collection of ϕ-ers, in
which case there are no individuation conditions that cover the thing both as a ψ-er and
as an organized collection of ϕ-ers, and there is no sense to the question of whether they
are identical. Items at different mechanistic levels are no doubt intimately related, but our
ordinary notions of token and type identity appear poorly equipped to describe that relation.
For these reasons, we prefer to speak of a particular organized ϕ-ing of X’s as exhaustively
constituting a particular ψ-ing at or over a time.
193
Instead of asking about identity, one might ask rather about whether phenomena have causal
powers in a given context that the organized collection of interacting parts do not have. As
noted above, it is relatively uncontroversial that wholes can do things their parts cannot do (see
Kim 1998, p. 85). Do realized phenomena have causal powers their realizing mechanisms do
not have in the same contexts? If one thinks of the phenomenon as identical to the behaving
mechanism, then the answer is no: the two have all and only the same powers. Non-reductive
physicalists have traditionally attempted to argue that although higher-level phenomena are all
ontically explainable in terms of more fundamental mechanisms, there remains a sense in which
the higher-level phenomena have causal powers of their own.
Again, issues of multiple realization arise. Suppose a neuron generates an action potential,
causing neurotransmitters to be released from the cell. Whenever it does, we suppose there
is a determinate set of entities (ions, channels, etc.) organized and interacting in determinate
ways that fully constitutes it at that time. Now suppose we ask: which feature of the world is
relevant to the release of neurotransmitters: the action potential or the determinate organiza-
tion and interaction of ions and channels? One standard test of causal relevance is to ask what
would happen if the cause had been different. Experimental investigation will reveal that the
action potential makes a substantial difference to the probability of neurotransmitter release by
raising membrane voltage. The precise arrangement of ions and ion channels makes no differ-
ence except insofar as it affects membrane voltage. Many differences among the realizers make
no difference to the effect in question. If causal relevance is a matter of making a difference
(Woodward 2003), the fine details about the realizing mechanism are not causally relevant. Each
action potential is exhaustively constituted by a mechanism, but the causally relevant difference
is the occurrence of the action potential, not the precise mechanism that produced it (see Craver
2007, chapter 6).
Just how this mechanistic perspective best fits with metaphysical views about, e.g., identity
and causal powers remains to be decided. A determinate view about multiple realization and its
implications requires metaphysical commitments that go beyond the skeletal framework of lev-
els presented here. Views on multiple realization differ depending, inter alia, on how one thinks
about properties (Heil 1999), the realization relation (Shapiro 2004), and theories (Klein 2013).
Similarly, the prospects for token physicalism depend, inter alia, on one’s account of events
(Kim 2012) and identity (Marcus 2006). One goal of future work is to integrate the mechanistic
perspective with a metaphysics that produces a coherent and plausible view of the structure of
the mechanistic world.
7. The lowest level

Questions also arise concerning the bottom of a hierarchy of mechanisms. Two options seem to
be available: either (1) the hierarchy has a bottom or (2) it does not, i.e. it continues ad infini-
tum. Though both options are metaphysically live, some facts about quantum mechanics seem
to favor the first.
It has been argued, for example, that the non-separability of ontological units in the
theories of quantum mechanics entails that there are no distinct objects to be identified as
components beneath a certain level of organization. Kuhlmann and Glennan (2014) argue
that for any local, classical phenomenon, there is a lowest classical level at which mechanistic
explanation bottoms out. Precisely where it bottoms out varies across phenomena because
decoherence—the physical process by which the quantum realm becomes classical—is itself
local. Below this level, where properties like spatial location are indeterminate, prospects for
194
mechanistic explanation are murky. Mechanistic explanation may still be possible if quantum
phenomena can be explained by causal organization among non-local “objects” (for exam-
ple, see the discussion of laser light (Kuhlmann and Glennan 2014, p. 354)), but in quantum
holistic cases where the properties of subsystems don’t appear to determine the properties of
the whole system (for example, EPR experiments), the constitutive ideal may be unreachable.
8. Conclusion
Our goal has been to sketch the mechanistic approach to levels, to contrast it with near neigh-
bors, and to explore some of its metaphysical implications. This perspective provides some
important content missing from standard metaphysical discussions of levels. Most notably, it
offers a clear sense of what it means for things to be at different levels. It also allows us to dis-
tinguish mechanistic levels from realization relations. And finally, it provides a clearer vision of
the structure of multilevel explanations, the evidence by which such explanations are evaluated,
and the form of scientific unity that results from building them. The mechanistic approach does
not, by itself, solve all the metaphysical questions one might have about physicalism. Yet it does
provide a skeletal framework for thinking about how the macroscopic phenomena of our world
are or might be related to its most fundamental parts and activities. Simon’s parable and its many
contemporary analogs give us reason to think that this skeletal picture captures a central and
important pattern in the causal structure of our world.
Note
1 Thanks to Mike Dacey, James Gulledge, Ben Henke, Eric Hochstein, Emily Prychitko, and Felipe
Romero for invaluable comments, and to Pamela Speh for the design of Figure 14.1.
References
Albert, R., H. Jeong, and A.-L. Barabási. (2000) “Error and Attack Tolerance of Complex Networks,”
Nature, 406, 378–382.
Bechtel, W. (2013) “Addressing the Vitalist’s Challenge to Mechanistic Science: Dynamic Mechanistic
Explanation,” in S. Normandin and C. T. Wolfe (eds), Vitalism and the Scientific Image in Post-
Enlightenment Life Science 1800–2010, Dordrecht: Springer, 345–370.
Bechtel, W. and J. Mundale. (1999) “Multiple Realizability Revisited: Linking Cognitive and Neural
States,” Philosophy of Science, 66, 175–207.
Bechtel, W. and R.S. Stufflebeam. (2001) “Epistemic Issues in Procuring Evidence about the Brain: The
Importance of Research Instruments and Techniques,” in W. Bechtel, P. Mandik, J. Mundale and
R.S. Stufflebeam (eds), Philosophy and the Neurosciences: A Reader, Blackwell: Malden, MA, 55–81.
Boogerd, F.C., F.J. Bruggeman, R.C. Richardson, A. Stephan and H.V. Westerhoff. (2005) “Emergence
and Its Place in Nature: A Case Study of Biochemical Networks,” Synthese, 145, 131–164.
Churchland, P.S. and T. Sejnowski. (1992) The Computational Brain, Cambridge, MA: MIT Press.
Couch, M. (2011) “Mechanisms and Constitutive Relevance,” Synthese, 183, 375–388.
Craver, C.F. (2001) “Role Functions, Mechanisms, and Hierarchy,” Philosophy of Science, 68, 53–74.
Craver, C.F. (2007) Explaining the Brain, Oxford: Oxford University Press.
Craver, C.F. (2014) “The Ontic Account of Scientific Explanation,” in M.I. Kaiser, O.R. Scholz,
D. Plenge, and A. Hüttemann (eds), Explanation in the Special Sciences: The Case of Biology and History,
Dordrecht: Springer, 27–52.
Craver, C.F. and W. Bechtel. (2007) “Top-down Causation without Top-down Causes,” Biology and
Philosophy, 22, 547–563.
Craver, C.F. and L. Darden. (2013) In Search of Mechanisms: Discoveries across the Life Sciences, Chicago:
195
Cummins, R. (1975) “Functional Analysis,” Journal of Philosophy, 72, 741–765.

Darden, L. and N. Maull. (1977) “Interfield Theories,” Philosophy of Science, 44, 43–64.
Eronen, M.I. (2015) “Levels of Organization: A Deflationary Account,” Biology and Philosophy, 30, 39–58.
Fodor, J. (1968) Psychological Explanation, New York: Random House.
Fodor, J. (1974) “Special Sciences: Or the Disunity of Science as a Working Hypothesis,” Synthese, 28,
97–115.
Gillett, C. (2002) “The Dimensions of Realization: A Critique of the Standard View,” Analysis, 62,
316–323.
Glennan, S. (2002) “Rethinking Mechanistic Explanation,” Proceedings of the Philosophy of Science Association,
3, S342–S353.
Harinen, T. (2014) “Mutual Manipulability and Causal Inbetweenness,” Synthese, 1–20.
Haugeland, J. (1998) Having Thought: Essays in the Metaphysics of Mind, Cambridge, MA: Harvard University
Press.
Heil, J. (1999) “Multiple Realizability,” American Philosophical Quarterly, 36, 189–208.
Hempel, C. (1965) Aspects of Scientific Explanation, New York, NY: Free Press.
Kauffman, S.A. (1970) “Articulation of Parts Explanation in Biology and the Rational Search for Them,”
PSA: Proceedings of the Biennial Meeting of the Philosophy of Science Association, 1970, 257–272.
Kim, J. (1998) Mind in a Physical World: An Essay on the Mind-Body Problem and Mental Causation, Cambridge,
MA: MIT Press.
Kim, J. (2012) “The Very Idea of Token Physicalism,” in C. Hill and S. Gozzano (eds), New Perspectives on
Type Identity: The Mental and the Physical, Cambridge: Cambridge University Press, 167–185.
Klein, C. (2013) “Multiple Realizability and the Semantic View of Theories,” Philosophical Studies, 163,
683–695.
Kuhlmann, M. and S. Glennan. (2014) “On the Relation Between Quantum Mechanical and Neo-
Mechanistic Ontologies and Explanatory Strategies,” European Journal for Philosophy of Science, 4,
337–359.
Levy, A. and W. Bechtel. (2013) “Abstraction and the Organization of Mechanisms,” Philosophy of Science,
80, 241–261.
Lewis, D. (2004) “Causation as Influence,” reprinted in J. Collins, N. Hall and L.A. Paul (eds), Causation
and Counterfactuals, Bradford: MIT Press, 75–106.
International Studies in the Philosophy of Science, 18, 27–39.
Machamer, P., L. Darden and C.F. Craver. (2000) “Thinking about Mechanisms,” Philosophy of Science,
67, 1–25.
Marcus, E. (2006) “Events, Sortals, and the Mind-Body Problem,” Synthese, 150, 99–129.
Marr, D. (1982) Vision, San Francisco: Freeman Press.
Martin, C.B. and M. Deutscher. (1966) “Remembering,” Philosophical Review, 75, 161–196.
Menzies, P. (2012) “The Causal Structure of Mechanisms,” Studies in History and Philosophy of Science Part
C: Studies in History and Philosophy of Biological and Biomedical Sciences, 43, 796–805.
Nagel, E. (1961) The Structure of Science: Problems in the Logic of Scientific Explanation, New York: Harcourt,
Brace and World, Inc.
Oppenheim, P. and H. Putnam. (1958) “Unity of Science as a Working Hypothesis,” in H. Feigl,
M. Scriven, and G. Maxwell (eds), Concepts, Theories, and the Mind-Body Problem, Minnesota Studies in
the Philosophy of Science II, Minneapolis: University of Minnesota Press, 3–36.
Polger, T. (2004) Natural Minds, Cambridge, MA: MIT Press.
Putnam, H. (1967) “Psychological Predicates,” in W.H. Capitan and D.D. Merrill (eds), Art, Mind, and
Religion, Pittsburgh: University of Pittsburgh Press, 37–48.
Richardson, R.C. and A. Stephan. (2007) “Emergence,” Biological Theory, 2, 91–96.
Romero, F. (2015) “Why There Isn’t Inter-level Causation in Mechanisms,” Synthese, 192, 3731–3755.
University Press.
Schaffner, K. (1993) Discovery and Explanation in Biology and Medicine, Chicago: University of Chicago
Press.
Schlosser, G. and G. Wagner (eds) (2004) Modularity in Development and Evolution, Chicago: University of
Chicago Press.
Shapiro, L. (2004) The Mind Incarnate, Oxford: Oxford University Press.
196
Simon, H. (1969) The Sciences of the Artificial, Cambridge, MA: MIT Press.
Van Gulick, R. (1993) “Who’s in Charge Here? And Who’s Doing All the Work?” in J. Heil and
A.R. Mele (eds), Mental Causation, Oxford: Oxford University Press, 233–256.
Wimsatt, W.C. (1976) “Reductionism, Levels of Organization, and the Mind-Body Problem,” in
G. Globus, G. Maxwell and I. Savodnik (eds), Consciousness and the Brain: A Scientific and Philosophical
Inquiry, New York: Plenum Press, 202–267.
Wimsatt, W.C. (1997) “Aggregativity: Reductive Heuristics for Finding Emergence,” Philosophy of Science,
64, 372–384.
Woodward, J. (2003) Making Things Happen, Oxford: Oxford University Press.
197
15
MECHANISMS AND
NATURAL KINDS1
Emma Tobin
Scientists divide the particulars that they study into kinds and classes. Chemists classify particular
pieces of gold, as members of the class gold. We cluster things together and find this useful for
the purposes of prediction and explanation. For example, we think we can study one piece of
gold and learn things about other pieces of gold. Debates about classification in philosophy have
traditionally focused on the delineation of natural kinds of substances. More particularly, we
ask whether these natural kinds reflect real natural divisions in nature or whether alternatively
they merely reflect the theoretical interests of the scientists that use them. The discussion has
more recently invoked a more generous notion of natural kind, going broader than substances
to consider whether we can have natural kinds of processes, relations, and/or property clusters
(Ellis, 2001).
As we will see, some views of natural kinds (Boyd, 1991) have even invoked the notion
of a mechanism to provide a robust realist account of natural kinds. In the meantime, the
mechanisms literature has occasionally discussed the reality of mechanisms, but has seldom
addressed the question of whether mechanisms actually form natural kinds. A closer analysis of
the relationship between natural kinds and mechanisms is therefore required. The analysis of
this chapter reframes the question as to what a more liberal account of classification might look
like, one which includes mechanisms. Moreover, I will show that thinking seriously about
how we classify mechanisms in scientific practice significantly alters our understanding of what
classification is supposed to do.
Boyd’s (1991) homeostatic property cluster (HPC) account is one much-discussed view of
natural kinds, and it invokes the notion of a mechanism as a factor sustaining the clusters of
properties that are considered to be natural kinds. On Boyd’s view, natural kinds are reliable for
the purposes of explanation and inductive inference, precisely because clusters of properties are
reliably sustained by homeostatic mechanisms. However, an ontological account of mechanisms
is not sufficiently developed by those holding this view, nor is the relationship between kinds
and mechanisms precisely cashed out. Kinds, we are told, are understood as property clusters
sustained by homeostatic mechanisms. Nevertheless, the relationship between the two, namely
what counts as a mechanism and how homeostasis is achieved, is not sufficiently developed.
To develop this view, a closer analysis of how mechanisms are involved in classification is
required, which directly addresses the relationship between kinds and mechanisms. An addi-
tional question is raised from considering this relationship, which is how we might classify
198
Mechanisms and natural kinds
mechanisms themselves. Can mechanisms be sorted into more and less similar mechanisms, also
allowing us to study one token mechanism and think we can learn something about similar
mechanisms? If so, perhaps we need to broaden further our notion of natural kinds to accom-
modate mechanisms as well as substances, processes, relations, and/or property clusters.
One ambition of addressing these questions would be to provide a more robust realist HPC
view of natural kinds, which precisely specifies the role of mechanisms in maintaining natural
property clusters, and my primary aim in this chapter will be to examine whether this is feasible.
Craver (2009) is skeptical because he claims that conventional elements are involved in deciding
when two mechanisms are of the same kind and also in identifying when one mechanism ends
and the other begins. Kind classifications are affected by facts about how humans find it useful
to cluster properties. This is a key conventionalist challenge to the HPC account, which under-
mines a realist interpretation of mechanisms and their natural property clusters. My secondary
aim in this chapter will be to examine how thinking seriously about classifying mechanisms can
illuminate both the mechanisms literature and the classification literature.
I am going to do two things. In section 1, I will bring together the classification literature
with the mechanisms literature in an attempt to show how the two debates impact on one
another. In section 2, I will address two pragmatic challenges facing mechanism classification. In
the mechanisms literature there has been a worry about drawing the boundaries of mechanisms.
This boundary problem, as I will call it, is about token mechanisms. Theorists are concerned
about how we can know where one mechanism begins and another ends. Related to this is the
concern about how we delimit token mechanisms from the environment. In contrast to the
mechanisms literature, boundary problems in the kinds literature have been about how we sort
things into kinds, or, in other words, how we sort tokens into types. Given that the HPC view
is an attempt at combining both mechanisms and natural kinds, a positive view will need to get
clear on these various kinds of boundary issues. I examine these issues looking at two case studies
from scientific practice in biochemistry and chemistry, showing how there are ways in which
classifying mechanisms runs into familiar problems from the classification debate. In particular,
I argue that there is an overlapping problem for mechanisms similar to the problem that has
been discussed in the natural kinds literature. I will also show that the dynamic and complex
picture of kinds and mechanisms that emerges from practice transforms how we should think
about classification.
In section 3, I return to views of classification, and explore an apparent circularity that infects
the HPC view. The difficulty is that mechanisms were introduced to help with the delineation
of the property clusters on the HPC account of natural kinds, but addressing the issues that
emerge when identifying mechanisms in scientific practice seems to require kinds of property
clusters. This is because the property clusters will be the phenomena that the mechanisms are
invoked to explain, and the phenomena partially individuate the mechanisms. In the end, I
argue that the HPC view must answer both of the boundary problems as well as the issues of
circularity if it is to succeed as an account of classification informed by scientific practice.
1. Philosophical approaches to natural kinds

The two traditional theoretical accounts of natural kinds are essentialism and conventional-
ism. Essentialists hold that there are essential properties that all and only members of a natural
kind share, which makes them categorically distinct (Ellis, 2001; Lowe, 2006). These proper-
ties provide necessary and sufficient conditions for kind membership. For example, chemical
elements are solely individuated by their atomic number, namely the number of protons in
their atomic nuclei. Silver is the element with 47 protons in its atomic nucleus. A substance
199
Emma Tobin
with an additional proton, 48 in total, will be the chemical element cadmium and not silver.
The boundaries between chemical elements are such that an atomic number is a necessary and
sufficient condition for kind membership.
Conventionalists hold that there is no objective distinction between natural kinds and any
other cluster of properties (Hacking, 1991a, 1999). Instead, the kinds that are deemed natural
are those that have theoretical and practical salience for the scientific practitioner. For example,
LaPorte (2004) uses the following example to make clear how the practical commitments of a
group of scientific practitioners can impact the decisions about natural kind membership. Jade is
microstructurally composed of two distinct chemical structures; jadeite and nephrite. For many
centuries the jade used in China was just nephrite. Toward the end of the eighteenth century
jadeite from Burma was introduced to China. Practitioners knew that it was different micro-
structurally, but nevertheless, decided to regard it as jade.2 A conventionalist would argue that
in this case contingent and conventional elements come into play in deciding what counts as a
member of the natural kind jade.
In Bird and Tobin (2008), we distinguish two forms of realism about natural kinds, which
sit on the spectrum between essentialism and conventionalism. Strong realism about natural kinds
is the view that there exist real entities that correspond to the natural kinds. Alternatively, weak
realism is just naturalism about natural kind classes. In terms of our earlier example, the difference
between silver and cadmium is not just a difference between two natural classes of thing, but is
a difference between two distinct and real entities on this view.
Boyd’s homeostatic property cluster account of natural kinds is often seen as a way between
the traditional essentialist and conventionalist account of natural kinds. Boyd originally pro-
posed this view of natural kinds as a way of articulating the kinds of the special sciences (e.g.
biology, psychology, the social sciences, etc.). In particular, Boyd’s view was introduced
to try to refute the view in the philosophy of biology that species should be understood as
individuals rather than as natural kinds (Ghiselin, 1974, 1987, 1997; Hull, 1976, 1978). He
acknowledged that essentialism could not accommodate special science kinds, and wanted to
formulate a weaker version of realism that would show how the kinds of the special sciences
are objective and sufficiently robust to support inductive inference and explanation, without
involving a commitment to essentialism. The kinds in the special sciences appear difficult to
accommodate on traditional accounts because they are changing and dynamic. For example,
take a particular species of “rabbit,” like the Cashmere Lop found in Britain. This is not static,
but has come into being at a particular time and place and continues to evolve. So an account
of natural kinds that is static and uncompromising does not appear fit for purpose for all kinds.
According to Boyd’s homeostatic property cluster view of natural kinds, natural kinds are
clusters of co-occurring properties. Boyd’s account is an extension of the traditional account
of natural kinds. He claims that this extension is appropriate just to the extent that the kinds in
question are employed for induction and explanation. Reference to a particular natural kind is
appropriate for certain sorts of inductive generalization, because it is explained by property correla-
tions that obtain independently of the understanding. These property correlations are sustained
by homeostatic mechanisms. Studying gold allows us to learn something about other samples
of gold, and studying a particular species of rabbit or a population of rabbits tells us something
about other rabbits or populations; this is why these are candidate natural kinds. Boyd’s view
then might be regarded as a strong form of realism, rather than a merely naturalist view, because
there are real entities, the property clusters that correspond to the natural kinds. These clusters
are complex and dynamic in nature, which is why it is so difficult to draw their boundaries.
The key question that arises is: why do these property correlations obtain? Boyd’s own
answer is that clusters of properties are found together in virtue of homeostatic mechanisms.
200
This view portrays a dualist ontic picture including (a) real property clusters and (b) real
homeostatic mechanisms. The HPC view argues that natural kinds are real property clusters
in the world because they are sustained by underlying mechanisms. For Boyd, a homeostatic
mechanism is defined loosely as a mechanism, which maintains a cluster of properties. For
example, in the case of a population (or species) of rabbits (e.g. the Cashmere Lop), we can
explain the clustering together of properties of members of the species by the homeostatic
mechanisms in virtue of which members all form the same population or species, such as
common ecological niche or gene exchange within this population.
Or take another example at the sub-organismic level: consider the heart as the means by
which the circulatory system pumps the blood around the body. In Boyd’s terms, the heart is a
clustering together of properties, which is maintained by the homeostatic mechanisms underlying
it. The property of having a regular heart rhythm is maintained by the heart’s natural pacemaker
in the sinoatrial node. The pacemaker sends out regular electrical impulses from the top chamber,
causing it to contract and pump blood into the lower ventricle. The pacemaking mechanism
allows the heart to contract in a continuous fashion, which allows it to pump blood to the body
and the lungs. These might be considered to be the underlying homeostatic mechanism which
allows the heart to be considered as a stable grouping of properties, namely as a HPC kind.
The attractiveness of this view is that it moves away from the traditional static view of clas-
sification and allows for change within the clustering of properties over time. It is therefore
consistent with the view that kinds like those in biology might evolve over time. The HPC
view allows a more dynamic version of classification, which sidesteps some of the issues dis-
cussed in philosophy of biology surrounding the so-called species problem.
The ontological addition of mechanisms might appear promising to those interested in pro-
viding a realist account of natural kinds, in so far as the mechanistic structure of the world could
serve as an objective foundation for an adequate taxonomy of kinds. In other words, kind con-
cepts cut nature at its joints, and to find nature’s joints, we must analyze how the homeostatic
mechanisms can set the boundaries of the property clusters. In cases where property clusters are
dynamic and may even change over time, the aspiration is that the mechanisms can help us to
sort token property clusters into types. This is not just of theoretical interest, but also provides
a practical rationale for how scientists might approach taxonomy in practice, in that it provides
hope of knowing how to go about identifying kinds in practice. There are also additional ben-
efits for classification, primarily because this view is less stringent than its essentialist alternative.
It allows for natural kinds to change some of their properties over time within any given cluster,
while remaining relatively stable over time. This is a very attractive view for taxonomists, par-
ticularly in disciplines like biology, where natural kinds evolve over time: it seems promising for
rabbits. In fact, it is the inability of the old essentialist view to straightforwardly accommodate
these changes, which has been the chief objection to it.
The HPC view, then, presents us with property clusters sustained—particularly sustained
together—by underlying mechanisms. Parallel to the classification literature, a significant litera-
ture about the nature of mechanisms has developed (Bechtel and Abrahamsen, 2005; Craver,
2007; Glennan, 1996, Machamer, Darden, & Craver, 2000). Glennan and Illari argue that these
new mechanists have a set of core commitments about what mechanisms are that are embodied
in a characterization of mechanisms they call minimal mechanism:
(Chapter 1 of this volume; see also Illari and Williamson,
2012; Glennan, forthcoming)
201
Emma Tobin
Here we have a mechanism responsible for a phenomenon, which aligns in many ways with the
idea of an underlying mechanism sustaining a property cluster.
In Chapter 7 of this volume, Glennan and Illari suggest that a scheme for classifying things
begins with a specification of the set of things to be classified. In the case of mechanism, the
things to be classified are captured by minimal mechanism, which they use as the beginning of
a classification of varieties of mechanisms. Minimal mechanism is an intentionally broad and
permissive characterization of mechanism, which does not necessarily reject the more restrictive
sense of mechanisms in earlier accounts, but rather treats these as types of mechanisms construed
more broadly. Taking this broader version of mechanism would seem to offer the best path for
providing a view of the complex and dynamic kinds that have proved difficult to characterize in
realist accounts of natural kind classification.
Craver (2009) is the first to directly bring together the recent mechanisms literature and the
HPC approach to kinds, and I will examine and develop his view in this through the rest of
this chapter. According to the HPC view, the taxonomies of natural kinds are adequate for the
purposes of explanation and induction only because the kinds in that taxonomy are sustained
by mechanisms. This view provides a justification for realism about natural kinds, because the
properties cluster together in virtue of these mechanisms. Natural kind concepts cut nature at its
joints and Craver suggests this fact can provide a normative constraint, if mechanisms have real
boundaries and so sustain real boundaries between the property clusters they sustain. In Craver’s
own term’s, “nature’s joints are located at the boundaries of mechanisms” (2009: 575).
An interesting question then arises of what exactly is meant by the boundaries of mecha-
nisms. In the mechanisms literature, the question of the boundaries of mechanisms is a pragmatic
challenge about how a practitioner would know which entities, activities, and organizational
features are part of a mechanism (or kind of mechanism) and which are not (see Chapter 9).
Now interestingly, the boundary question in the kinds literature (as we saw earlier) is a ques-
tion about how to sort token kind members into types. These are two separate questions.
What makes the HPC view so interesting is that it attempts to use mechanisms to solve the
kinds boundary problem; namely, the reason why we can sort property clusters into types is
because they are sustained by homeostatic mechanisms. However, this will require additionally
an answer to the first mechanisms boundary question. So, to say which homeostatic mechanisms
are responsible for property clusters, we will need to be able to clearly draw the boundaries of
token mechanisms, and perhaps also sort those mechanisms into kinds of mechanisms. This
serves to make clear that there is a serious challenge in conflating the boundary issues, and a
better understanding of the relationship between property clusters and mechanisms will require
a much deeper analysis than has been provided in the literature thus far.
The addition of mechanisms successfully provides an important pragmatic strategy for allow-
ing the practitioner to divide the world into kinds. This strategy, elaborated by Craver (2009), is
called the lump and split methodology: if you find that a single cluster of properties is sustained
by more than one kind of mechanism, split the cluster into subset clusters, each of which is
explained by a single kind of mechanism. Alternatively, if you find that two or more distinct
kinds are explained by the same kind of mechanism, lump the kinds into one. For example, take
the traditional yet controversial example of species delineation in biological taxonomy. In the
case of a speciation event where reproductive isolation occurs, reproductive isolation changes
the mechanism for the subset of the species population that becomes isolated. In this case, the
practitioner is forced to split the species into two separate species given the operation of this
new mechanism.
Despite the initial attractiveness of the view, Craver (2009) raises two concerns, which sug-
gest a huge impact of conventional elements. In the end, he argues that conventional elements
202
come into play in deciding (a) when two mechanisms are mechanisms of the same type, and (b)
where one particular mechanism ends and another begins. In the next section, I will look at how
the issues of overlapping have been dealt with in the classification literature so far. Additionally,
I will address how issues of overlap impact some real cases from scientific practice. An analysis
of these cases will prove helpful in looking at the relationship between mechanisms and property
clusters in practice.
2. Overlapping: mechanism and kind classification

In the classification literature, there has been much discussion of the no-overlap principle,
which is intended to allow a categorical distinction between natural kinds from a realist per-
spective. As discussed earlier, chemical elements have clear boundaries and cannot overlap. The
atomic number of an element makes it categorically distinct from the elements that come either
side of it on the periodic table. The hierarchy thesis for natural kinds is one way for natural kinds
realists to accommodate apparent cases of the overlapping of kind members. Theorists claim that
natural kinds form a hierarchy; if any two kinds overlap, then one must be subsumed under the
other like a species under a genus (Kuhn, 2000: 228–52, Ellis, 2001: 67–76, 97–100). In Tobin
(2010a), I discussed the following case, which would appear to support both the categorical
distinctness of natural kinds and the hierarchy thesis. The elements magnesium (Mg) and pro-
methium (Pm) are classified together as metals. The elements lanthanum (La) and promethium
(Pm) are classified together as lanthanides. Therefore, in accordance with the hierarchy thesis,
magnesium and lanthanum must also be classified together. This is the case because all lantha-
nides are classified as metals.
However, ipso facto cases of crosscutting natural kinds in scientific practice provide a serious
challenge to the no-overlap principle for classification (Khalidi, 1998; Dupré, 1993; Tobin,
2010a). For example, in Tobin (2010a), I examine how crosscutting occurs in the case of pro-
teins and enzymes, which appears to contradict both categorical distinctness and the hierarchy
thesis. For example, albumin and renin can be classified together as proteins. Renin and the
hairpin ribozyme can be classified together as enzymes. But the hairpin ribozyme and albumin
cannot be classified together at all, because albumin is not an enzyme. So, enzymes and proteins
are not categorically distinct. As we have seen with the problem of boundaries in the last sec-
tion, the problem of boundaries recurs with mechanisms, and is essentially the issue pointed out
by Craver when he claims that there is a difficulty in deciding where one particular mechanism
begins and another ends. As stated earlier, there are two different issues in that the boundary
problem for natural kinds would appear to be one of how we group tokens into types, whereas
the boundary problem for mechanisms appears to be one about how we draw the boundaries
between the token mechanisms themselves. Given that the HPC view involves both kinds and
mechanisms, the HPC view sits awkwardly at the juncture of these two boundary problems.
Lumping and splitting provides a practical methodology for coping with the boundary issue
for mechanisms. But it is worth pointing out that the strategy of lumping and splitting also
adheres to the same kind of no-overlap principle as we saw in the natural kinds literature;
namely, mechanisms that overlap should be lumped together, and those that do not overlap
should be split. For example, in cognitive science, spatial selective attention and spatial work-
ing memory have largely been studied in isolation and so traditionally they have been split into
two separate mechanisms, because they were thought not to overlap (Jonides and Awh, 2001).
The HPC view introduces mechanisms to provide an ontological entity, which allows us
to draw the boundaries between property clusters. However, it would seem that this kicks
the pebble down the path somewhat, since the boundary problem for kinds is replaced by an
203
Emma Tobin
additional boundary problem for mechanisms. Moreover, even if the boundary problem for
token mechanisms can be answered or at least dealt with pragmatically, then an additional
sorting problem of how we group token mechanisms into types also arises. The mechanisms
literature has rather loosely conflated these issues and separating them out will allow for a much
more nuanced discussion of the relationship between mechanisms and natural kinds.
Despite these theoretical worries, the lump and split strategy is one used and in fact introduced
by scientific practitioners themselves. The strategy was originally described by the geneticist
Victor McKusick in 1969. Additionally, it is worth noting that a number of arguments in favor
of the lump and split strategy have been provided in the mechanisms literature. The lump and
split strategy is successful to the extent that mechanisms are categorically distinct, they develop
independently of one another, they have evolved separately, and they can be manipulated or
suffer loss of function independently of one another (Craver, 2009: 581).
However, on closer analysis it looks as though scientific practice also suggests a problem
when mechanisms are added. Putting it in the terms of the classification debate, clearly mecha-
nisms overlap, but in some cases they are not lumped together or subsumed one under the
other, and in other cases they are not split. For example, take the case above from cognitive
science. A more recent study has discovered that actually the mechanisms for spatial attention
and spatial working memory interact and overlap in performing early sensory processing. In fact,
spatial memory recruits top-down processes from spatial attention and these modulate the earli-
est stages of visual analysis. Despite this clear overlap, these mechanisms continue to be treated
as distinct. Craver’s argument essentially is that sometimes the decision to lump or split involves
conventional and contextual elements.
The problem with the strategy of lumping and splitting is that there is an assumption that
once a mechanism is found to be responsible for a property cluster, this alone is sufficient for
delineating the boundaries of that cluster. However, in fact it is more likely that there are a
number of mechanisms that are responsible for a given clustering of properties. The boundary
problem in the mechanisms debate has been concerned with drawing the boundaries between
these related mechanisms. In other words, it might be possible in some cases to lump or split
the same property cluster differently depending on which mechanism the practitioner uses or
is focusing on. This is made worse by the fact that, in scientific practice, it is in fact normal
that multiple mechanisms might sustain the same property cluster. For example, Craver notes
that there might be multiple etiological pathways, which could result in the same clustering of
properties. This is certainly the case for psychiatric disorders, which often have multiple distinct
etiological mechanisms. For example, prenatal smoking exposure, low birth weight and expo-
sure to high amounts of lead are distinct risk factors for attention-deficit/hyperactivity disorder
(ADHD). In other words, there might be multiple causal routes that could result in a similar
functional output. Moreover, there might be different kinds of mechanisms, which produce the
same property cluster, some etiological and some constitutive, and depending on which one we
are using, the outcome of whether to lump or split may be different.
The problem of overlapping will be made clearer by considering some cases from scientific
practice. By considering two cases in detail, I hope to be able to show how the different bound-
ary problems play out, but also give some clarity on whether these problems are at the level of
tokens or at the level of types. Protein synthesis is a much-discussed case study in the mecha-
nisms literature, since the complexity of protein folding is illustrative of mechanistic behavior. I
would like to focus here on the case of protein classification more particularly. The protein data
bank is the primary repository of protein structures. Scientists use experimental techniques such
as X-ray crystallography, NMR spectroscopy and cryo-electron microscopy to determine the
location of each atom relative to the others in the molecule. Then, this information is deposited
204
in the protein data bank. However, the sheer volume and complexity of this data mean that the
scientific community has had to design hundreds of secondary databases that categorize the data
according to different criteria, to serve their different purposes.
Before using any one of these different databases to classify (e.g. SCOP, CATH, or
DDD), scientists have to split the proteins into domains. Domains are distinct globular parts,
which are considered to function independently. These globular parts are tertiary structures
which are the result of a simple mechanism where the molecule’s apolar amino acids are
bound to the molecule’s interior and the polar amino acids are bound outwards, allowing
dipole-to-dipole interactions. In effect, they are the simple mechanisms of protein folding.
Once a newly solved structure is discovered, the information is placed, with its unique PDB
identifier, in the protein data bank.
In incremental classification, given a newly solved structure, the new protein structure
is partitioned into domains, and each new domain is compared to domains of the exist-
ing classification to identify the most similar folds. If no such fold exists (dependent on the
parameters of the classification), then a new fold is added to the classification and the particu-
lar domain is added to it. In other words, if the domain cannot be lumped then it is split. In
practice, at least superficially, this looks like a case of support for the lumping and splitting
strategy, in so far as there appears to be a clear methodology for when we should lump and
split protein domains.
However, this makes the case look more straightforward than it is. In practice, scientists
use many different algorithms for domain identification. Splitting proteins into domains makes
the boundary problem for mechanisms quite clear. Domains are not distinct parts of proteins,
which can be clearly delineated and are categorically distinct; there are different ways to draw
the boundaries, because of the different mechanisms of protein folding.
In other words, the concept of an independent functioning unit is not enough to allow us
to delineate the boundaries of protein domains absolutely. Scientists end up having to design
algorithms to interrogate the primary data so as to artificially agree upon a boundary for the
protein domains. This is made more problematic by the fact that the algorithms used do not
always agree on what the boundaries are and often “domain choice,” as scientists call it, is
relatively subjective.
So, before the practitioner can even make the decision to lump and split, the scientific
practitioner must make a domain choice. The domain is an artificial freeze frame of the part
of a protein that is sustained by the underlying mechanisms of protein folding. Depending on
the algorithm or category used (e.g. family, superfamily) plus which database is used (there are
over 100), then it looks as though quite a few conventional elements come into the picture in
practice. This case makes clear the complexity of the boundaries of the protein domains and the
practical tools scientists need to use before the work of classification can even begin.
There is a further problem that we might consider when drawing the boundaries for token
mechanisms found in different contextual settings. In Tobin (2010b), I discussed the case of
moonlighting proteins. Moonlighting proteins are proteins that have different functional roles
in different parts of the organism. Another example is crystallins, which have a structural func-
tion in the lens of the eye, but also aid catalytic activity in cells. Crystallins are examples of
what scientists call intrinsically unstructured proteins, which means that the mechanism for
protein folding differs slightly depending on the context. One and the same primary structure
can assemble different kinds of macromolecules (quaternary structures) with different functional
roles in different parts of the organism. This is because of the structural malleability of the region
involved in binding, which allows for different protein folds because of the interactions with
other biochemicals found in different environments.
205
Emma Tobin
The question raised by this case is whether the two-tokens mechanism can be grouped
under the same mechanism type—namely mechanisms for crystallins—when they are found
to be distinct in different parts of the organism. However, it would appear that the decision
to lump together the two mechanisms under the same type—as mechanisms for crystallins—is
taken because they are considered to be two tokens of the same type, despite their different
functional roles in different contexts.
To sum up, drawing the boundaries between mechanisms for proteins is extremely difficult.
The first issue is the difficulty of creating a type of protein domain (which is an idealized model
of the dynamic part of a protein sustained by the background mechanisms of protein folding).
The second issue is that protein folding itself can differ in different contexts, to the extent that
we sometimes have to split proteins because of their great functional difference in different con-
texts. Importantly, the biochemical milieu is intimately connected to the mechanisms of protein
folding, so there is a real issue of drawing the boundaries between the mechanisms and their
environment in these complex biochemical cases.
A similar and interesting case is the use of computational methods in the classification of
chemical reactions. Ratcliffe (2015) examines the following case in chemical classification. She
discusses the use of Quantitative Structure Activity Relations (QSAR) and neural networks in
the classification of chemical reactions. For classification by QSAR and neural networking to
take place, a preliminary classification of reactions into general types of changes in entities takes
place. Scientists seek to identify the structural properties that correlate with the instantiation
of certain types of chemical reactions. Scientists have to use this methodology because of the
infinite number of chemical reactions that are possible. They seek to find a mechanism that
is responsible for the general change perceived in a chemical reaction of a certain type. This
is then further refined to identify the subtypes of reactions within that type. This case allows
us to see the complexity involved in the mechanisms for chemical reactions. As a result of
this complexity, scientists in practice have designed and used pragmatic techniques to classify
dynamic entities and their activities. Scientists need to apply QSAR methods to identify the
relevant data at the level of token mechanisms for chemical reactions before classification into
types can even commence.
Ratcliffe also (2015: 74) discusses the ambiguity in deciding the exact physicochemical prop-
erties that best determine the route of any kind of chemical reaction. These will in fact vary
on a case-by-case basis because identifying the correct properties is a matter of determining the
structure-activity relation governing a particular reaction. In some cases, the difference in partial
atomic charges, which describe the polarity of the bond, will be responsible, while in other
cases it will be effective bond polarizability. In other words, the chemical reaction can occur
via different causal routes, which amount to distinct mechanisms even though they determine
one and the same kind of chemical reaction. Again, which causal route operates will be highly
context sensitive.
Both of these cases make clear how fuzzy the boundaries of mechanisms can be. The mecha-
nisms that regulate kinds in practice are so dynamic and complex that scientists in both these
cases have had to develop techniques to identify the relevant data before classification can begin.
In both cases, conventional and contextual elements often dictate the choice of the boundaries
of the mechanisms involved.
Recall that on the HPC view, mechanisms are introduced to give a rationale for the bounda-
ries of property clusters. It is these sustaining mechanisms which make the natural kinds stable
and robust for the purposes of induction and prediction. However, looking at the relation-
ship between the two cases in scientific practice, it would appear that drawing the boundaries
206
of the underlying mechanism is itself often dictated by conventional elements and pragmatic
techniques designed by the scientists to cope with the complexity and dynamism both of the
mechanism themselves and of the clusters that are maintained by those mechanisms.
3. Circularity
Consideration of the cases in scientific practice suggests there is a problem of circularity that is
faced by the HPC view of natural kinds. Given that the boundaries of mechanisms are vague
and can be drawn in different nuanced ways, it is perfectly legitimate to ask: when can we say
that token mechanisms are mechanisms of the same type? In particular, when distinct mecha-
nisms are responsible for a property cluster, do we need to look at the resulting cluster to say
whether the mechanisms are mechanisms of the same kind? This would be particularly true in
cases where quite distinct mechanisms might both regulate a property cluster.
To make this clear, let’s return to our population of rabbits. We saw earlier that multi-
ple mechanisms might be involved in holding a property cluster together. So, for example,
in the case of rabbits, shared ecological niche and gene exchange in the population are two
mechanisms which make the property cluster a reliable one for drawing the boundaries of the
population. But these two heterogeneous mechanisms are not mechanisms of the same type.
They have in common only that they sustain the same property cluster. This evokes a problem
of circularity given that the mechanisms were introduced to help us to delineate the property
clusters on the HPC account. But now it looks as though one way of dealing with the boundary
problem for mechanisms is to look at what mechanisms are responsible for sustaining; namely,
the resulting property clusters.
We saw earlier that the mechanisms literature has traditionally been focused on solving bound-
ary problems for token mechanisms, whereas the kinds literature has been more interested in how
we group tokens into types or classes. An interesting question that emerges from considering
these cases in chemical practice is whether this type-level organization of mechanisms should be
considered to have any real ontological significance, because many heterogenous mechanisms are
often responsible for sustaining a property cluster and these can also differ depending on the envi-
ronment and context. Token-level differences between the mechanisms, which are responsible
for kind clusters in practice, make the identification of the boundaries for types of mechanisms
very difficult (perhaps even impossible). I have also shown that because of the complexity of
mechanistic classification, practitioners are often forced to use pragmatic techniques to interro-
gate voluminous data about these complex causal routes. In fact, these considerations suggest that
grouping mechanisms into types may be mostly a matter of epistemic parsimony and convenience.
If this is the case, then it now becomes much clearer that the “homeostatic mechanism” part
of the HPC view is not a simple type-level component, but rather a catchall for a very complex
set of mechanism tokens, which in scientific practice more often than not requires very com-
plex technologies for the boundaries to be drawn. The impression from scientific practice, in
fact, is that we often draw the boundaries in an artificial or idealized way, depending on which
technology is used. In the case of protein domains, for example, different algorithms will result
in distinct drawing of boundaries.
Originally, the aspiration of the HPC view was to allow a strong form of realism about
dynamic property clusters, allowing a more liberal account of natural kinds. However, if my
analysis of scientific practice is correct, then the addition of mechanisms brings with it some
deep challenges about how we can draw the boundaries. If it is a form of realism, given all of the
pragmatics that come into play, it is a very promiscuous form of realism indeed.
207
Emma Tobin
4. Conclusion
Thus far, I have discussed the HPC view of natural kinds by attempting a more in-depth
discussion of the relationship between property clusters and mechanisms. Several problems have
been raised for this relationship. I have argued for theoretical reasons that there is a problem of
circularity in the introduction of mechanisms for solving the boundary problem for kinds. I have
also shown that in scientific practice the boundary problems are much more problematic than
has been acknowledged in the mechanisms literature and that a conflation of token and type-
level classification has proved unhelpful. This means that the “homeostatic mechanism” part of
Boyd’s account is not nearly clearly enough defined.
One possible response from the new mechanists would be to argue that within the new
mechanistic framework there is no need for HPC kinds at all. Some new mechanists (Bogen,
2005; Machamer, 2004; Glennan, forthcoming) claim that both causes and mechanisms are
singular, not general or universal. However, if the circularity worries above are correct, then to
solve the boundary problem for mechanisms, we need (at the very least for pragmatic reasons)
to group those mechanisms into property clusters or mechanism kinds. In fact, this is precisely
what scientists do when using methods like QSAR and neural networking in the case of chemi-
cal reactions. The new mechanists need to either provide an argument for why we do not need
property clusters or mechanism types, or they need to show why these groupings are merely
epistemic and not of ontological significance.
Another puzzle for the mechanists is what I would call the problem of generality for a singu-
larist conception of mechanisms. A core consensus view of the mechanisms literature in general
is that we should do without laws of nature in many of their traditional roles. Particularly in
the early literature on mechanisms, mechanisms are contrasted explicitly with laws of nature
(Bechtel, 1988; Bechtel and Abrahamsen, 2005; MDC, 2000) and the mechanisms literature
can rightly be viewed as an alternative view to laws-based accounts of explanation. This is cer-
tainly the case in the philosophy of biology literature in the discussion of ceteris paribus biological
generalizations. On the mechanisms view, the resilience and necessity of these generalizations
are because of the resilience of the mechanisms. However, if we allow for promiscuity in terms
of the overlapping of mechanisms and we accept that we have to rely either on type-level mech-
anisms or the resulting property clusters which the mechanisms sustain, then are we not back to
formulating generalizations about these, which look very much like laws of nature understood
as universal generalizations?
Future work will need to addresses the question of whether the boundary problems for
mechanisms commit the new mechanist to laws of nature. There are already some movements
in this direction in recent literature. Glennan (forthcoming, ch. 4) makes the interesting sug-
gestion that models are the tools that allow practitioners to classify particulars into kinds. In
Chapter 12 of this volume, Andersen also offers an argument for how laws and mechanisms
relate without either conceptually or ontologically reducing to one another.
In conclusion, despite the initial ambition of having a more liberal view of classification,
which would allow for the joint classification of mechanisms and kinds, considerations from
scientific practice together with theoretical concerns appear to undermine the HPC account
significantly. Cases in scientific practice show that there is a boundary problem for locating the
vague boundaries of mechanisms. I have argued that the mechanisms, which regulate kinds in
practice, are so dynamic and complex that scientists in practice are often forced to develop tech-
niques to identify the relevant data before classification can even begin. Moreover, depending
on the techniques used, the boundaries are often drawn differently. Secondly, I argue that there
is a theoretical problem of circularity because the vagueness of mechanism boundaries requires
208
classification into kinds to avoid ambiguity and overlap. It seems clear from a closer examination
of the relationship between mechanisms and kinds that mechanisms alone do not have the ontic
resources to allow us to identify nature’s joints.
Notes
1 I would like to thank Professor Stuart Glennan and Dr Phyllis Illari for excellent guidance and
feedback on earlier drafts of this chapter. Their editorial role has been crucial in producing the final
article. I am also grateful to Professor Alexander Bird and to staff and students at the Department
of History and Philosophy of Science in Cambridge who all commented on earlier versions of
this chapter.
2 It is worth noting that LaPorte uses this case not as an argument for conventionalism about natural kinds,
but to reject Kripke’s view that the essences of natural kinds are discovered a posteriori. I am using it
here as a case to show how a conventionalist might argue that pragmatic and conventional elements play
a role in the delineation of natural kinds.
References
Bechtel, W., 1988, Philosophy of Science: An Overview for Cognitive Science, Hillsdale, NJ: Erlbaum. Italian
translation: Filosofia della scienza e scienza cognitiva, Gius. Laterza & Figli, 1995.
Bechtel, W. and A. Abrahamsen, 2005, “Explanation: A Mechanistic Alternative”, Studies in History and
Philosophy of the Biological and Biomedical Sciences, 36: 421–441.
Bird, A. and E. Tobin, 2008, “Natural Kinds”, Stanford Encyclopedia of Philosophy, https://plato.stanford.
edu/entries/natural-kinds/.
Bogen, J., 2005, “Regularities and Causality: Generalizations and Causal Explanations”, Studies in History
and Philosophy of Science Part C, 36(2): 397–420.
Boyd, R., 1991, “Realism, Anti-Foundationalism and the Enthusiasm for Natural Kinds”, Philosophical
Studies, 61: 127–148.
Boyd, R., 1999a, “Homeostasis, Species, and Higher Taxa”, in R. Wilson (ed.), Species: New Interdisciplinary
Essays, Cambridge, MA: MIT Press: 141–186.
Boyd, R., 1999b, “Kinds, Complexity and Multiple Realization”, Philosophical Studies, 95: 67–98.
Craver, C.F., 2001, “Role Functions, Mechanisms & Hierarchy”, Philosophy of Science, 68(1): 53–74.
Craver, C.F., 2007, Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience, Oxford:
Clarendon Press.
Craver, C.F., 2009, “Mechanisms and Natural Kinds”, Philosophical Psychology, 22: 575–594.
Craver, C.F. 2015, “Mechanisms and Emergence: A Reply to Denis, C. Martin”, in T. Metzinger &
J.M. Windt (eds), Open MIND, Frankfurt am Main: MIND Group. doi: 10.15502/9783958571099.
Dupré, J., 1993, The Disorder of Things: Metaphysical Foundations of the Disunity of Science, Cambridge, MA:
Ellis, B., 2001, Scientific Essentialism: Cambridge Studies in Philosophy, Cambridge: Cambridge University Press.
Ghiselin, M.T., 1974, “A Radical Solution to the Species Problem”, Systematic Zoology, 23: 536–544.
Ghiselin, M.T., 1987, “Species Concepts, Individuality and Objectivity”, Biology and Philosophy, 2:
127–144.
Ghiselin, M.T., 1997, Metaphysics and the Origin of Species, Albany: State University of New York Press.
Glennan, S., 1996, “Mechanism and the Nature of Causation”, Erkenntnis, 44: 49–71.
Glennan, S., 2002, “Rethinking Mechanistic Explanation”, Philosophy of Science, 69: S342–S353.
Glennan, S., Forthcoming, The New Mechanical Philosophy, Oxford: Oxford University Press.
Hacking, I., 1991a, “A Tradition of Natural Kinds”, Philosophical Studies, 61: 109–126.
Hacking, I., 1991b, “On Boyd”, Philosophical Studies, 61: 149–154.
Hacking, I., 1999, The Social Construction of What? Cambridge, MA: Harvard University Press.
Hull, D., 1976, “Are Species Really Individuals?”, Systematic Zoology, 25: 174–191.
Hull, D., 1978, “A Matter of Individuality”, Philosophy of Science, 45: 335–360.
Illari, P., and J. Williamson, 2012, “What Is a Mechanism?: Thinking about Mechanisms Across the
Sciences”, European Journal for Philosophy of Science, 2: 119–135.
209
Emma Tobin
Jonides, J. and Awh, E., 2001, “Overlapping Mechanisms of Attention and Spatial Working Memory”,
Trends in Cognitive Science, 5(3): 119–126.
Khalidi, M., 1998, “Natural Kinds and Crosscutting Categories”, Journal of Philosophy, 95: 33–50.
Kuhn, T., 2000, The Road Since Structure, Chicago: University of Chicago Press.
LaPorte, J., 2004, Natural Kinds and Conceptual Change, Cambridge: Cambridge University Press.
Lowe, E.J., 2006, The Four-Category Ontology: A Metaphysical Foundation for Natural Science, Oxford:
Clarendon Press.
Machamer, P., 2004, “Activities and Causation: The Metaphysics and Epistemology of Mechanisms”,
International Studies in the Philosophy of Science, 18: 27–39.
Machamer, P.K., L. Darden, and C.F. Craver, 2000, “Thinking about Mechanisms”, Philosophy of Science,
67: 1–25.
Ratcliffe. S., 2015, A Metaphysics for the Classification of Chemical Reactions in Practice, PhD Thesis, UCL
Discovery. Available at: http://discovery.ucl.ac.uk/1464574/1/Steph%20Ratcliffe%20PhD%20
Final%20copy.pdf.
Tobin, E., 2010a, “Crosscutting Natural Kinds and the Hierarchy Thesis”, in Helen Beebee and Nigel
Sabbarton-Leary (eds.), The Semantics and Metaphysics of Natural Kinds, London: Routledge: 1–179.
Tobin, E., 2010b, “Microstructuralism and Macromolecules: The Case of Moonlighting Proteins”,
Foundations of Chemistry, 12(1): 41–54.
210
PART III
Mechanisms and the

philosophy of science
16
MECHANISTIC EXPLANATION
AND ITS LIMITS
Marta Halina
1. Introduction
Many attempts have been made by philosophers to provide a satisfying account of what con-
stitutes an explanation in the sciences. Over the last few decades, particular attention has been
given to the role of mechanisms in providing such an account in the life sciences. Mechanistic
explanations have proved to provide a powerful account of both the practices of biologists and
the normative constraints on explanation. Even so, many philosophers are wary of explanatory
hegemony, holding instead that a plurality of accounts of explanation will be needed to capture
the diversity of scientific practice (Godfrey-Smith 2003).
This chapter examines the strengths and limitations of mechanistic explanation. It does
this by considering the advantages of the mechanistic account over previous models of sci-
entific explanation, as well as its descriptive adequacy with respect to various aspects of
scientific practice. Concerning the latter, I focus on how mechanistic explanation accounts
for the following three aspects of science: the appeal to non-mechanistic explanations, the
use of abstract and idealized models, and the generality of explanation (for more discussion
along these lines, see Chapter 17). Recently, critics of mechanistic explanation have cited
these areas as posing problems for the mechanistic account. Examining whether and why this
is the case will allow us to probe the boundaries of mechanistic explanation—highlighting
its strengths and potential weaknesses. In each of these three cases, I discuss tools that the
mechanist might use to address the problems advanced. As these areas represent active
domains of research, I do not attempt to adjudicate them here. This chapter instead aims
to introduce the reader to the current state of the art on mechanistic explanation, as well as
potential directions for moving forward.
I begin in section 2 by introducing the idea of a general philosophical account of expla-
nation and the various considerations and constraints that factor into the construction and
evaluation of such an account. In section 3, I briefly introduce two precursors to mechanistic
explanation before turning to mechanistic explanation in section 4. Section 5 examines the
potential limitations of this model of explanation as they manifest in the three areas of scientific
practice noted above.
213
Marta Halina
2. What is explanation?
To provide an account of mechanistic explanation, it is important to first understand what is
meant by “explanation.” Generally, an explanation can be understood as an answer to a why
question. We might ask why the sky is blue or why human skin wrinkles when submerged in
water. The thing in need of an explanation (the blue sky or wrinkled skin) is the explanandum,
whereas the explanation for this phenomenon is the explanans. Scientists may seek explanation
as an end in itself—that is, to better understand a given phenomenon—or as a means to other
ends, such as prediction and intervention (for a general introduction to scientific explanation,
see Woodward 2014).
In developing a philosophical account of explanation, one must be clear about the standards
that will be used to evaluate such an account. Should a philosophical account of explanation,
for example, capture the way in which the term “explanation” is used in ordinary life? Or
should it instead focus on the way it is used in a particular science, such as theoretical physics
or molecular biology? Alternatively, should philosophers base their account of explanation on
accounts of evidence or truth or justification, or should their main concern be capturing the
explanatory practices of scientists? To date, philosophers have taken a variety of approaches to
developing and evaluating general accounts of explanation. This has led to some confusion, as
one might develop an account using one standard, such as capturing the normative practices
of a particular scientific field, which might then be critiqued using a different standard, such as
capturing ordinary language use (Ruben 2012).
For the purposes of this chapter, I adopt explanatory demarcation and normativity as the
main standards by which to evaluate an account of explanation (see Craver 2014). Explanatory
demarcation is the practice of distinguishing explanation from other activities, such as descrip-
tion or prediction. For example, identifying a bird as belonging to a particular species involves
description and categorization; however, scientists do not view this as explanatory. Explanatory
normativity, on the other hand, is the practice of distinguishing good explanations from bad.
Biologists reject the claim that wet skin wrinkles because water enters the epidermis through
osmosis, causing it to swell. Instead, they hold that this phenomenon is due to vein constric-
tion triggered by the sympathetic nervous system. At present, the former is considered a bad
explanation and the latter good.
The above criteria are widely adopted within philosophy of science as evaluative stand-
ards for philosophical accounts of explanation (Lipton 2004). One reason for this is that they
enable philosophers to construct accounts of explanation that remain true to the practices of
scientists. Although an account of explanation might be constructed that accurately describes
our ordinary language use (see Wright 2012), such practices may fail to conform to the
explanatory standards used in science. Insofar as this is the case, and our goal is to develop an
account of scientific explanation, common sense and ordinary language practices may lead us
astray. Within the broad realm of scientific practice, however, this chapter endorses a plural-
ist stance toward explanation. Discussions of explanation have often presumed that there is
one explanatory form or relation that will capture the explanatory practices of all scientific
disciplines (Woodward 2014). I follow others in rejecting this assumption (Godfrey-Smith
2003; Kellert, Longino, and Waters 2006). Perhaps we will discover that explanation in phys-
ics and anthropology and molecular biology all take the same form, but we should not begin
our inquiry assuming that this is the case. With this in mind, the focus of this chapter will be
on biological explanation (although some general examples will be provided for illustrative
purposes). Lastly, it is important to note that explanatory demarcation and normativity are not
the only features of explanation discussed by philosophers. Peter Lipton (2004), for example,
214
Mechanistic explanation and its limits
highlights that explanations commonly have a self-evidencing nature (the explanandum may
serve as evidence for the explanans) and lack explanatory regress (something can serve as an
explanation even when it itself is unexplained). Although these may be important features of
scientific explanation, I do not discuss them here due to limited space and the fact that they
do not feature highly in the debates under consideration.
Before introducing precursors to mechanistic explanation in the next section, it would be
useful to highlight a feature of explanation that is receiving increasing attention in the literature.
This is the question of whether explanation should be understood as an epistemic activity or
ontic feature of the world. Broadly, epistemic explanation concerns the models, representations,
and activities used to communicate and elicit understanding of a target system. When I com-
municate to a group of students how photosynthesis works, I explain in the epistemic sense.
Ontic explanation, in contrast, refers to that aspect of the world that explains why some worldly
event happened (Craver 2007; Strevens 2008; Craver 2014). When I say that the ice on the road
explains the car accident, I am using explanation in this ontic sense. The term “explanation”
is often used in both ways (Illari and Williamson 2011; Colombo, Hartmann, and van Iersel
2014). For the purposes of this chapter, I adopt an ecumenical account that accepts both senses
of explanation as important for understanding the explanatory practices of science (Illari 2013;
but see Sheredos 2015). Distinguishing them is important, however, as they impose different
constraints on good explanation, as we will see in section 5 (see also Illari 2013).
3. Precursors to mechanistic explanation

Contemporary discussions of scientific explanation find their origins in the development of
the deductive-nomological (DN) model of explanation developed in the mid-1900s by Carl
Hempel and others (Hempel and Oppenheim 1948; Hempel 1965). One of the main rivals to
the DN model of explanation is the causal-mechanical account of explanation (Salmon 1984).
A descendant of causal accounts of explanation, mechanistic explanation emerged at the turn of
the century and has since become the dominant philosophical account of scientific explanation
(Bechtel and Richardson 1993; Glennan 1996; Machamer, Darden, and Craver 2000). This sec-
tion introduces the DN and causal models of explanation, reviewing their strengths and weak-
ness, before turning to mechanistic explanation in the following section.
The DN model was the dominant model of scientific explanation in the twentieth century.
According to this model, to successfully explain a phenomenon, one must derive it from a set
of premises—premises that tell us why the conclusion is true. Under this view, the premises
constitute the explanans, the conclusion the explanandum, and they relate to each as parts of
a deductive argument. Not every deductive argument counts as an explanation, however. To
count as an explanation, Hempel (1965) maintained that the premises must include at least one
generality or law of nature. Hence the “deductive-nomological” account, where “nomologi-
cal” derives from the Greek word “nomos” for law. For example, if someone were to point to
a pot of boiling water and ask, “why is that liquid boiling?” an explanation would take the form
of noting that water heated to 100°C at 1 atmosphere of pressure boils (law of nature), that the
liquid in the pot is water (particular fact), and that it has been heated to 100°C at 1 atmosphere
of pressure (particular fact). This explains why the liquid is boiling.
The DN model has many virtues. One of the virtues is that it captures the epistemic advan-
tages of both explanation and prediction. Under the DN model, there is a symmetry between
explanation and prediction: I can explain why the liquid is boiling by citing the above laws of
nature and particular facts, but these laws and facts can also be used to predict what will happen
when I heat water to the conditions of 100°C at 1 atmosphere of pressure. As noted in section 2,
215
Marta Halina
explanation is often thought to be valuable as a means for prediction and intervention. The DN
model provides a nice account of why this is the case: explanation simply is a form of prediction.
Although the DN model has many virtues, it is now widely rejected as an account of expla-
nation because of its well-known shortcomings (see Woodward 2014 for an overview). The
most commonly noted shortcomings are that it is both over-permissive and too restrictive in
what counts as an explanation. The DN model is over-permissive in that it includes many things
as explanatory, which we generally do not accept as explanations. We take a drop in atmos-
pheric pressure as explaining a storm, but not a drop in the reading of a barometer, even when
both of these variables reliably vary with the onset of storms. However, there is nothing in the
DN model that excludes us from explaining a storm by appealing to the regularity that a low
barometer reading tends to precede storms. Also, we can insert irrelevant information into the
premises of a deductive argument without affecting its validity and soundness. For example, I
could add to our water boiling argument the premise that it is Wednesday. However, we would
not want to say that the fact that it is Wednesday explains why the water boils. Again, the DN
model is too permissive in counting this as explanatory. On the other hand, the DN model is
too restrictive in that it requires the explanans to include a regularity or law of nature. Scientific
explanations do not always include such laws, however (Sober 1997). Biological explanations,
for example, often include complex, idiosyncratic systems that do not exhibit law-like regulari-
ties (but see section 5).
The shortcomings of the DN model have played a key role in the development and
acceptance of the causal-mechanical account of explanation. Broadly, this account holds that
a phenomenon is explained by its causes (Salmon 1984). Not only does this account over-
come the problems of over-permissiveness and over-restrictiveness of the DN model, but it also
explains why explanatory information is delimited in the ways described above. Why does the
barometer reading not explain the storm? Why does the fact that it is Wednesday not explain
the boiling water? According to the causal-mechanical account of explanation, the reason is that
these factors play no causal role in producing the phenomena to be explained. Low atmospheric
pressure is causally involved in the production of a storm and heating water to 100°C is caus-
ally involved in its boiling. It is in virtue of these facts that the phenomenon is explained. The
causal-mechanical account also avoids the over-restrictiveness of the DN model because it does
not require laws of nature for explanation. The First World War might have been caused—and
thus explained—by the fact that the driver of Archduke Franz Ferdinand took a wrong turn in
Sarajevo regardless of the fact that this was a one-time and unlikely event.
Mechanistic explanation is a form of causal-mechanical explanation in that it maintains that
mechanisms are explanatory and causes are one of the four features of mechanisms (Craver and
Tabery 2015). It also extends this account, however, to include constitutive causal-mechanical
explanations in addition to etiological ones (Craver 2007). An etiological causal explanation
involves explaining an event by citing its antecedent causes, such as when we explain the car
accident by citing the ice on the road. A constitutive explanation, in contrast, explains a phe-
nomenon by citing the mechanism responsible for it, such as explaining cancer by citing the
mechanisms responsible for abnormal cell growth. Explanations of both varieties are used widely
in the biological sciences (Bechtel and Richardson 1993). The rest of this chapter explicates the
notion of mechanistic explanation and considers its limitations.
4. Mechanistic explanation
Contemporary notions of mechanistic explanation began to emerge in the 1990s, and vari-
ous accounts have been developed over the last few decades (for an overview, see Craver
216
and Tabery 2015; Chapter 1 of this volume). For our purposes, we can adopt the ecumenical
definition of mechanism advanced by Illari and Williamson (2011): “A mechanism for a phe-
nomenon consists of entities and activities organized in such a way that they are responsible for
the phenomenon” (120). Philosophers have discussed and debated many aspects of the nature
of mechanisms, including how to determine its boundaries, define its parts, identify its activities
and functions, and more (see Part II of this volume). For our purposes, it is sufficient to identify
the way in which mechanisms are taken to be explanatory. Generally, mechanistic explanations
are taken to “explain why by explaining how” (Bechtel and Abrahamsen 2005, 422). One can,
for example, explain why bones become brittle in old age by specifying the biological factors
and processes involved in the reduction of bone mineral density. By providing a description
of the mechanisms responsible for a phenomenon, one provides an explanation for why that
particular phenomenon occurs and why it has the properties that it does. The purpose of
identifying operating parts and determining their organization is to go beyond describing the
phenomenon to showing how the working entities cause and constitute the phenomenon.
Mechanistic explanations inherit many of the advantages of the causal-mechanical account of
explanation. They capture the asymmetry of explanation: those causes or mechanisms respon-
sible for a phenomenon explain that phenomenon, but not vice versa (the influenza virus
explains swine flu, but swine flu does not explain the influenza virus). They also account for
why irrelevant information is not explanatory. The fact that the water boiled on Wednesday
might be true, but it is not explanatory because it is not part of the causal mechanism respon-
sible for the water boiling. Like the causal-mechanical account, mechanistic explanation also
does not require laws of nature, so it is not restrictive in this respect. Mechanistic explanation
has many of the virtues of the DN model as well. It provides an account of the epistemic
advantages of prediction and intervention: understanding how something works gives us
knowledge about how to effectively intervene on that system and predict its future states or
how it will behave in a new context.
5. The limits of mechanistic explanation

Discussions concerning the possible limitations of mechanistic explanation are usefully framed
in terms of the two criteria for a successful account of explanation introduced in section 2. A
philosophical account of explanation should capture the scientific practices of distinguishing
explanation from other activities (explanatory demarcation) and good explanations from bad
(explanatory normativity). In what follows, I consider whether mechanistic explanation is suc-
cessful in these two respects. Rather than attempting an exhaustive survey, however, I focus
on those aspects of mechanistic explanation that have been points of recent controversy. These
concern the use of non-mechanistic explanation in the sciences, abstract and idealized explana-
tory models, and the generality of scientific explanation. Examining these areas will give us a
good idea of the current perceived limitations of mechanistic explanation and how these poten-
tial limitations might be overcome.
One criticism of mechanistic explanation is that it excludes certain things as explanatory when
those things are indeed explanatory. This is a failure of explanatory demarcation insofar as mech-
anists claim that something is not an explanation when it in fact is. One of the most prominent
criticisms of this type comes from advocates of dynamical modeling (see Chapter 20). Dynamical
models track and predict the behavior of complex systems using mathematical tools such as
difference and differential equations. These models can be extremely powerful in both their pre-
dictive success and ability to unify disparate phenomena. This has led some theorists to maintain
that they are explanatory (Chemero and Silberstein 2008; Stepp, Chemero, and Tuvey 2011).
217
Marta Halina
Mechanists have rejected dynamical explanation, however, for many of the same reasons that
have led to the rejection of the DN model. They maintain that dynamical models are either
explanatory in virtue of conveying information about the mechanism responsible for the phe-
nomenon or not explanatory at all (Kaplan and Bechtel 2011; Kaplan and Craver 2011; Chapter
20 of this volume). If dynamical models fall in the former category, then they are simply instances
of mechanistic explanation; if they fall in the latter, then they do not present a counterexample
to mechanistic explanation because they are not explanatory. Mechanists hold that models in the
latter category—those that describe and predict a phenomenon without providing information
about how it works—are problematic because descriptions and predictions are well known to be
insufficient for explanation (for the reasons discussed in section 3). Although dynamical models
may be insufficient for explanation, it is important to note that this does not mean they do not
contribute to our explanatory practices at all. As Bechtel and Abrahamsen (2010, 2011) show,
dynamical modeling is often used to understand the behavior of complex mechanisms in fields
like chronobiology and cognitive science. Under this view, tools like quantitative computational
modeling are essential for explanation because they enable researchers to investigate how the
properties of the parts and operations of a mechanism dynamically change over time.
Batterman and Rice (2014) and Ross (2015) argue that dynamical models are explanatory
not in virtue of their descriptive and predictive accuracy alone, but also in virtue of abstracting
from irrelevant details (see also Batterman 2001). They call such explanations “minimal models”
and argue that they constitute not only an alternative to mechanistic explanation, but also to all
“common features accounts” of explanation or those that take the explanatory power of a model
to derive from its representation of the processes responsible for the explanandum phenomenon.
According to these authors, minimal models explain by showing that a subset of possible and
actual systems (including a minimal model and a target system) fall within the same universality
class. This is despite the systems differing in their microdetails. Although the minimal model
and target system will have other features in common (besides the macrobehavior in question),
the proponents of this view hold that it is not these common features that explain the macrobe-
havior of the target system, but rather that these features are a “by-product of the mathematical
delimitation of the universality class” (Batterman and Rice 2014, 362).
Whether the minimal model is a genuine alternative approach to mechanistic explanation is
unclear, however. As Lange (2015) argues, minimal model explanations appear to be another
form of common features account. First, as mentioned above, explanations are generally asym-
metrical: the explanans explains the explanandum, but not vice versa. Minimal models appear
to lack this property. If a minimal model explains a target system in virtue of falling into the
same universality class, it is not clear why the target system cannot also explain the minimal
model. This is inconsistent with the explanatory practices of scientists, however, as they gener-
ally do not take target systems to be explanatory in this way. One could solve this problem by
interpreting minimal model explanations as explaining in virtue of identifying those features that
lead to the production of the macrobehavior in question. Under this view, the minimal model
explains because it includes all and only those features necessary to produce the macrobehav-
ior (as opposed to the target, which includes many other features). Insofar as the target system
has these features too (i.e., insofar as it shares the features exhibited by the minimal model),
the explanation is successful. This account preserves the idea that minimal models explain by
abstracting and idealizing away from irrelevant detail, but it does so by interpreting these models
in terms of a common features account.
Before closing this discussion of potential non-mechanistic explanations, it is important to
note that mechanists do not maintain that all scientific explanations are mechanistic. That is, they
hold that there are many structures and relations that we take to be explanatory, but which are
218
not causal or mechanistic. For instance, Craver (2014) maintains that, “attractors, final causes,
laws, norms, reasons, statistical relevance relations, symmetries, and transmission of marks” might
all be considered explanatory, depending on the field of inquiry (29). Strevens (2008) also cites
things like moral and aesthetic relations, holding that the explanatory relevance of these rela-
tions stems from something other than their causal import (see also Lipton 2004). The dispute
about non-mechanistic explanation then is not whether they exist, but rather what form they
take and what renders them explanatory. Mechanists deny that dynamical and minimal models
are explanatory independently of the information they convey about the workings of the target
system. Advocates of these models need to do more to show that this is indeed the case.
One of the great strengths of the mechanist account of explanation has been its ability
to elucidate a wide range of scientific practices—practices that were difficult to account for
under traditional views of explanation, such as the DN model (Bechtel and Abrahamsen
2005). Recently, however, this aspect of mechanistic explanation has come under criticism,
particularly with respect to the use of abstract and idealized models in science. The criticism
holds that mechanistic explanation is committed to a representational ideal of completeness—
the ideal that the more features and causal relationships a model represents, and the more
accurately it represents them, the better the model (see Weisberg 2013 and Chapter 17 of this
volume for discussion). If this is the standard to which we should hold explanatory models,
however, it is in conflict with the actual practices of scientists. It is well known that most
scientific models are highly abstract and idealized. They are abstract in the sense of omitting
detail about the target system and idealized in the sense of distorting elements of that sys-
tem. Explanatory models are thus intentionally incomplete and inaccurate; further, this is not
viewed as a deficit by practicing scientists. So the representational ideal of completeness is not
something to which scientists appear to adhere.
This ideal of completeness is attributed specifically to the views of mechanists like Machamer,
Darden, Craver, and Kaplan (see Levy and Bechtel 2013; Chirimuuta 2014; Batterman and
Rice 2014; Levy 2014; Love and Nathan 2014). For example, Levy and Bechtel (2013) write,
“Machamer, Darden, and Craver appear to treat abstractions as, at best, templates for explana-
tion. They regard the filling in of concrete detail as a hallmark of explanatory progress” (258).
Similarly, Chirimuuta (2014) attributes to Craver and Kaplan the claim that, “the hypothetical,
maximally complete and detailed representation of the mechanism is the one best explanatory
model onto which all others aim to converge” (132). Incomplete models are sketches that fall
short of their explanatory goal; the more detailed and precise they get, the better. This is what
Chirimuuta refers to as the “More Details the Better” assumption (132). Levy (2014) also attrib-
utes the representational ideal of completeness to Craver, writing, “sketches are typically steps
along the way to a better explanation. If all goes well, the gaps are filled and the mechanism is
described in full detail. Once that occurs, the sketch is transformed into a satisfactory explana-
tion” (479; see also Batterman and Rice 2014, 352).
The same point is made against mechanistic explanation with respect to idealization. The
representational ideal of completeness requires that explanatory models are not only complete,
but also accurate. This, however, is at odds with the widespread use of idealized explanatory
models. As Love and Nathan (2014) write: “The idealization of causal relations—the intentional
misrepresentation of how the mechanism produces the phenomenon—means that these models
do not show how the mechanism actually works. Mechanistic explanations thus appear to fail
according to their own criteria” (13).
The critics hold then that mechanists such as those above are committed to a representational
ideal of completeness and that this renders their account of explanation descriptively inadequate.
Abstract and idealized models are taken to be explanatory by practicing scientists despite the fact
219
Marta Halina
that they are incomplete and inaccurate. If the critics are right, this represents a crucial limitation
of the mechanistic account. Mechanistic explanation fails to capture the practices of the explana-
tory normativity of science: what scientists take to constitute a good explanation is at odds with
what the mechanist account maintains.
Are the critics right about this shortcoming of mechanistic explanation? I argue that they are
not for the reason that completeness is not the sole ideal that guides mechanistic explanation.
To see this, recall that under the mechanistic account, it is the causal mechanism that produces,
constitutes, or maintains a phenomenon of interest that explains that phenomenon—the pro-
cess by which plants convert solar energy into usable carbohydrates explains photosynthesis,
for example. For those who hold the ontic view of explanation (introduced in section 2), it
is the actual entities and activities found in plants that explain photosynthesis. Descriptions of
photosynthetic processes (in the form of diagrams, linguistic descriptions, physical models, etc.)
are explanatory in a derivative sense: we take them to be explanatory because they convey
information about the real-world explanation. For those who hold the epistemic view, it is
the models and communicative acts that we use to convey information about the mechanisms
responsible for photosynthesis that constitute the explanation. These models and acts might be
explanatory in virtue of conveying information about the target mechanism, but the mecha-
nism itself is not explanatory. Both ontic and epistemic views, however, hold that mechanisms
are the targets of explanatory models and that the explanatory power of a model is in part a
function of the information that model conveys about this target (Wright 2012; Illari 2013;
Craver 2014). Accurately representing the target mechanism then is an important success con-
dition for explanatory models.
Mechanisms constrain those communicative acts, texts, and representations that we
take to be explanatory. If this were the only dimension along which explanatory models
were assessed, then the practices of abstraction and idealization in science would indeed
be puzzling. Why settle for incomplete and inaccurate models when we can do better?
Crucially, however, this is not the only constraint on explanatory models. As Kaplan and
Craver (2011) write,
the idea of an ideally complete how-actually model, one that includes all of the rele-
vant causes and components in a given mechanism, no matter how remote, negligible,
or tiny, without abstraction or idealization, is a philosopher’s fiction. Science would
be strikingly inefficient and useless both for human understanding and for practical
application if it dealt in such painstaking minutiae.
(609–10)
An explanatory model serves many aims. In addition to conveying information about the target
mechanism, such a model must often be intelligible, easily shared with a broader community,
practical to work with, and more. Some of these desiderata pull in different directions. For
example, making a model more intelligible will often require simplifying it by removing detail.
This may decrease the amount of information that the model conveys about the target mecha-
nism, but this does not mean that we should avoid such simplifications. If explanatory models
were unintelligible, researchers would be unable to use them to communicate findings, design
experiments, interpret results, train students, etc. All of these activities are important for advanc-
ing research, including the further discovery of mechanisms.
Various accounts of scientific practice support the claim that multiple desiderata constrain
the construction and evaluation of explanatory models. For example, Weisberg (2013) discusses
how different goals give rise to different forms of idealization (where an idealization is the
220
intentional distortion of a target system as mentioned above). Researchers engage in Galilean

idealization when their goal is to increase the tractability of a model. In this case, elements of
the target system are distorted not because they are unimportant, but because they interfere with
a researcher’s ability to understand and work with the model. In contrast, minimal idealization
involves introducing distortions for the sake of capturing the core causal factors of a target sys-
tem. Here distortions are introduced not for the purpose of making a model more intelligible,
but to remove information about those parts of the world that are taken to be explanatorily irrel-
evant. Weisberg notes that desiderata such as simplicity and accuracy will often trade off of each
other, in which case multiple models may be used for attaining one’s research goals. Especially
when the target is a complex system, we should expect that multiple models will be deployed,
each of which will capture some aspect of the core causal mechanism, while simplifying other
aspects to increase intelligibility. Taken together, these models may advance us toward capturing
the target mechanism in a way that can be understood and applied.
Even the critics’ descriptions of explanatory practice suggest the operation of multiple
desiderata. As Love and Nathan (2014) note, “These kinds of [highly abstract and idealized] dia-
grammatic representations are common in textbooks. However, in more advanced discussions,
we find increasingly detailed representations of eukaryotic gene expression and more precise
narrative descriptions of the mechanism” (8). This suggests that intelligibility may take priority
in pedagogical contexts; while conveying information about the target mechanism may become
more important in those contexts where advanced researchers are attempting to understand and
intervene on a target system.
Are mechanistic explanations generalizable? This is an important question because good
explanations seem to be those that can be applied to a variety of systems and contexts. This
question is all the more pressing in biology, where most mechanistic models are developed by
investigating only a small handful of organisms, those known as “model organisms” (Weber
2005; Ankeny and Leonelli 2011). The mechanistic explanations developed by studying model
organisms are often done so under the assumption that they will apply to a wide variety of
organisms, not simply to those being investigated. Insofar as explanations in the sciences are
general, a philosophical account of explanation should be able to account for this.1
Glennan (2002) holds that mechanistic explanations are generalizable, writing, “although
any particular mechanism will occupy a particular region of space-time, it is an important
feature of our world that it often contains many tokens of a single type of mechanism”
(S345). One might ask what “important feature of our world” is responsible for the fact that
similar mechanisms tend to appear again and again (for more discussion on the regularity
of mechanisms, see Chapter 12). When it comes to biological organisms, evolution seems
to provide an answer. Mechanisms that have been inherited by descendants of a common
ancestor will have features in common. Hence, if an explanation is provided for a mechanism
that is evolutionarily conserved in this way, that explanation will likely apply to a variety
of organisms (Bechtel 2009; Halina and Bechtel 2013). Does this mean that mechanistic
explanations should be confined to those mechanisms that are highly evolutionarily con-
served? This would be unreasonable. Scientists provide mechanistic explanations for a wide
variety of phenomena, whether they know them to be evolutionarily conserved or not.
Furthermore, even conserved mechanisms exhibit variation.
Bechtel and Abrahamsen (2005) suggest that mechanisms might be categorized by how
similar they are, rather than by whether they are exactly the same. This approach is promis-
ing as an account of how scientists identify mechanisms as belonging to a particular type (see
Glennan forthcoming). Also, it is important to note that recognizing the similarity between
mechanisms is a useful investigatory and explanatory tool, even if the mechanisms differ in
221
Marta Halina
other respects. Consider, for example, the use of the model organism Drosophila to study
human ovarian cancer. In Drosophila, border cells in the ovary migrate in preparation for
egg fertilization. Studying the migratory behavior of these cells has proven useful for under-
standing particular aspects of ovarian cancer in humans—namely, the movement of cancer
cells from one site to another. The spread of ovarian cancer is a complicated process involv-
ing many variables of which the migration of the malignant cancer cells is only one part.
However, the study of normal-functioning border cells in Drosophila has the potential to
shed valuable light on at least this aspect of metastasis (Naora and Montell 2005). Although
cells in Drosophila and human ovarian cancer cells are neither completely similar nor evolu-
tionarily conserved processes, a mechanistic explanation of the former can still be usefully
applied to the latter.
In addition to similarities serving as heuristics for mechanism discovery, differences can also
play such a role. For example, one strategy for discovery in circadian rhythms research is to
search for the homolog of a protein found in one species in another species. If this homolog
then plays a different role in the second species, researchers identify this new role. This in
turn leads them to look for a protein that would perform this new role in the original spe-
cies (see Halina and Bechtel 2013 for a specific example of such a case). Such back-and-forth
research fueled by the discovery and pursuit of homologs is common in the biological sciences
(see Bechtel 2009; for more on heuristics for mechanism discovery, see also Craver and Darden
2013 and Chapter 19 of this volume).
Thus, to the extent in which mechanisms are evolutionarily conserved, we can expect that
mechanistic explanations can be generalized. However, formulating models of mechanisms
that are not evolutionarily conserved, or that exhibit variation across individuals or species, is
still useful as a heuristic for discovery. Such heuristics are crucial for mechanistic explanations,
as such explanations are explanatory precisely in virtue of providing information about the
target mechanism. Little work has been done on the generality of mechanistic explanation
to date, but it would be premature to say that this account is unable to capture this aspect
of scientific practice, at least in biology. Whether the same is true of other sciences, such as
physics and chemistry, is an open question. Additional candidates for generating similarities
across mechanisms might include convergent evolution, self-organization, cultural transmission,
and intentional design (Glennan forthcoming).
6. Conclusion
Mechanistic explanation has played a pivotal role in developing our understanding of explana-
tory practices in biology and other sciences. This chapter has highlighted the advantages of the
mechanistic approach over other accounts of explanation, as well as its descriptive adequacy
with respect to the use of idealized, abstract, and general models. Our understanding of explana-
tory practices in sciences like biology is still rudimentary. As work in this area continues, the
successful application and limitations of mechanistic explanation will become clearer. Until
then, mechanistic explanation might serve as its own heuristic for discovery—finding the ways
in which it coincides and diverges from the explanatory practices found in science will no doubt
improve our understanding of those practices.
Note
1 Although I will not employ this distinction here, see Sheredos 2015 for a discussion of the difference
between generality and scope and its implications for mechanistic explanation.
222
References
Ankeny, R.A. and Leonelli, S. (2011) “What’s So Special about Model Organisms?” Studies in History and
Philosophy of Science, 42, 313–323.
Batterman, R.W. (2001) The Devil in the Details: Asymptotic Reasoning in Explanation, Reduction, and
Emergence, New York: Oxford University Press.
Batterman, R.W. and Rice, C.C. (2014) “Minimal Model Explanations,” Philosophy of Science, 81,
349–376.
Bechtel, W. (2009) “Generalization and Discovery by Assuming Conserved Mechanisms: Cross-Species
Research on Circadian Oscillators,” Philosophy of Science, 76, 762–773.
Bechtel, W. and Abrahamsen, A. (2005) “Explanation: A Mechanist Alternative,” Studies in History and
Philosophy of Biol & Biomed Sci, 36(2), 21–21.
Bechtel, W. and Abrahamsen, A. (2010) “Dynamic Mechanistic Explanation: Computational Modeling of
Circadian Rhythms as an Exemplar for Cognitive Science,” Studies in History and Philosophy of Science
Part A, 41(3), 321–333.
Bechtel, W. and Abrahamsen, A. (2011) “Complex Biological Mechanisms: Cyclic, Oscillatory, and
Autonomous,” in C.A. Hooker (ed.) Philosophy of Complex Systems. Handbook of the Philosophy of
Science, Volume 10, New York: Elsevier, 257–285.
Bechtel, W. and Richardson, R.C. (1993) Discovering Complexity: Decomposition and Localization as Strategies
in Scientific Research, Princeton, NJ: Princeton University Press.
Chemero, A. and Silberstein, M. (2008) “After the Philosophy of Mind: Replacing Scholasticism with
Science,” Philosophy of Science, 75, 1–27.
Chirimuuta, M. (2014) “Minimal Models and Canonical Neural Computations: The Distinctness of
Computational Explanation in Neuroscience,” Synthese, 191, 127–153.
Colombo, M., Hartmann, S. and van Iersel, R. (2014) “Models, Mechanisms, and Coherence,” The British
Journal for the Philosophy of Science, 66, 181–212.
Craver, C. (2007) Explaining the Brain, Oxford: Oxford University Press.
Craver, C.F. (2014) “The Ontic Conception of Scientific Explanation,” in M.I. Kaiser, O.R. Scholz,
D. Plenge, and A. Hütteman (eds.) Explanation in the Special Sciences: The Case of Biology and History,
Dordrecht: Springer, 27–52.
Craver, C.F. and Darden, L. (2013) In Search of Mechanisms: Discoveries across the Life Sciences, Chicago: The
Craver, C. and Tabery, J. (2015) “Mechanisms in Science,” in E.N. Zalta (ed.) The Stanford Encyclopedia of
Philosophy, https://plato.stanford.edu/.
Glennan, S.S. (1996) “Mechanisms and the Nature of Causation,” Erkenntnis, 44, 49–71.
Glennan, S. (2002) “Rethinking Mechanistic Explanation,” Philosophy of Science, 69, S342–53.
Glennan, S. (forthcoming) The New Mechanical Philosophy.
Godfrey-Smith, P. (2003) Theory and Reality: An Introduction to the Philosophy of Science, Chicago: University
of Chicago Press.
Halina, M. and Bechtel, W. (2013) “Mechanism, Conserved,” in W. Dubitzky, O. Wolkenhauer,
H. Yokota, and K.H. Cho (eds.) Encyclopedia of Systems Biology, Dordrecht: Springer, 1201–1204.
Hempel, C.G. (1965) Aspects of Scientific Explanation and Other Essays in the Philosophy of Science, New York:
Free Press.
Hempel, C.G. and Oppenheim, P. (1948) “Studies in the Logic of Explanation,” Philosophy of Science, 15,
135–175.
Illari, P. (2013) “Mechanistic Explanation: Integrating the Ontic and Epistemic,” Erkenntnis, 78, 237–255.
Illari, P.M. and Williamson, J. (2011) “Mechanisms Are Real and Local,” in P.M. Illari, F. Russo, and
J. Williamson (eds.) Causality in the Sciences, New York: Oxford University Press, 818–844.
Kaplan, D.M. and Bechtel, W. (2011) “Dynamical Models: An Alternative or Complement to Mechanistic
Explanations?” Topics in Cognitive Science, 3, 438–444.
Kaplan, D.M. and Craver, C.F. (2011) “The Explanatory Force of Dynamical and Mathematical Models
in Neuroscience: A Mechanistic Perspective,” Philosophy of Science, 78(4), 601–627.
Kellert, S.H., Longino, H.E. and Waters, C.K. (2006) “Introduction: The Pluralist Stance,” in
S.H. Kellert, H.E. Longino, and C.K. Waters (eds.) Scientific Pluralism, Minneapolis: University of
Minnesota Press, vii–xxix.
Lange, M. (2015) “On ‘Minimal Model Explanations’: A Reply to Batterman and Rice,” Philosophy of
Science, 82, 292–305.
223
Marta Halina
Levy, A. (2014) “What Was Hodgkin and Huxley’s achievement?” The British Journal for the Philosophy of
Science, 65(3), 469–492.
Levy, A. and Bechtel, W (2013) “Abstraction and the Organization of Mechanisms,” Philosophy of Science,
80(2), 241–261.
Lipton, P. (2004) Inference to the Best Explanation, London: Routledge.
Love, A.C. and Nathan, M.J. (2014) “The Idealization of Causation in Mechanistic Explanation,” paper
presented at the Philosophy of Biology at Madison Workshop, University of Wisconsin-Madison,
5 May-1 June.
Machamer, P., Darden, L., and Craver, C.F. (2000) “Thinking about Mechanisms,” Philosophy of Science,
67(1), 1–25.
Naora, H. and Montell, D.J. (2005) “Ovarian Cancer Metastasis: Integrating Insights from Disparate
Model Organisms,” Nature Reviews Cancer, 5(5), 355–366.
Ross, L. (2015) “Dynamical Models and Explanation in Neuroscience,” Philosophy of Science, 82(1), 32–54.
Ruben, D. (2012) Explaining Explanation: Second Edition, London: Routledge.
University Press.
Sheredos, B. (2015) “Re-reconciling the Epistemic and Ontic Views of Explanation (Or, Why the Ontic
View Cannot Support Norms of Generality),” Erkenntnis, 1–31. doi: 10.1007/s10670-015-9775-5.
Sober, E. (1997) “Two Outbreaks of Lawlessness in Recent Philosophy of Biology,” Philosophy of Science,
64, S458–S467.
Stepp, N., Chemero, A. and Tuvey, M.T. (2011) “Philosophy for the Rest of Cognitive Science,” Topics
in Cognitive Science, 3, 425–437.
Strevens, M. (2008) “Approaches to Explanation,” Depth: An Account of Scientific Explanation, Cambridge,
MA: Harvard University Press.
Weber, M. (2005) “Model Organisms: Of Flies and Elephants,” in his Philosophy of Experimental Biology,
Cambridge: Cambridge University Press, 154–187.
Weisberg, M. (2013) Simulation and Similarity: Using Models to Understand the World, Oxford: Oxford
University Press.
Woodward, J. (2014) “Scientific Explanation,” in E.N. Zalta (ed.) The Stanford Encyclopedia of Philosophy,
https://plato.stanford.edu/.
Wright, C.D. (2012) “Mechanistic Explanation without the Ontic Conception,” European Journal for
Philosophy of Science, 2(3), 375–394.
224
17
MODELS OF MECHANISMS1
John Matthewson
1. Introduction
Mechanisms and models are two of the most intensively studied areas in recent philosophy of
science, and there are clear points of connection between these fields. However, apart from
some notable examples (for instance Glennan 2005; Craver 2006; Darden 2007; Levy and
Bechtel 2013), there has been surprisingly little work on modeling in mechanistic sciences
(see also Kaplan 2011: 346). As noted by Stuart Glennan over a decade ago, contemporary
mechanistic philosophy has not been overly concerned with issues regarding modeling:
Perhaps because of the realist tendencies of the philosophers involved, most of the literature
has focused on the properties of mechanisms themselves and has not said much about the
relationships between mechanisms and their models or theoretical representations.
(2005: 443–4)
In fact, even when theoretical representation in mechanistic science is directly addressed, it is still
not obvious how this work relates to the literature regarding scientific modeling. For example,
mechanistic philosophy often considers the representational roles of mechanism sketches,
mechanism schemas, and their instantiations. While these clearly relate to scientific models as
discussed in philosophy of science more broadly, exactly how they connect to one another is
not entirely certain.2
Similarly, the modeling literature has tended to neglect the centrality of mechanistic models
in many branches of science, usually focusing instead on highly theoretical mathematical mod-
els. These lacunae are unfortunate, and represent genuine missed opportunities. Thinking about
models casts light on the practice of mechanistic science, and thinking about how mechanisms
are modeled is an essential component of understanding scientific models in general.
The next two sections of this chapter will describe some common and important topics in
the philosophical literature regarding scientific modeling. The final two sections illustrate how
these topics arise in the study of mechanistic science specifically. These latter sections are not
intended to present a definitive list of the relevant issues, or a definitive treatment of those issues
I do discuss. Hopefully, however, they will give a feel for the landscape where models and
mechanisms interact, and where further work is required.
225
John Matthewson
2. Some clarifications
Scientific models have been subject to philosophical scrutiny for the last fifty years or so, and the
extent of this scrutiny continues to grow. This attention has resulted in an increasing number of
involved and subtle positions, each with their own distinctions and terminology. Additionally,
although much of the work regarding models is closely interconnected, subtle differences in the
use of the word “model” can sometimes obscure important differences regarding what is actu-
ally under discussion. It is therefore important to clarify that our focus in this chapter is not the
role models play in the structure of scientific theories as discussed in the semantic account, but
the way models are used to represent mechanisms.3 So the primary sense of “scientific model” in
this context refers to the representations employed by scientists to describe, predict, or explain
some system of interest (Downes 1992; Godfrey-Smith 2014: 19).
However, it is also worth noting a related but different philosophical project involving
models, more concerned with the practice of modeling in science. Here investigators adopt an
intentionally indirect methodology, investigating the system or phenomenon of interest by cre-
ating and exploring a model of that system. These models will usually be much simpler than the
system they stand in for. For example, biologists interested in population growth may consider
a “model population,” where immigration and emigration are ignored, and size increases con-
tinuously rather than discretely. Nevertheless, this model population can then be used to make
inferences about the growth of real populations on the basis of similarities between the model
and actual populations in the world (Godfrey-Smith 2006; Weisberg 2007, 2013). There is a
great deal of overlap between the study of models as representations and this practice of indirect
modeling, and I will use insights from both of these traditions in what follows.
3. Models and scientific representation

There is a vast range of models used in scientific theorizing. Philosophical discussion often
focuses on mathematical models described by equations and often depicted as trajectories
through a state-space, or groups of entities, each acting according to a set of simple rules
and represented on computer screens as changing patterns on a lattice. However, science also
employs models that are totally different to these theoretical representations: living fruit flies and
particular species of flowering plants, for example, or physically constructed miniature canals
and boats (Sterrett 2002). Still other models are simply imagined scenarios, sketched out in a
way akin to philosophical thought experiments (Godfrey-Smith 2006).
As pointed out by Stephen Downes (1992, 2009), in the face of such diversity the chal-
lenge is to find a middle ground between the claim that models are one single specific type
of thing (which is clearly false) and allowing anything at all to function as a model (which
would tell us nothing useful or interesting about scientific models). One way to approach
this is to consider the role these models all play. A quite standard place to turn to here is the
work of Ronald Giere (1988, 2004). Giere points out that the descriptions used in science
often do not appear to be directly about the subject of interest. Scientific discussion is often
concerned with entities such as pendulums operating in perfectly uniform gravity, instanta-
neously growing populations, or perfectly informed consumers. These aren’t descriptions of
any actual pendulums, populations, or consumers. Rather, they are descriptions of models of
these entities. Giere argues that these models are useful to scientists to the extent that they are
relevantly similar to the phenomenon under consideration. An understanding of the model
and the ways in which it is similar to this phenomenon allows scientists to articulate, predict,
or explain that particular part of the world.
226
Models of mechanisms
Figure 17.1 Diagram of the relations between description, model, and target, modified from Giere
(1988: 83)
Note that according to this account, we distinguish between models and their descriptions.
Models are not the equations or diagrams used to describe them any more than any other
object can be equated with its description. Scientific modeling therefore involves three
relata: the description of the model, the model itself, and the part of the world under inves-
tigation (the “target system”). The description is true of the model, while the model is
connected to the target system via a relation of similarity (see Figure 17.1).
I will focus on three prominent issues regarding this framework: first, the ways in which
models can be abstract and idealized—how they diverge from being complete and veridical
representations of their targets; second, the relation of similarity that is meant to hold between
model and target system; third, the ontology and individuation of different types of models,
especially in the case of models that are mathematical or highly abstract. These three issues are
key topics in the philosophy of modeling generally, but each of them will also prove important
when we go on to consider mechanistic models in particular.
There are many different ways of understanding the terms “idealization” and “abstraction”
(for example, see Cartwright 1989; Leonelli 2008). Here I will follow Peter Godfrey-Smith’s
framework, as it draws a distinction that will be important for some of the discussion that follows
in later sections (Godfrey-Smith 2006; Love and Nathan forthcoming).
For Godfrey-Smith, idealization occurs when the model represents something known to
be false of the target system, while abstraction involves the omission of certain aspects of that
system. On this way of understanding the terms, only some representations are idealized, while
essentially every representation must be abstract to some extent, given the impossibility of artic-
ulating every aspect of a target.
An example Godfrey-Smith uses to illustrate the difference is a model population where the
investigators are only concerned with the density of its members, and so information regarding
the population’s actual size is not included. This is an abstract representation of some popula-
tion, because it expresses only truths (let’s say), but not all of the truths regarding that population.
This differs from a model in population genetics that is stipulated to be infinitely large to ignore
drift effects. Here we have a description that is false of any actual population, and is known by
any scientist using it to be false. It is therefore a case of idealization.
Idealizations can be distinguished further, between what Michael Weisberg calls “Galilean”
and “minimal” idealization (2007a, 2013: 98–103). In the former case, simplifications are
introduced into the model for pragmatic reasons such as computational tractability or ease of
use. In the latter, idealization is used in the service of explanatory generality. By removing
227
John Matthewson
some details, a model may gain generality, and thereby—on some accounts at least—gain
explanatory power.
Weisberg notes that these two strategies will often generate the same models: minimizing the
details in a model may make it both more general and easier to use. However, we can still differ-
entiate these kinds of idealization according to their ultimate goals. Galilean idealization is usually
something to be removed over time, as scientists and their technology become better able to
manage complexity. In contrast, minimal idealizations may remain even as the science progresses,
as this type of idealization is made for in-principle reasons. On this view, generality can be a
desirable feature of the model, regardless of whether a more complex model would be tractable.4
Now we turn to consider similarity. As noted multiple times in the literature, it is not very
illuminating to simply claim that representations or models must be similar to their targets
(Goodman 1972; Godfrey-Smith 2006: 733). Rather, we have to understand the ways and
extent to which a model must be similar to its target to successfully represent that part of the
world. To complicate issues, these demands on similarity may vary depending on the phenom-
ena studied and the purpose of the model.
To help clarify, I will again draw on concepts developed by Michael Weisberg (2007b,
2013), although my terminology is somewhat different. One way a model can be similar to its
target is in terms of its behavior. In this sense, a model is similar to the phenomenon of interest
to the extent that their outputs match when given the same inputs or initial conditions. A model
that is similar to its target in this way will be useful for describing and predicting the behavior
of the system it represents.
However, it is important to note that the underlying structure of such a model will not
necessarily reflect any of the real structural elements or dependencies in the target system. We
can see this in the setting of “phenomenological” models, where the model simply codifies cor-
relations between the values of certain variables, with no commitment to the structure of the
model matching anything particular in the world (Craver 2006). For example, we might con-
struct a perfectly serviceable phenomenological model that connects the probability of a local
storm to barometer readings. Such a model will allow us to make defeasible predictions about
the weather, while making no claims that it resembles what actually underlies the probability of
storm occurrence.
Alternatively, one might construct a model that includes cloud cover, air pressures, ocean
temperature etc., and maps the interactions of these features to generate an output regarding
the probability of a storm in a particular location. To the extent that this model’s parts and their
interactions match the actual processes involved in the production of local weather conditions,
this model will thereby be similar to its target with respect to those underlying processes.
If a scientist is only interested in predicting the outputs of some target, then a purely behav-
iorally similar model may be exactly what they are after. There is no reason for the scientist to
care about a match between the underlying structures of model and target, as long as the outputs
are correct. However, if the modeler’s objective is to understand what actually underlies that
behavior, then the model will need to be structurally similar to its target in at least some impor-
tant respects. As we will see, this type of similarity is central when we consider the adequacy
conditions for mechanistic modeling.
It is also important to note that similarity to target might not be the modeler’s sole objective.
For example, a model may be intended to represent or unify a broad class of phenomena rather
than a single system. Depending on the complexity and heterogeneity of these phenomena, a
trade-off between generality and similarity may force scientists to be satisfied with a limited match
between their model and any specific target (Levins 1966; Odenbaugh 2003; Matthewson and
Weisberg 2009; Matthewson 2011).
228
The upshot of the foregoing is that the manner and extent to which a model should resem-
ble its target will vary according to the task at hand. However, as pointed out by Susan Sterrett
(2006: 72), this doesn’t make the relevant similarities subjective, just contextual. Given a specific
setting and goal, a scientist will have particular similarity requirements that the model needs to
meet to be adequate for the task.
I now turn to the ontology of models. This is relatively clear-cut when the model is a physi-
cal object, such as a scale model of a boat or building. However, things are more problematic
when we consider models that are not concretely instantiated in the world, such as frictionless
projectiles or infinitely large populations. I will consider two prominent views regarding the
ontology of these theoretical models: first that they are mathematical objects, and second that
they should be treated as imaginary or hypothetical entities akin to literary fictions.
Many theoretical models are described by equations that relate the values of different vari-
ables and are depicted as (sets of) trajectories through a state-space. The values of these variables
represent certain properties of the model system. Each variable can be ascribed a dimension in
a space, so a full specification of all the model’s variables can be described by a point in this
space. One of the variables will often stand for time, and variable values will evolve over time to
trace a trajectory through the space. According to some views regarding non-physical models,
the model just is this mathematical object: a set of trajectories through an n-dimensional space,
described by a collection of equations, and corresponding to a set of possible inputs (for more
on this, see Weisberg 2004, 2007b, 2013; Odenbaugh 2008).
At least some modelers do appear to think of their models in this way at least some of the
time. However, we might wonder how well this fits with what Martin Thomson-Jones (2006)
calls the “face-value practice” of modeling: how scientists generally act and speak about these
parts of their day-to-day work. Peter Godfrey-Smith (2006: 734–5) points out that the “folk
ontology” of scientists (the term is from Deena Skolnick Weisberg) often tends to treat models
more like concrete entities; bone fide populations or cell gates, for example, rather than a set of
trajectories through an abstract multidimensional space.
Godfrey-Smith argues that for this and other reasons, such theoretical models are bet-
ter thought of as “imagined concrete” objects (2006: 104, 734). These models might not
physically exist (or at least are assumed to not exist), but if they did exist, they would be
concrete entities.
There is a great deal that can be said about this, and it is certainly a controversial topic (see,
for example, Giere 2009; Fine 2009; Frigg 2009). For example, the “imagined concrete” view
arguably captures the richness and flexibility of the model-target similarity relation better than
the “mathematical object” view. Unfortunately, it also appears to generate further problems
of its own, such as how scientists could use a purely imaginary object to learn something new
about the world (Godfrey-Smith 2009). However, rather than exploring the ontology of mod-
els further here, we will now turn to consider how this and the other issues discussed above
manifest in the setting of mechanistic science, beginning with the question of what counts as a
distinctively mechanistic, or mechanical model.
4. Models in mechanistic science

In the paper “Modeling Mechanisms,” Stuart Glennan presents the view that “A mechanical
model is (not surprisingly) a model of a mechanism” (Glennan 2005: 445). This is an intuitive
starting point: what makes a model mechanical is that it represents a mechanism. However, at
least two important issues immediately arise. First—as Glennan goes on to argue—for a model
to truly be “about” a mechanism, it cannot merely mirror the mechanism’s behavior, but must
229
John Matthewson
also represent the mechanism itself (i.e. its parts, activities, and their organization). This central
idea is discussed further below, when we turn to mechanistic representation.
Second, although the ontology of models is controversial, views based on Giere’s work gen-
erally agree that models are entities in their own right with their own properties. This raises the
possibility that models might be mechanistic (or not) in and of themselves, regardless of their
target (see also Craver and Tabery 2015: section 3.3).
An example used in (Matthewson and Calcott 2011) is the Meccano model of a VW
gearbox designed and built by Alan Wenbourne (2006). The Wenbourne model is itself a
mechanism, made of metal struts and fastenings, organized in such a way that they interact to
produce a specific output. It would seem that the label “mechanistic model” is appropriate
here in at least some sense, independent of what the model is ultimately used to represent. In
this case, when we claim the model is mechanical or mechanistic, we might be saying some-
thing about the target of the model, but we might also be saying something about the model
itself. So to avoid possible ambiguity or misunderstanding when discussing “mechanistic
models,” models of mechanisms should not be conflated with models that are mechanisms.
(For further discussion regarding this way of understanding the ontology of mechanistic
models, see Matthewson and Calcott 2011.)
Here the earlier discussion regarding model ontology becomes important. A model that is
mechanistic “in its own right” must be able to exhibit the properties of a mechanism: it must
be composed of identifiable, separable interacting parts that are organized in a way that brings
about some output behavior. Although this is quite intuitive in the case of a physical model like
the Wenbourne gearbox, it seems we should allow for non-physical models to be mechanistic in
this sense also. After all, the mechanistic sciences aren’t restricted to the use of concrete models.
Unfortunately, how a purely theoretical model can possess the properties of a mechanism is
not entirely clear. One way to resolve this would be to adopt one of the views regarding theo-
retical models we considered in the previous section: that such models should be considered
hypothetical concrete objects. On this understanding, were a mechanistic model to be realized
concretely, it would be a mechanism in the same way actual concrete models can be mechanisms.
For example, a gearbox does not need to be modeled with pieces of metal. One might instead
describe the model with a series of equations articulating the various interactions involved, or
draw a diagram of such a mechanism. In this case, the model so described isn’t a concrete object,
simply because it has not been realized concretely. However, if it were to be realized, it would
be a mechanism.
In discussions of mechanistic modeling, however, authors are usually more concerned with
the issue of how models can be mechanistic in Glennan’s sense above: what does it take for a
model to successfully represent a mechanism?
Recall that a model can be similar to its target with respect to behavior or underlying struc-
ture. In the case of mechanistic representation, the model must be capable of exhibiting both
kinds of similarity, again as noted by Glennan:
A mechanical model consists of (i) a description of the mechanism’s behavior (the

behavioral description); and (ii) a description of the mechanism that accounts for that
behavior (the mechanical description).
(2005: 446)
A purely phenomenological model reflects some dependencies that produce the target’s behav-
ior, but only those that involve external inputs or initial conditions. Whether such a model can
be explanatory at all is controversial, but it is uncontroversial that such models cannot provide
230
mechanistic explanations. To represent and explain mechanistically is to show how a mecha-

nism produces the phenomenon of interest in virtue of its parts and their organized interactions.
This requires a model that is not only similar to its target with respect to its behavior, but also
with respect to its structure.
Carl Craver and David Kaplan have made similar points (Craver 2006; Kaplan 2011; Kaplan
and Craver 2011). In recent work, they call this the “model-mechanism-mapping” constraint,
and explicitly note how this might be understood in the setting of mathematical models:
A model of a target phenomenon explains that phenomenon to the extent that (a) the
variables in the model correspond to identifiable components, activities, and organi-
zational features of the target mechanism that produces, maintains, or underlies the
phenomenon, and (b) the (perhaps mathematical) dependencies posited among these
(perhaps mathematical) variables in the model correspond to causal relations among
the components of the target mechanism.
(Kaplan 2011: 347; see also Kaplan and Craver 2011)
This all looks reasonably straightforward, details aside, and a great deal of mechanistic modeling
can be understood in this way. However, there are of course complications. Up to now we
have been talking as though the distinction between pure behavioral similarity and structural
similarity is always clear-cut. This is not the case. We might ask when a model is “merely”
phenomenological, and when it might be, or become, a truly mechanistic model. For example,
debates regarding the famous Hodgkin–Huxley model of axon depolarization are at least in part
concerned with when and how a model based on data might be confirmed in the right ways to
represent the underlying mechanism (Weber 2008; Craver 2008; Levy 2013).
A further potential complication involves exactly what aspects of structural similarity must be
met to explain mechanistically. For example, similarity of mechanistic structure might require
an accurate spatiotemporal match between the parts of both model and target. However, struc-
tural similarity might also mean something less constrained, such as preserving just the causal
mapping between parts and their interactions, or perhaps something even more abstract than
this (see for example Illari and Williamson 2010).
A favorite example of mine here is the MONIAC: a model that uses fluid movements
through valves and plastic tubing to represent a national economy. This is a mechanistic model
in the sense that it is itself a mechanism, but furthermore, it is thought to explain the economy’s
behavior on the basis of its causal structure. The MONIAC’s parts, such as idle balances, sav-
ings, and revenue, combine via their interconnections to show how certain economic outcomes
might arise (Phillips 2000; Barr 2000).
Here is a case that has the hallmarks of mechanistic explanation—parts, interactions, and
organization—and there are at least some abstract structural similarities between target and
model. However, this mapping is not at all straightforward. For example, the MONIAC is
made up of parts that are discrete and spatially separate, entirely unlike these “parts” in a real
economy. Spatial separation in the MONIAC is used to represent causal independence in the
model, but this comes at the cost of fidelity to the target in just these spatial respects. Indeed,
it is not entirely clear what being spatially discrete would mean in economic terms. In turn,
this means the structures of model and target are dissimilar in some ways that at least intuitively
might be thought essential to mechanistic explanation. Nevertheless, the MONIAC does appear
to explain its target by way of its internal structure (see also Matthewson and Calcott 2011).
This issue is potentially even more marked in the setting of mathematical models. Here,
certain aspects of mechanistic structure may not even be applicable, where part-hood and
231
John Matthewson
organization may be expressed through the modularity of, and interactions between, the
equations that express the model.
Such mismatches regarding spatial discreteness and location might appear to constitute a sig-
nificant problem for representation in mechanistic sciences. However, it is at least not obvious
that mechanistic representation and explanation require that the parts in a model are spatiotem-
porally similar to those of its target. For sure, spatiotemporal features will be essential in at least
some instances of mechanistic explanation, but we have seen that similarity requirements can
vary according to context, and mechanistic explanation of at least a kind seems possible when
abstract causal structure is all that is similar between model and target. (For examples of this idea
in the setting of systems biology, see Chapter 27.)
In summary, although it is correct that successful mechanistic representation requires struc-
tural similarity between model and target system, different instances of mechanistic explanation
may have different requirements regarding what this comes to. This means a full picture of
mechanistic modeling will be more complicated (and possibly more varied) than just stating that
the model must be structurally similar to its target. With this in mind, we now turn to consider
the role of abstraction and idealization in mechanistic science.
Abstraction is certainly present in mechanistic models, for the reasons discussed above, and
at least some prominent mechanists are conscious of the presence and possible utility of abstrac-
tion in scientific work. For example, Lindley Darden has discussed abstraction in science in a
number of settings, including the relationship between abstraction and generality (1996), and
how abstract representations can function as parts of theories by subsuming various cases. These
abstract representations are then to be instantiated with the relevant details to explain particular
phenomena (2006: chapter 10).
However, given this, the general position in the philosophical literature appears to be that
abstraction is not in itself a desirable aspect of mechanistic models (Machamer, Darden, and
Craver 2000; Darden and Craver 2002), at least as far as explanation is concerned. Rather, it
seems to be more a practical constraint or present for reasons of convenience; not something to
be valued for its own sake.
This idea has recently been challenged by Arnon Levy and Bill Bechtel, who argue that
abstraction allows for certain explanatory benefits, and so sometimes “less can be more” (Levy
and Bechtel 2013: 241). For example, emphasizing causal structure at the expense of details
regarding the specific parts of a mechanism may highlight features that truly make a difference to
the phenomenon of interest. Levy and Bechtel’s position is more in keeping with the literature
regarding modeling and explanation more generally, where abstraction is often seen as a positive
goal in at least some settings (e.g. Levins 1966; Wimsatt 1987; Strevens 2004, 2008).
This debate is only just beginning to be addressed in the context of mechanistic models,
but clarifying this type of issue is exactly the kind of insight that philosophy of modeling may
bring to the philosophy of mechanistic science. Might the inclusion of further detail sometimes
actually reduce the representational adequacy of these models? Regardless of the final answer
to this question, the outcome will be of interest to all parties. If mechanistic models are always
improved by the inclusion of more detail when possible and practical, this is of relevance to
arguments regarding the proper objectives of modeling (Levins 1966; Orzack and Sober 1993;
Odenbaugh 2003; Potochnik 2007). On the other hand, if increased detail is sometimes a bad
feature of mechanistic models, then this finding will substantially add to discussion regarding
representation in the mechanistic sciences.
Idealization in mechanistic models has received even less attention than abstraction, but
again the standard position appears to be that it is something generally to be avoided. On this
view, the more that a mechanistic model diverges from the structure of its target, the worse that
232
model will be. There may be pragmatic or pedagogical reasons to produce simplified models,
but this is paying an epistemic price for an increase in ease of use, rather than a positive aspect
of the model per se.
This position has also recently been questioned. In the paper “The idealization of causation
in mechanistic explanation,” Alan Love and Marco Nathan argue that mechanistic models
are sometimes altered through judicious simplification to improve explanatory power. This
can sometimes even include misrepresenting features known to be genuine difference-makers
for the phenomenon of interest. For example, Love and Nathan note that many standard
mechanistic explanations of gene expression represent the process in extremely simplified
ways, omitting various necessary co-factors and converting multiple steps involving multiple
molecules into a series of simple, unitary causal processes. Nevertheless, they maintain (and
the scientific literature appears to agree) that these idealized representations are successfully
explanatory (Love and Nathan forthcoming).
In fact, it is not difficult to find examples of quite markedly idealized mechanistic models.
Furthermore, this idealization can occur in fields usually thought to lend themselves to a
straightforward “machine-like” representation of the target system (Levy 2014). For example,
many cellular phenomena are modeled as proceeding through a series of steps from start-up
conditions to an end point, with each entity playing a specific role in the process. However,
cellular behaviors such as protein synthesis and signaling pathways are often quite noisy “biased
random walks” (Moore 2012: 8), and the relationships between cellular entities and their roles
are often many–many, in ways not reflected in standard models (see, for example, Raser and
O’Shea 2005; Viney and Reece 2013; Moore 2012; Zhao et al. 2009). This does not necessarily
mean that a close match between model and target isn’t usually desirable in mechanistic science.
However, idealization is clearly present in many mechanistic models, and merits more attention
than it has received so far.
5. Relevance for contemporary debates in the mechanisms literature

We have seen how the study of mechanistic models intersects with a number of concerns in the
broader modeling literature. I will now outline a couple of examples to briefly illustrate how
the consideration of modeling may speak to some core issues in mechanistic philosophy. First I
will address the question of what counts as a mechanism in the setting of natural selection, and
then the issue of whether mechanistic explanation is ontic or epistemic.
In section 4 of this chapter I argued that there can be successful instances of mechanistic
modeling where the model and target are dissimilar in certain ways, including their spati-
otemporal organization. One example of this occurs when scientists construct mechanistic
models of populations. In such cases, although the parts and interactions which underlie certain
population-level behaviors might arguably occur at the level of individual members, it may
be possible to mechanistically model these behaviors using population-level properties. The
earlier case of the MONIAC is one instance of this, but it is by no means the only one.
This type of scenario is interesting for a number of reasons, not least because it may have a
bearing on the limits of mechanistic representation. For example, in the article “Thinking about
Evolutionary Mechanisms: Natural Selection” (2005), Robert Skipper and Roberta Millstein
address the question of whether natural selection can be represented as a mechanism in a way
that fits standard philosophical accounts. They point out that scientists call natural selection a
mechanism, and engage with natural selection as though it is a mechanism in much of their
work. However, the then standard accounts of mechanisms (Machamer, Darden, and Craver
2000; Glennan 2002) seem unable to accommodate this, since—among other reasons—the
233
John Matthewson
entities involved and their interactions do not exhibit the right kinds of stability and regularity
to qualify. This concern has since led to considerable discussion regarding whether natural selec-
tion can legitimately be considered a mechanism (e.g. Barros 2008; Havstad 2011; Levy 2012).
Our earlier points regarding how mechanistic models might misrepresent the discreteness
and spatiotemporal structure of their target may be relevant to this issue. The “mechanism” of
natural selection might be represented at the level of the population, where the effects of the
environment, breeding, and so on can be modeled as though they occur discretely in space and
time, washing out messy details of the actual ongoing individual events that underlie them.
Even though such a model misrepresents certain aspects of the process of natural selection, it
is certainly not a phenomenological model. Thinking of natural selection as a population-level
mechanism still explains the process in a way that relies on organization and causal structure (see
also Illari and Williamson 2010).
Here we see the importance of recognizing that similarity requirements can vary according
to scientific purpose. For mechanistic explanation to be possible, there must be structural simi-
larity between model and target. However, as argued above, at least sometimes this similarity
may elide many of the spatiotemporal details. In this way, Skipper and Millstein can be right
that natural selection is not strictly mechanistic as outlined in standard views (I do not make
any claims either way here, but see Chapter 22 for more on this), while it is still the case that
for some purposes, at least, scientists can legitimately think of and describe natural selection as a
mechanism, and thereby gain some explanatory insights.
Another prominent internal disagreement within mechanistic philosophy is whether expla-
nation is “ontic” or “epistemic.” Proponents of the ontic view claim that “mechanisms in
the world” explain phenomena, while their opponents instead see explanation as dependent
on epistemic artifacts such as scientific representations (Machamer, Darden, and Craver 2000;
Craver 2007; Bechtel and Abrahamsen 2005; Bechtel 2006).
It is not absolutely clear what this disagreement turns on (see Illari 2013 and Chapter 16,
this volume, for further discussion), but one way to frame the issue is as an argument regarding
priority. Everyone can agree that both mechanisms-in-the-world and their representations are
required for mechanistic explanation, but there is still the question of how these are related. For
example, if the mechanism itself is key to explanation, then the role of scientific representation
will be merely to articulate that mechanism. On the other hand, if the epistemic artifact is cen-
tral, then there may be cases where successful explanatory representations articulate something
different to, or other than, the relevant mechanism.
Now the prior discussion regarding abstraction and idealization becomes important. Recall
that mechanistic models are always at least somewhat abstract, and at least sometimes idealized.
It might seem that these findings immediately militate in favor of the epistemic view. However,
the issue is subtler than this, and the reasons that underlie such abstraction and idealization are
relevant to the debate. For example, if the abstraction is only intended to omit non-essential
features to make the relevant mechanism more transparent, this is certainly consistent with the
ontic view. However, if abstraction occurs in the service of greater unification or generality,
even at the expense of certain relevant mechanistic details, this would fit more neatly within an
epistemic framework.
Even some kinds of idealization may be consistent with an ontic approach. If idealization in
mechanistic science is generally Galilean (and so simply to serve pragmatic concerns such as trac-
tability), this is perfectly in keeping with a mechanism-first account of explanation. Conversely,
minimal idealization seems genuinely inconsistent with the ontic account. If mechanistic models
can be improved through intentional misrepresentation of the mechanism for in-principle rea-
sons, this finding would be in favor of the epistemic view (Weisberg 2013: 102).
234
For example, if Love and Nathan are correct, scientists consider at least some models of gene
expression to be superior when they represent the relevant mechanism in ways known to be
misleading. Intentionally including falsehoods for explanatory gain seems primarily an epistemic
exercise, and it is not obvious how such an approach could be accommodated by the ontic
view. The more that mechanistic models are thought to be improved by omitting parts of the
relevant mechanism, or by misrepresenting it, the less convincing it seems that the mechanism
itself is the primary driver of a model’s explanatory force.
Once again, these points are not intended to decide this extremely complex issue, and the
arguments will turn on careful examination of particular cases of mechanistic representation.
However, it is hopefully clear how consideration of concerns embedded in the philosophy of
modeling can provide traction regarding key problems in the philosophy of mechanisms. These
examples represent only part of the work that has been done here, let alone the work that
remains to be done. The intersection of modeling and mechanisms will undoubtedly prove to
be an area of increasingly important work in philosophy of science.
Notes
1 My sincere gratitude to Stuart Glennan and Phyllis Illari for their patience, advice, and support.
2 For example, in a recent article Dana Matthiessen treats schemas as at least akin to scientific models
(Matthiessen 2015), while Lindley Darden has noted that any of these representations of mechanisms
might be treated as models in a general sense (Darden 2007: 145). To be fair, difficulties in establishing a
clear or consistent view here are likely to be at least partly due to the various ways the term “model” is
used in the philosophical literature.
3 According to the semantic account, theories are sets of models, where these models are interpretations of,
or structures that satisfy, the sentences of the prior “syntactic” account identified with scientific theory
(Suppe 1977; van Fraassen 1980; Odenbaugh 2008). As a potential contrast to the semantic view, Lindley
Darden has suggested that theories may be characterized as sets of mechanism schemas (2007: 142).
4 Weisberg also identifies a third type of idealization he calls “multiple model” idealization, which I won’t
consider here. However, this kind of idealization has been discussed in the specific setting of mechanistic
models (Love and Nathan forthcoming).
References
Barr, Nicholas. 2000. “The History of the Phillips Machine.” In A. W. H. Phillips: Collected Works in
Contemporary Perspective, edited by R. Leeson, 89–114. Cambridge University Press.
Barros, D. Benjamin. 2008. “Natural Selection as a Mechanism.” Philosophy of Science 75 (3): 306–22.
Bechtel, William. 2006. Discovering Cell Mechanisms: The Creation of Modern Cell Biology. Cambridge
University Press.
Bechtel, William, and Adele Abrahamsen. 2005. “Explanation: A Mechanist Alternative.” Studies in History
and Philosophy of Biological and Biomedical Sciences 36 (2): 421–41.
Cartwright, Nancy. 1989. Nature’s Capacities and Their Measurement. Oxford University Press.
Craver, Carl. 2006. “When Mechanistic Models Explain.” Synthese 153: 355–76.
——. 2007. Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience. Oxford University Press.
——. 2008. “Physical Law and Mechanistic Explanation in the Hodgkin and Huxley Model of the Action
Potential.” Philosophy of Science 75 (5): 1022–33.
Craver, Carl, and James Tabery. 2015. “Mechanisms in Science.” In The Stanford Encyclopedia of Philosophy,
edited by Edward N. Zalta, Winter 2015. http://plato.stanford.edu/archives/win2015/entries/science-
mechanisms/.
Darden, Lindley. 1996. “Generalizations in Biology.” Studies in History and Philosophy of Science Part A 27 (3):
409–19.
——. 2006. Reasoning in Biological Discoveries: Essays on Mechanisms, Interfield Relations, and Anomaly
Resolution. Cambridge University Press.
——. 2007. “Mechanisms and Models.” In The Cambridge Companion to the Philosophy of Biology, edited by
David L. Hull and Michael Ruse, 139–59. Cambridge University Press.
235
John Matthewson
Darden, Lindley, and Carl Craver. 2002. “Strategies in the Interfield Discovery of the Mechanism of
Protein Synthesis.” Studies in History and Philosophy of Science Part C 33 (1): 1–28.
Downes, Steven. 1992. “The Importance of Models in Theorizing: A Deflationary Semantic View.”
Proceedings of the Philosophy of Science Association 1: 142–53.
——. 2009. “Models, Pictures, and Unified Accounts of Representation: Lessons from Aesthetics for
Philosophy of Science.” Perspectives on Science 17: 417–28.
Fine, Arthur. 2009. “Science Fictions: Comment on Godfrey-Smith.” Philosophical Studies 143: 117–25.
Frigg, Roman. 2009. “Models and Fiction.” Synthese 172 (2): 251–68.
Giere, Ronald. 1988. Explaining Science: A Cognitive Approach. University of Chicago Press.
——. 2004. “How Models Are Used to Represent Reality.” Philosophy of Science 71 (5): 742–52.
——. 2009. “Why Scientific Models Should Not Be Regarded as Works of Fiction.” In Fictions in Science:
Philosophical Essays on Modeling and Idealization, edited by Mauricio Suárez, 248–58. Routledge.
Glennan, Stuart. 2002. “Rethinking Mechanistic Explanation.” Proceedings of the Philosophy of Science
Association 3: 342–53.
——. 2005. “Modeling Mechanisms.” Studies in History and Philosophy of Biological and Biomedical Sciences
36 (2): 443–64.
Godfrey-Smith, Peter. 2006. “The Strategy of Model Based Science.” Biology and Philosophy 21: 725–40.
——. 2009. “Models and Fictions in Science.” Philosophical Studies 143: 101–16.
——. 2014. Philosophy of Biology. Princeton University Press.
Goodman, Nelson. 1972. Problems and Projects. Bobbs-Merrill.
Havstad, Joyce C. 2011. “Problems for Natural Selection as a Mechanism.” Philosophy of Science 78 (3):
512–23.
Illari, Phyllis. 2013. “Mechanistic Explanation: Integrating the Ontic and Epistemic.” Erkenntnis 78 (2):
237–55.
Illari, Phyllis, and Jon Williamson. 2010. “Function and Organization: Comparing the Mechanisms of
Protein Synthesis and Natural Selection.” Studies in History and Philosophy of Biological and Biomedical
Sciences 41 (3): 279–91.
Kaplan, David Michael. 2011. “Explanation and Description in Computational Neuroscience.” Synthese
183 (3): 339–73.
Kaplan, David Michael, and Carl Craver. 2011. “The Explanatory Force of Dynamical and Mathematical
Models in Neuroscience: A Mechanistic Perspective.” Philosophy of Science 78 (4): 601–27.
Leonelli, Sabina. 2008. “Performing Abstraction: Two Ways of Modelling Arabidopsis Thaliana.” Biology
and Philosophy 23: 509–28.
Levins, Richard. 1966. “The Strategy of Model Building in Population Biology.” American Scientist 54 (4):
421–31.
Levy, Arnon. 2012. “Three Kinds of New Mechanism.” Biology & Philosophy 28 (1): 99–114.
——. 2013. “What Was Hodgkin and Huxley’s Achievement?” The British Journal for the Philosophy of
Science 65: 469–92.
——. 2014. “Machine-Likeness and Explanation by Decomposition.” Philosophers’ Imprint 14 (6): 1–15.
Levy, Arnon, and William Bechtel. 2013. “Abstraction and the Organization of Mechanisms.” Philosophy
of Science 80 (2): 241–61.
Love, Alan, and Marco Nathan. Forthcoming. “The Idealization of Causation in Mechanistic Explanation.”
Philosophy of Science.
Machamer, Peter, Lindley Darden, and Carl Craver. 2000. “Thinking about Mechanisms.” Philosophy of
Science 67 (1): 1–25.
Matthewson, John. 2011. “Trade-Offs in Model-Building: A More Target-Oriented Approach.” Studies
in History and Philosophy of Science 42: 324–33.
Matthewson, John, and Brett Calcott. 2011. “Mechanistic Models of Population-Level Phenomena.”
Biology and Philosophy 26: 737–56.
Matthewson, John, and Michael Weisberg. 2009. “The Structure of Tradeoffs in Model Building.”
Synthese 170: 169–90.
Matthiessen, Dana. 2015. “Mechanistic Explanation in Systems Biology: Cellular Networks.” The British
Journal for the Philosophy of Science, doi:10.1093/bjps/axv011.
Moore, Peter B. 2012. “How Should We Think about the Ribosome?” Annual Review of Biophysics 41
(1): 1–19.
Odenbaugh, Jay. 2003. “Complex Systems, Trade-Offs and Mathematical Modeling: Richard Levins’
‘Strategy of Model Building in Population Biology’ Revisited.” Philosophy of Science 70: 1496–1507.
236
——. 2008. “Models.” In Blackwell Companion to the Philosophy of Biology, edited by S. Sarkar and
A. Plutynski, 506–24. Blackwell.
Orzack, Stephen, and Elliot Sober. 1993. “A Critical Assessment of Levins’s ‘the Strategy of Model
Building in Population Biology’ (1966).” Quarterly Review of Biology 68 (4): 533–46.
Phillips, A. W. 2000. “Mechanical Models in Economic Dynamics.” In A. W. H. Phillips: Collected Works
in Contemporary Perspective, edited by R. Leeson, 68–88. Cambridge University Press.
Potochnik, Angela. 2007. “Optimality Modeling and Explanatory Generality.” Philosophy of Science 74 (5):
680–91.
Raser, Jonathan, and Erin O’Shea. 2005. “Noise in Gene Expression: Origins, Consequences, and
Control.” Science 309 (5743): 2010–13.
Skipper, Robert, and Roberta Millstein. 2005. “Thinking about Evolutionary Mechanisms: Natural
Selection.” Studies in History and Philosophy of Biological and Biomedical Sciences 36 (2): 327–47.
Sterrett, Susan. 2002. “Physical Models and Fundamental Laws: Using One Piece of the World to Tell
about Another.” Mind and Society 3: 51–66.
——. 2006. “Models of Machines and Models of Phenomena.” International Studies in the Philosophy of
Science 20: 69–80.
Strevens, Michael. 2004. “The Causal and Unification Approaches to Explanation Unified—Causally.”
Nous 38 (1): 154–76.
——. 2008. Depth: An Account of Scientific Explanation. Harvard University Press.
Suppe, Frederick. 1977. “The Search for Philosophic Understanding of Scientific Theories.” In The
Structure of Scientific Theories, edited by Frederick Suppe, 1–241. Urbana–Champaign, IL: University
of Illinois Press.
Thomson-Jones, Martin. 2006. “Models and the Semantic View.” Philosophy of Science 73: 524–35.
van Fraassen, Bas. 1980. The Scientific Image. Oxford University Press.
Viney, Mark, and Sarah E. Reece. 2013. “Adaptive Noise.” Proceedings of the Royal Society of London B:
Biological Sciences 280 (1767), doi:10.1098/rspb.2013.1104.
Weber, Marcel. 2008. “Causes without Mechanisms: Experimental Regularities, Physical Laws, and
Neuroscientific Explanation.” Philosophy of Science 75 (5): 995–1007.
Weisberg, Michael. 2004. “Qualitative Theory and Chemical Explanation.” Philosophy of Science 71 (5):
1071–81.
Weisberg, Michael. 2007a. “Three Kinds of Idealization.” The Journal of Philosophy 104 (12): 639–59.
——. 2007b. “Who is a Modeler?” The British Journal for the Philosophy of Science 58: 207–33.
——. 2013. Simulation and Similarity: Using Models to Understand the World. Oxford University Press.
Wenbourne, Alan. 2006. “Computer Controlled DSG Transmission—South East London Meccano
Club.” Accessed January 13, 2016. http://www.selmec.org.uk/article_0004_computer_controlled_
dsg_transmission.aspx.
Wimsatt, William. 1987. “False Models as a Means to Truer Theories.” In Neutral Models in Biology, edited
by M. Nitecki and A. Hoffmann, 23–55. Oxford University Press.
Zhao, Ping, Lu Yang, Jamie A. Lopez, Junmei Fan, James G. Burchfield, Li Bai, Wanjin Hong, Tao Xu,
and David E. James. 2009. “Variations in the Requirement for v-SNAREs in GLUT4 Trafficking in
Adipocytes.” Journal of Cell Science 122 (19): 3472–80.
237
18
EXPLAINING VISUALLY USING
MECHANISM DIAGRAMS
Adele Abrahamsen, Benjamin Sheredos,
and William Bechtel
Scientists are prolific purveyors of diagrams and other visual representations. Their publications
are replete with them; their lab meetings are organized around them. With the mecha-
nistic turn in philosophy of science, many of the New Mechanists and Social Scientific
Mechanists assembled in this handbook under the broad “minimal mechanism” umbrella
(Chapter 1; Illari and Williamson 2012; Glennan in press) have incorporated some of
these as figures within their otherwise text-heavy case studies. Several have devised their
own diagrams as well, to help make sense of the science (see Chapters 9, 13, and 14). Our
own chapter therefore could have aimed to explore the visualizations used in the physical
sciences, social sciences, and philosophy of science in all of their diversity. Instead, we
have chosen to focus on the one type of visual representation that most directly captures
the machine metaphor at the heart of the researchers’ endeavor, which we call mechanism
diagrams. These have deep historical roots (which are unearthed as far back as Descartes’
biology; see Chapter 3).
In constructing a mechanism diagram, a scientist lays out the parts of a proposed mecha-
nism and the operations (activities) they perform so as to highlight spatiotemporal organization
(e.g., Figure 13.1 in Chapter 13). Our focus on mechanism diagrams reflects an epistemic,
rather than ontological, construal of mechanism, beginning with what we call basic mecha-
nistic explanation (relying on strategies of decomposition and localization) and progressing
to dynamic mechanistic explanation (which adds computational modeling of the mechanism’s
dynamics to its otherwise merely ordinally specified temporal organization; see Bechtel and
Abrahamsen 2012). Recently we have dubbed these Mechanism 1.0 and 1.1 and argued for
a Mechanism 2.0 in which stable parts and other strictures of the machine metaphor can be
abandoned (Levy and Bechtel forthcoming). Since most scientists’ diagrams reside comfort-
ably within Mechanism 1.0, however, most of this chapter focuses on basic mechanistic
diagrams (sections 1–5). It is bookended with consideration of a broader range of diagrams
and other visual representations: first, a brief review of work on the functions diagrams serve
for scientists and others (section 1); later, an annotated mechanism diagram linking a basic
mechanism to a computational model of its dynamics (section 7); and finally, visual represen-
tations even further afield—such as those displaying data from experiments, or causal relations
between variables—which potentially may contribute to mechanistic understanding despite
not portraying parts and operations (section 8).
238
Explaining visually using mechanism diagrams
1. Functions served by diagrams

The key function of diagrams, shared by almost of their specific uses, is to provide an external aid
that supports and expands the reach of mental work. (An exception is the archival function, but
even that preserves the potential for diagrams to aid future mental work.) More specific functions
include the use of a diagram by an individual (or interactively by a group) as an aid in discovering
or modifying a scientific account, solving a problem, and understanding a particular mechanistic
explanation or other sort of content. Mechanism diagrams in particular not only support concep-
tualizing and reasoning about a proposed mechanism but also help with relating the proposal to
existing explanations, considering the evidence for new components, trying out a reorganization
of existing components, and so forth (see Sheredos and Bechtel in press-b). For example, after
adding an arrow for a recently discovered molecular interaction to a diagram, a scientist may stare
and sketch for some time considering what further changes to the previous explanation this entails.
Diagrams can also be used for communication, either one-to-one or one-to-many; for
example, an organizational chart may be handed to a new executive or emailed to all. For
scientists, if a journal article includes a mechanism diagram it typically is the first figure (raising
a question) or last figure (displaying a mechanism that fits the findings ) or, increasingly, serves
as a graphical abstract (capturing the crux of the contribution). Multiple mechanism diagrams
may be included in an integrative review or oral presentation. When instead the communica-
tion is from teacher to student—pedagogy—the diagram may have a different design that takes
into account students’ limited knowledge of both content and diagram conventions in the field
under study (Tversky 2005) and congruity with the desired mental representation.
A few philosophers of science have offered insightful inquiries into scientists’ various uses of
diagrams, especially for discovery. Construing discovery as an extended, iterative process, Darden
and colleagues have introduced mechanism schemas and the related notions of mechanism sketches
and schema types. Their case studies of the discovery of particular biological mechanisms, many
of which are brought together in the book by Craver and Darden (2013), include some of the
mechanism diagrams scientists have used to convey proposed schemas. In Chapter 19 of this
handbook, Darden offers a very clear précis (sans diagrams). Another window into the discovery
process is provided by unpublished diagrams produced over the course of a single investigation
(Burnston, Sheredos, Abrahamsen, and Bechtel 2014). Incorporating both published and unpub-
lished diagrams, and arguing that they form a coupled system with mental models, Nersessian
(2008) contributed a penetrating account of Maxwell’s development of his theory of electro-
magnetism. Moving forward from discovery, Perini (2005) emphasized the final stage in which
scientists include diagrams and other visual representations in their publications to support their
claims. She later compared three visual formats used by biologists (pictorial, schematic, and com-
positional) with respect to the different ways each related form to content. She pointed out that
this relationship was least transparent for the compositional format, as illustrated by a basic mecha-
nistic diagram of ATP synthesis in which wedge shapes represented the (not wedge-shaped)
enzyme ATPase (Perini 2013). Finally, Woody (2014) explored yet another use of diagrams—as
problem-solving aids—by focusing on the periodic table and other chemistry diagrams.
These philosophical inquiries tie into the much larger literature on diagrams in cognitive
science in which experimenters present research participants with tasks involving diagrams. In
a pioneering study, Larkin and Simon (1987) had participants perform problem-solving tasks
in which the same information was provided in different formats. Inferences were made more
effectively with diagrams than with sentential representations because, they argued, diagrams
enabled faster search and the use of perceptual recognition. One of the most ambitious later
studies of this type (Cheng 2011) demonstrated superior learning by students completing a
239
Abrahamsen, Sheredos, and Bechtel
two-session mini-curriculum in which they learned to use probability space diagrams to solve
probability problems. In a review Hegarty (2011) noted that visual-spatial displays, such as the
weather maps promoting inference in one of her own studies, support cognitive activities by
organizing information spatially, storing it externally, and enabling it to be processed visu-
ally. Tversky is another cognitive scientist providing astute theoretical roadmaps in addition to
empirical studies of diagrams (e.g., Tversky 2005, 2011).
2. A simple diagram of a proposed circadian mechanism

We employ as exemplars of mechanism diagrams those published in a single field, circadian biology,
to show the variety of options available even when the potential content is the same. Daily oscilla-
tions in organisms’ behaviors, such as plants folding and unfolding their leaves, have been observed
since ancient times. It was not until the twentieth century, however, that researchers established that
these rhythms are internally generated but entrainable to local day–night cycles.
The rise of molecular biology in the mid-twentieth century provided the tools to start
uncovering the molecular mechanisms responsible for a wide range of phenomena, including
behavioral rhythms. The first circadian gene, per (period), was identified in the 1970s in fruit flies.
Like most genes, it is transcribed into mRNA, which is transported into the cell’s cytoplasm
and translated there into a protein. But for per, concentrations of these products oscillate on
a 24-hour cycle (with per mRNA about 8 hours ahead of the PER protein). To explain these
oscillations, Hardin, Hall, and Rosbash (1990) proposed a negative feedback mechanism.
To make their explanation accessible to nonbiologists, we designed a mechanism diagram
showing the parts, operations, and functional organization of the simplest version of this molec-
ular clock. (Their own working diagram contains alternative loops and a number of question
marks—a good example of Darden’s mechanism sketches and useful for supporting the scien-
tists’ own reasoning, but too complicated and idiosyncratic for the purpose here.) As shown
in Figure 18.1, after PER molecules have been produced, they are transported back into the
nucleus and inhibit further expression of per (though less so as they degrade, and later found
to require the help of additional components). The result of this negative feedback is a cycle
in which increases in PER lead to increasing inhibition, which leads to decreases in PER and
hence less inhibition, which leads to increases in PER once again—a new turn of the cycle.
If the specific molecular labels are removed, the remaining generic labels and arrows consti-
tute a schema type (see Darden’s discussion in Chapter 19)—the simplest version of the large
family of transcription-translation feedback loops exemplified by the per loop. While Hardin, Hall,
and Rosbash proposed this feedback loop to explain circadian rhythms in fruit flies (Drosophila),
its basic organization is remarkably well conserved. There is a similar mechanism within the
neurons of the suprachiasmatic nucleus (SCN) in mouse brains, as we will see in the mechanism
diagrams in sections 4–6. These have more loops, but two of them involve homologs of per
(labeled as mPer1 and mPer2 in Figure 18.2 and—omitting the optional “m” for mouse—as
Per1/2 in Figure 18.3). For all mammals, including humans, the SCN relies on these molecular
loops to serve its function as a central clock that adapts its timing to input from the retina and
sends output signals to synchronize the cycles of other cells throughout the body.
Figure 18.1 illustrates some common graphical conventions: regular arrows for most opera-
tions, a flat-edged arrow for inhibitory operations (negative feedback), and an irregular arrow
for degradation. The additional interacting loops uncovered since 1990 inspired graphical inno-
vations that provide a variety of ways to fit multiple feedback loops into a single mechanism
diagram. We will encounter some of these in sections 4–6, but begin by sorting through the
toolbox of components from which both simple and complex diagrams are constructed.
240
transcription transport translation

gene mRNA mRNA protein
inhibition
degradation
transport
protein
degradation nucleus cytoplasm
Figure 18.1 A simple mechanism diagram showing the key parts and operations in the transcription-
translation feedback loop proposed by Hardin, Hall, and Rosbash (1990) to account for
circadian rhythms. (The roles of additional interacting loops were discovered later.)
3. Components of mechanism diagrams

Most mechanism diagrams are composed of several different kinds of elements. Some of the
most powerful are in a category that Tversky (2011) calls glyphs. These are simple elements
that are “prevalent across a wide range of graphics” and are “readily understood in context”;
cytoplasm
nucleus
Clk
?
+ Bmal1
mCry
+
P mPer1
C B
mPer2
mPer3
−
Figure 18.2 A diagram of the molecular clock mechanism for circadian rhythms in mammals,
exemplifying several types of elements used in mechanism diagrams. From Reppert and
Weaver (2001), figure 1. Reproduced with permission of Annual Review of Physiology,
Volume 63  2001 by Annual Reviews, http://www.annualreviews.org. See plate 1
241
moreover, “Like words in language, [they] can be combined in various ways to create
varying meanings. Like words in language, there are constraints on how they can be combined”
(Tversky 2005, p. 141). In the following subsections we first describe the glyphs that are most
prevalent in mechanism diagrams: shapes, lines, and arrows. We note how color can combine
with these, and then describe three more specialized types of elements (text, iconic symbols,
and graphic symbols). We introduce these by discussing the choices made by Reppert and
Weaver (2001) in their diagram of the mammalian circadian clock mechanism (Figure 18.2;
see plate 1). We also point out alternative choices made by others in Figures 18.3–18.5 (more
fully discussed in sections 4–6) and note one additional type of element—embedded data
graphs—within Figure 18.4. Finally, what we call a conventional complex is a multi-element
building block seen in Figures 18.2–18.5.
Shapes
Different types of parts in a mechanism typically are represented by different geometric shapes,
which sometimes are labeled or color-coded to further distinguish them. At the center and
upper right of Figure 18.2 are several oblong boxes, each labeled in color with the abbrevi-
ated name of the gene it represents (mCry, mPer1, etc.). Elsewhere, shapes in matching colors
represent proteins produced from these genes: gold-colored diamonds for the protein mCRY,
blue ovals for mPER1, red rectangles for mPER2, and light-green hexagons for mPER3. The
two slightly larger ovals with the letters C and B in them represent the proteins CLOCK and
BMAL1. Thus, two different ways of using shapes were chosen here for the same basic task of
representing proteins: distinctive shapes with redundant color-coding for the four central pro-
teins vs. a plain shape with a distinctive initial letter for two other proteins. (An additional shape,
the small gray circle, represents a hypothesized molecule.) This diagram also illustrates how the
positioning of shapes can convey relationships between entities: the ovals labeled C and B touch
each other to indicate that CLOCK and BMAL1 have dimerized (formed a two-part complex).
Lines and arrows

Lines and arrows can have a variety of interpretations, depending on context or stipulation.
Among those for arrows are simple temporal sequencing, abstract causal relationships, and
transformations without change of location. However, the arrows in this chapter have several
different more specific interpretations, requiring viewers to keep track of which arrows have
which interpretation—even within a single diagram. This is easy for the single feedback loop in
Figure 18.1, because each arrow represents one operation and is labeled. However, most dia-
grams of circadian mechanisms include multiple loops, pushing designers toward more com-
pact representations of each loop. Figure 18.2 makes demands on viewers by mixing different
strategies in representing its six loops. All of them incorporate the convention of a small arrow
bent at a 90° angle next to a gene to represent its transcription. For Clk and Bmal1, the sub-
sequent parts and operations in their loops are combined into a single glyph—a large, angled
green arrow—directed at the C and B ovals that represent the resulting CLK:BMAL1 dimer.
The green arrow simplifies Figure 18.2 visually, but makes demands on the viewer’s back-
ground knowledge—a frequently encountered tradeoff. Here, the viewer needs to know not
only the parts and operations involved but also that the CLK:BMAL1 dimer, having been trans-
ported to the nucleus, activates expression of the four central genes.
A different strategy was used to depict the central genes’ loops. The four genes are bundled
together by two vertical lines so they can share the thick black arrow representing the two
242
operations following transcription—transport and translation of mRNA—as well as the curved

line from P representing phosphorylation of the resulting proteins. A single vertical line bundles
together the color-coded glyphs for the proteins, which are arranged to indicate the complexes
they form (dimers and trimers). The fact that some tokens of each type of complex are trans-
ported back to the nucleus, where they inhibit the activation activity of the CLK:BMAL1
dimers, is separately represented by the long red arrow with a minus sign (an alternative to the
flat-edged inhibitory arrows in Figures 18.1, 18.3, and 18.5). Other tokens return to the nucleus
for more specialized operations only partly understood in 2001, as indicated by the two arrows
accompanied by copies of the relevant complexes and question marks.
To make the use of lines and arrows in Figure 18.2 understandable, we have alluded to the
use of color and such elements as labels, question marks, and minus signs. These are among the
elements discussed in the rest of this section.
Colors
Color can interact with glyphs in a variety of ways that aid understanding. Often one shape rep-
resents a class of objects (e.g., ovals for proteins), and color is used to make distinctions within
the class. For example, in Figure 18.3 (see plate 2), proteins and polypeptides are represented
by circles that are colored gold for PER, green for CRY, and violet for VIP. Reppert and
Weaver took a different approach: each protein in Figure 18.2 has its own particular shape as
well as color. This visual redundancy facilitates tracking each protein and linking it to its gene
(which is labeled in the same color). For example, focusing on gold takes a viewer from the gene
mCry to the protein CRY and then to its various interactions with other proteins. Yet another
way of color-coding glyphs is showcased in Figure 18.5, where color identifies functionally
similar components of the circadian mechanism not only in the same organism but also across
species (discussed in section 6; see plate 4). As for lines and arrows, they are usually black, but
Figure 18.2 (as well as Figure 18.5) illustrates the convention that a green arrow indicates activa-
tion and a red arrow inhibition. This helps viewers see the overall organization of the mecha-
nism as combining positive and negative feedback.
Text
Although diagrams are quintessentially visuospatial, rather than linguistic, most incorporate textual
elements. At one extreme, Figure 18.1 makes extensive use of text and only limited use of space
and glyphs. Abbreviations alone identify parts (except for the oval around “PER”), and arrows are
labeled to identify operations. At the other extreme, Figure 18.2 has color-coded shapes depicting
parts (adding text labels only where necessary) and unlabeled arrows. Both figures have an enclosed
nucleus with word labels distinguishing nucleus from cytoplasm. A final use of text in Figure 18.1,
prompted by our pedagogical aims in crafting it, is for classifying the parts as gene, mRNA, or
protein. Various other styles of incorporating text can be seen in Figures 18.3–18.6, ranging from
single characters to multiword phrases and (in the original articles) lengthy captions.
Iconic symbols (icons)

As an alternative to text, iconicity is a powerful tool for conveying meaning with visual imme-
diacy. An iconic symbol (icon) resembles its referent in certain respects, and its meaning is
specific and stable across contexts. Designers can draw upon a variety of conventional icons or
invent their own.
243
Every transcription arrow in Figure 18.2 points toward a conventional icon: either a wavy
line indicating that the gene’s transcription activity exhibits circadian rhythmicity or (for Clk
only) a straight line indicating continuous transcription. Figure 18.4 (discussed in section 5; see
plate 3) shows one variation on another convention: the use of a large 24-hour clock face to
mark timing for the rest of the diagram. (A common alternative involves two small icons: the
sun marking daytime vs. the moon marking nighttime.) Figure 18.4 also has good examples of
less conventional icons: line drawings of mice and humans asleep or awake, with the obvious
meanings. More than purely aesthetic features, they enable quick interpretation of the diagram
without having to consult textual labels or figure captions.
Graphic symbols
Certain text elements, by taking on a life of their own, have acquired the impact of iconic
symbols despite their noniconicity. Figure 18.2 includes four examples of single characters that,
at least for biologists, are visually compelling and immediately interpretable: + for activation, -
for inhibition, P for the phosphate groups that bind with circadian proteins, and ? for gaps in
knowledge. The + and - symbols have stable core meanings (positive vs. negative), but their
more specific meanings depend on context (for example, in arithmetic, rating scales, chemistry,
or electricity). The phosphate symbol (P) is familiar on its own or inside a small circle (and in
Figure 18.4, tiny circles are used without P). Finally, the question mark (?) occurs with sur-
prising frequency in mechanism diagrams and is perhaps the most interesting of the graphic
symbols. Those in Figure 18.2 indicate uncertainty about the next operations performed by
known proteins (bottom) or an unidentified, hypothesized protein (top left). In Figure 18.5 they
mark proteins and operations that are unspecified, hypothesized, or for which evidence is weak
(with readers left to consult the text for the precise import of each question mark). Occasionally
a diagram with question marks will be the first figure in the article, setting the questions to
be addressed by the research that follows. Yet another function of question marks is to aid
ongoing reasoning, which is especially evident in working diagrams produced throughout the
research process and sometimes even published (e.g., that of Hardin, Hall, and Rosbash 1990, as
mentioned in section 2).
Embedded data graphs

We have distinguished mechanism diagrams from data graphs, but occasionally data graphs
are inserted into mechanism diagrams. One purpose is to display quantitative evidence about
the operation of a given part. Burnston (2016) argues that such graphs play an explanatory
role complementary to the basic mechanistic proposal.) Another purpose is to characterize
the phenomenon explained by the mechanistic proposal. For example, in Figure 18.4 (see
plate 3), the two thumbnail plots show that SCN neural firing is less intense in nighttime
than daytime—one of the phenomena to be explained by the intracellular mechanism in the
center of the diagram. When graphs are abstracted even further, they may give rise to iconic
symbols; for example, line graphs of transcription activity across 24 hours inspired the wavy
and straight-line icons in Figure 18.2.
Conventional complexes
Diagrams are created by flexibly combining different kinds of elements. Efficiency trumps
flexibility, though, when certain combinations of elements become standardized and take on a
244
life of their own. Consider the boxes for genes and the lines and bent arrows that accompany
them in Figure 18.2. Molecular biology students learn to see these arrangements (and vari-
ations on them in other figures) as skeletal depictions of the spatial layout of different func-
tional units along a DNA strand. For each gene in Figure 18.2, the box is the coding region
and the horizontal line is noncoding DNA (including the promoter regions—not specifically
depicted here, but distinguished by labeled glyphs in Figure 18.3 and thickened line segments
in Figure 18.5). The bent arrow’s placement on the line loosely represents the transcription
start site between the promoter and coding regions. Since this whole arrangement has a spe-
cific, stable meaning in diagrams throughout molecular biology, we are inclined to count it as
a conventional complex rather than as an arrangement of glyphs—and given its spatiality, one
that is somewhat iconic. The complex does retain some degree of flexibility in which shapes
and labels do or do not get added to the horizontal line (and for Figure 18.2, the addition of
wavy vs. straight-line icons as well).
4. Space in mechanism diagrams: representing

physical vs. functional organization
Since a diagram is laid out in space and a mechanism occupies space, it is natural to use the two-
dimensional space of a diagram to situate parts of the proposed mechanism in a way that approx-
imates its spatial organization. In Figure 18.2, for example, in addition to the one-dimensional
use of space in complexes representing genes, a well-delineated region representing cytoplasm
surrounds a smaller region representing the nucleus. Though simplified, the spatial relation of
enclosure in the diagram corresponds to that in a cell. The various molecules are placed as appro-
priate in either the nucleus or cytoplasm, and their movements between these two regions are
depicted by arrows. All of these design decisions use space as space. Within the two regions of
the diagram space, however, the molecular parts and operations are laid out to satisfy nonspatial
objectives, such as clearly conveying their functional organization into feedback loops, grouping
together parts that behave similarly, and avoiding crossed arrows. In the nucleus, for example,
two sets of genes play different roles: the two peripheral genes produce proteins that mediate the
auto-inhibitory loop of the four central genes. Grouping their glyphs in different locations, and
using shared arrows where possible for each group’s operations, makes this functional organiza-
tion clear. Likewise, the glyphs and arrows for the four types of proteins produced by the central
genes are the minimum needed to show the major operation they have in common (red inhibi-
tory arrow) as well as their distinctive operations. Finally, there is no attempt to show that in
the physical cell, numerous proteins of each type are distributed through much of the physical
space of the cytoplasm and nucleus. These last few design decisions show how space can be used
to clearly convey the functional (rather than spatial) organization of the mechanism. Once again
there is a tradeoff: the clarity gained by showing types rather than tokens vs. the risk of down-
playing the stochasticity of molecular mechanisms (see Chapter 13—especially its mechanism
diagram, Figure 13.1, which shows multiple tokens of the component types).
When a diagram is used to represent activities in multiple spaces, it is often necessary to sim-
plify what is shown in each space. For example, in Figure 18.3 (see plate 2) DeWoskin, Myung,
Belle, Piggins, Takumi, and Forger (2015) attempt to relate the familiar molecular mechanism
within individual cells to molecular pathways that extend to the extracellular environment.
The large ovals represent the cytoplasm (lilac) enclosing the nucleus (purple). It is notable that
only a select few of the parts and operations of the intracellular clock mechanism are shown:
the loops involving Clk and Bmal1 are absent here, as are numerous additional components
that had been discovered between 2001 and 2015. Only the Per and Cry homologs and their
245
Output
Input
VIP
Ca2+ VP
VIP AC
2
Ca2+ CREB
V Per 1/2
BCRE EBOX
Cry 1/2
(RRE) EBOX mRNA
K+
GABA
Figure 18.3 A mechanism diagram showing some of the key parts of the molecular clock mechanism
and how, in SCN neurons, they interact with pathways extending into the extracellular
environment. From DeWoskin, D., Myung, J., Belle, M. D., Piggins, H. D., Takumi, T.,
and Forger, D. B. (2015) “Distinct roles for GABA across multiple timescales in
mammalian circadian timekeeping,” Proc Natl Acad Sci USA, figure 1, 112(29): E3911-9.
See plate 2
mRNA and proteins are shown, using different conventions than Figure 18.2. Each pair of
homologs shares one minimalist representation of gene (the vertical and upper horizontal line)
and adjacent DNA (lower horizontal line); the bent arrow indicating transcription is omitted.
However, there are labels for the genes (Per 1/2 for Per1 and Per2; Cry 1/2 for Cry1 and Cry2)
and for specific promoter regions indicated by colored ovals (EBOX, CRE, and RRE). The
transcription-translation feedback loops are shown schematically: (1) The cyclic transcription of
each pair of genes is denoted by two color-coded wavy lines terminating in the word “mRNA.”
(2) Solid arrows represent transport of the mRNA into cytoplasm and translation there into pro-
teins (denoted by color-coded circles). (3) Two merging arrows indicate dimerization. (4) The
negative feedback, in which PER1:CRY1 and PER2:CRY2 dimers inhibit their own genes’
transcription, is represented by an inhibitory arrow that splits, with each branch terminating on
one of the EBOX glyphs. (As partly shown in Figures 18.2, 18.4, and 18.5, but not here, the
inhibition is mediated by removal of CLK:BMAL1 from the EBOX promoters.)
By omitting the operations through Clk and Bmal1 from their diagram, DeWoskin et al.
gained the space to show pathways for modulating Per expression that extend out to the
extracellular environment. Potassium ions (K+) play a role, but we will focus on two path-
ways that can be thought of as beginning with the entry of calcium ions into the neuron and
ending with CREB proteins enhancing Per transcription by binding to the CRE promoter
region. In the simplest case, the calcium ions directly induce phosphorylation of CREB.
The more elaborate pathway involves cell-to-cell communication via VIP. Each neuron
signals its activity by releasing VIP into the extracellular environment (with vesicle release
aided by calcium). To the extent that such signals are received in turn from other neurons
via the VPAC2 receptors, CREB is up-regulated (though the inhibitory arrow indicates it
is also subject to cyclic down-regulation). Since DeWoskin et al. show just one cell, it is up
to the viewer to imagine this mechanism replicated in numerous SCN cells, such that the
level and timing of gene expression in each cell is regulated in part by that of other cells
246
via intercellular VIP signaling. This intercellular level of organization is only hinted at here
by the words “Output” (indicating where VIP is excreted to affect other cells) and “Input”
(signifying where the VIP produced in other cells affects this one cell).
5. Representing time in static diagrams

Mechanisms function in time. This is conveyed minimally by arrows in a mechanism diagram. In
Figure 18.3, for example, the arrows from calcium to CREB and from CREB to CRE inform
us of two operations occurring in sequence. However, there is no indication of the duration of
each operation or its onset and offset times. Do they overlap, occur in immediate sequence, or
does one end hours before the other begins? Also, are there changes in their rate or intensity?
For circadian rhythms, timing is critical: the oscillations generated by the transcription-translation
feedback loop must have a period of approximately 24 hours. Given the speed with which most
chemical reactions occur, delays must be built into the molecular clock. None of this is conveyed
in the figures above. Animations provide one solution: beyond using space as space, they can also
CT6
SCN firing
z z zzz
z
mRNA
CB
Anticipated drawn
Anticipated dusk
rip
nsc tio
op ra
na
T
CT0 CB CB CT12
l fe
ed
back lo
CB
protein
SCN firing
CT18
Current Biology
Figure 18.4 Using a 24-hour clock face to show the timing of behavioral and neural activity (gray
region) and the underlying molecular clock (inner circle). Appropriately located iconic
symbols contrast the activity of humans (diurnal) with that of mice (nocturnal). Reprinted
from Current Biology, Volume 18, Hastings, M. H., Maywood, E. S., and O’Neill, J. S.,
Cellular circadian pacemaking and the role of cytosolic rhythms, figure 2,  (2008), with
permission from Elsevier. See plate 3
247
use time as time—more or less veridically, and more or less sped up or slowed down. Circadian
animations are well worth viewing, but for printed books and articles circadian researchers have
introduced some apt strategies for representing time indirectly (and perhaps more effectively;
see Tversky 2011, p. 526). For example, some figures lay out separate subdiagrams for the same
mechanism in different phases, arranged around a circle representing a 24-hour cycle. (A similar
strategy produces circular data graphs, which are among the specialized circadian data graphics
discussed by Bechtel, Burnston, Sheredos, and Abrahamsen 2014.)
The circular displays in mechanism or data diagrams have their roots in familiar physical
clocks with the numbers 1–12 placed in a circle, but instead use a 24-hour clock face. When
researchers place an organism into constant conditions (for animals, typically by switching from
light-dark cycles to constant darkness), time is conventionally reported in circadian time (CT).
CT0 is the time at which the organism’s internal clock should expect light (“anticipated dawn”).
In Figure 18.4 (from Hastings, Maywood, and O’Neill 2008; see plate 3), CT0 is placed at the
left side of the clock face. Labels for CT6, 12 (“anticipated dusk”), and 18 provide a timescale
for the rest of the figure, in which a gray region depicting behavioral and neural activity sur-
rounds an inner circle depicting molecular activity. The variable shading of the gray region
provides visual support for the CT labels: lightest at anticipated midday (CT6) and darkest at
anticipated midnight (CT18). Within it, diagram elements are appropriately indexed to CT6
and CT18: (a) icons show mice and humans as awake vs. asleep at opposite times, and (b) neural
firing rate is indicated via thumbnail data graphs obtainable from single-cell recording in mouse
SCN (~10 Hz top, ~1 Hz bottom).
In the center of Figure 18.4, multiple representational tools are used creatively and com-
pactly to convey information about the state of the molecular mechanism at different phases of
the circadian cycle. Two arrows, labeled “transcriptional feedback loop,” indicate the type of
molecular mechanism. The parts and operations are represented not by a single detailed dia-
gram, but rather by four minimalist versions of the conventional complex used to depict genes
and their transcription. These are indexed to circadian times 0, 6, 12, and 18 and emphasize
changes in the state of the mechanism. Each depiction uses a shared, unlabeled horizontal line
to represent all of the Per and Cry genes and adjacent DNA plus the thickness of the attached
bent arrow to represent different rates of transcription. They also include (a) labeled oval glyphs
for the proteins CLK and BMAL1, to be understood as a dimer binding to the promoter
regions for Per and Cry, and (b) small color-coded circles for the proteins PER and CRY and
Casein Kinase. The binding of PER and CRY to the dimer around CT12 (day)—indicated
by contact—is followed by their phosphorylation and consequent degradation around CT18
(night). Yet another diagrammatic technique—the positioning of crescents—is used to show
how changes in concentration of Per and Cry mRNA (tan crescent) precede those of PER and
CRY proteins (blue crescent).
Overall, once this diagram is unpacked and its various representational strategies interpreted,
it conveys considerable information about circadian phenomena (in the gray region) and the
time-locked phases of the molecular mechanism responsible for them (in the center).
6. Comparing two or more mechanisms

Diagrams provide a unique and powerful tool for comparing mechanistic accounts. In circadian
rhythm research, proposed molecular clock mechanisms are similar enough across different
species and even different biological orders that insights from research conducted on one spe-
cies has often guided research on distant species—quite a valuable shortcut for the discovery
process. For diagrams to facilitate this kind of research guidance, as well as integrative thinking
248
A KaiB C
PER TIM
per/tim
KaiC
KaiA
CYC CLK ?
? clk
KaiA KaiB KaiC
outputs
CRY1 CRY2
B FRQ D
wc-1 mcry1,2 PER1 PER3
PER2 ?
WC-1 WC-2
mper1,2,3
outputs ? ? frq
?
BMAL CLK ?
bmal
outputs
Figure 18.5 Comparison of the molecular clock mechanisms, as understood in 2001, in

(A) cyanobacteria, (B) fungi, (C) fruit flies, and (D) mice. From Harmer, S. L., Panda, S.,
and Kay, S. A. (2001), figure 2. Reproduced with permission of Annual Review of
Cell and Developmental Biology, Volume 17  2001 by Annual Reviews, http://www.
annualreviews.org. See plate 4
across the phylogenetic spectrum, consistency is needed in the choice of elements and use of
space. (To see how difficult comparisons would be otherwise, consider how differently the
Per loop was diagrammed in Figures 18.1–18.4.) Figure 18.5 (see plate 4) shows how Harmer,
Panda, and Kay (2001) used consistently designed diagrams to compare the clock mechanism in
(A) cyanobacteria, (B) fungi, (C) fruit flies, and (D) mice. At the time all of these were thought
to centrally involve transcription-translation feedback loops.
Each diagram uses the familiar horizontal line to represent each gene and its associated DNA
regions, with the line thickened for the noncoding promoter region. (Homologs and other
genes playing similar roles, such as per and tim in panel C, share a single line.) The gene(s) are
denoted by a label under the line, and transcription by the usual bent arrow. The remain-
ing gene expression operations are denoted by a color-coded dashed arrow terminating at the
glyph(s) for the resulting protein(s). These glyphs use distinctive shapes and colors to highlight
which proteins play comparable roles not only within a species (e.g., PER1,2,3 homologs in
mice) but, crucially, in different species. Most noticeable is the use of light-blue rounded rec-
tangles for the key proteins: KaiC in cyanobacteria, FRQ in fungi, the PER:TIM dimer in fruit
flies, and dimers (indicated by bidirectional arrows) of PER1,2,3 and CRY1,2 in mice. Each
of these proteins indirectly inhibits its own further production as its concentration increases,
shown visually by the red inhibitory arrow directed at one or two other proteins (labeled yel-
low ovals). The role of these other proteins—KaiA, WC-2, CYC:CLK, and BMAL1:CLK in
A–D respectively—is to activate the gene producing the key protein. This is depicted by a green
arrow pointing from the yellow ovals to smaller yellow ovals, which indicate binding of the
proteins to the gene’s promoter region. That this binding is disrupted in the inhibitory phase
must be inferred. Each panel also includes rose-colored shapes for other proteins (known or
hypothesized) and black lines for protein–protein interactions.
249
The use of color-coding is particularly important in supporting the comparison across species,
since it did not prove feasible to locate the corresponding parts of each mechanism in the same
region of the spatial layout. This was especially helpful for seeing ways in which the cyanobac-
teria clock (panel A) is similar to those of the other species. (The diagrams were less successful in
conveying how it differs.) For the other species (panels B–D), an important benefit can be seen
in the question marks. Although some simply flag issues not yet settled by research within one
species, others are motivated by exactly the kinds of inter-species comparisons Figure 18.5 aims
to facilitate. In the mammalian case (D), the arrow to the gene labeled “outputs” has no ques-
tion mark, since the researchers were confident that the genes generating the clock mechanism’s
output to other systems were directly activated by the BMAL1:CLK dimer. This motivates a
corresponding arrow in the fruit fly diagram (C) from CYC:CLK to the output genes there.
Harmer et al. (2001) made this arrow dotted (rather than solid or dashed) and appended a
question mark, inviting researchers to seek evidence of this direct link or, failing that, to find
unknown intermediaries. The text of the article must be consulted for many details and nuances,
but these diagrams uniquely serve as a visual heuristic by which researchers can offload some of
the cognitive burden of tracking similarities and differences, hence facilitating understanding as
well as discovery. The diagrams also make a theoretical contribution by highlighting mechanism
schema types held in common across disparate species but sometimes filled in differently—one
way of moving toward “unification in biology . . . through abstraction from causal details for
the purpose of identifying generic organizational patterns” (Chapter 27, p. 371).
7. Linking computational models to mechanism diagrams

Here we return to the simple per feedback loop, as understood by fruit fly researchers in the
early 1990s, to illustrate the benefits gained when computational models (discussed as well in
Chapters 17, 20, and 33) are linked to mechanism diagrams. Goldbeter (1995) proposed a com-
putational model to capture the cyclic dynamics of the per loop, using what we call a computation-
ally annotated mechanism diagram (Figure 18.6) to link variables and parameters in his model to the
mechanistic account. It is similar to Figure 18.1 in its austere reliance on arrows and text labels,
but has mathematical symbols appended. Next to each word denoting a part is a variable track-
ing its changing concentration (M, P0, P1, P2, PN). Next to each arrow denoting an operation is
a parameter. Three of the parameters directly represent the rate of the corresponding operation:
vs for the accumulation of per mRNA in cytoplasm as it is transported from the nucleus, k1 for
transport of PER2 into the nucleus, and k2 for a new proposed operation, return of some of the
PER2 back into cytoplasm. The other parameters represent enzyme actions influencing the rate
of their operations.
The computational model comprises five differential equations, one for each variable.
Running the model using parameter values that he regarded as biologically plausible, Goldbeter
demonstrated that the values of the relevant variables exhibited sustained oscillations—a success-
ful simulation of per’s rhythmicity. In the first equation, for example, oscillations in the value
of M (mRNA concentration) are obtained by subtracting a term that includes vm from a term
that includes PN and vs. (Each term also incorporates constants that would have cluttered the
diagram if shown.)
Figure 18.6 emphasizes those components of the mechanism that figure most prominently in
the five equations of the model. For example, it explicitly distinguishes the two partly reversible
phosphorylation steps that result in fully phosphorylated PER2 since these steps correspond to
two equations in the model. In contrast, gene transcription and its inhibition are represented
sparsely and somewhat idiosyncratically within the dashed rectangle, with per not even explicitly
250
per transcription nuclear PER (PN)
vs k2
k1
V1 V3
ks vd
per mRNA (M) PER0 PER1 PER2
(P0) (P1) (P2)
V2 V4
vm
Figure 18.6 Goldbeter’s computationally annotated mechanism diagram of the transcription-translation

feedback loop for per. Goldbeter, A. (1995) A model for circadian oscillations in the
Drosophila period protein (PER), Proceedings of the Royal Society of London. B: Biological
Sciences, 261, 1362, figure 1, by permission of the Royal Society
shown as a part of the mechanism. This is because only products of per have varying concentra-
tions, not the gene itself, and oscillations in its transcription are incorporated in the nonlinear
term of the equation tracking M.
Pursuing mechanistic and computational modeling in tandem produces what we call dynamic
mechanistic explanation (see Chapter 20 and Bechtel and Abrahamsen 2012, 2013). Rather than
viewing the two approaches as unrelated or as competitors, researchers can reap the benefits
of building a whole that is greater than the sum of its parts. If at least a sketch of a mechanistic
account is in place first, it can suggest candidate variables or parameters for a computational
model of the mechanism’s dynamics. In the example here, these quantified the concentrations
of parts and rates of operations; much work remained to select which to track, which to use
in which equations, which to omit, which would be variables vs. parameters, what additional
parameters to include and how, and so forth. For a given mechanism diagram, many different
computational models can be developed—but the diagram provides a starting point. It can then
be modified and annotated repeatedly, with the positioning of the mathematical symbols serv-
ing the modeler as a locality aid (Jones and Wolkenhauer 2012). On the other hand, when a
computational model comes first or very early, it can help shape an initial mechanism sketch or
schema and influence what research is undertaken to build on it.
For some purposes, such as assessing whether a mechanistic model can generate the right
dynamics, this kind of coordination is not just advantageous but essential in moving discovery
forward. Without computational modeling, mechanistic researchers rely on a strategy researched
by Hegarty (1992): mentally animating their diagrams. This is satisfactory for sequential mecha-
nisms with no feedback loops or nonlinearities, and can even yield a qualitative understanding
of how oscillations are produced in a more complex system (as we narrated for Figure 18.1).
Showing that such oscillations are sustained, however, requires the quantitative precision of
computational models like that of Goldbeter and its successors.
8. Other visual representations relevant to mechanistic explanation

Computational modeling is not the only research approach that can contribute to mechanistic
explanation. Most of the others bring their own conceptualizations and visual formats that do
not dovetail so closely with mechanism diagrams, but can be interpreted in ways that point
251
toward or elucidate an underlying mechanism. Consider Dynamical Systems Theory (DST),

which offers concepts and visualization tools for better understanding patterns in the relation
between variables over time. For example, the oscillations obtained by Goldbeter’s (1995) dif-
ferential equations can be understood as a limit cycle in state space and visualized in a phase
portrait, providing a deeper understanding of the circadian mechanism’s dynamics. (Chapter 20
further discusses mechanisms and DST; see especially Figure 20.1.)
Data graphics are a more general family of formats for the spatial representation of data, rang-
ing from generic line and bar graphs to specialized formats such as spectrograms. They comprise
most of the figures in published research articles, but only occasionally (as in our Figure 18.4) are
integrated into a mechanism diagram. As we have discussed elsewhere (Burnston et al. 2014),
although data graphics are thought of as merely describing what needs to be explained, often
they serve explanatory ends as well. This is the case when the variables in a graph correspond to
properties of certain parts and operations in a mechanistic explanation—measuring, for example,
the changes in concentration of one type of molecule in a metabolic mechanism or the rate
of one of its operations. But in his work on explanatory relations, Burnston (2016) emphasizes
relations between variables that help explain the phenomenon of interest while not linking so
simply to a mechanism diagram. Biologists might find it important, for instance, that the value
of a variable peaks at a particular time, that two variables are in phase with each other, or that
quantities vary proportionally to one another. Clearly, there is much more work for mechanists
to do on the various roles and visualizations of data.
Finally, there are a number of research strategies that focus solely on causal, predictive, or
other relations between variables. The variables might later be recognized to correspond to
properties of parts and operations in a mechanistic account, but it is the variables that are named,
perhaps manipulated, measured, analyzed, and reported. In structural equation modeling, for
example, the strength of each pairwise relation between variables is the product of the analy-
sis, often reported in a causal graph displaying the strongest links (see Chapter 26). As well,
Coleman’s boat diagram offers an interesting combination of causal and mechanistic understand-
ing (see Chapters 30 and 32).
9. Conclusion
We have suggested that as philosophers of science turn their attention to mechanistic explana-
tion, it is important to examine not only what scientists write but also the graphics they use as
tools for their own understanding and for communication. Data graphics predominate numeri-
cally and play important descriptive and explanatory roles, but in this chapter we have focused
on the special nature and advantages of mechanism diagrams. Using as exemplars several dia-
grams of the molecular mechanisms responsible for circadian rhythms, we have seen how dif-
ferent kinds of elements arranged in two-dimensional space can be used to (a) display the parts
and operations of a mechanism; (b) convey their spatial, functional, and temporal organization;
(c) facilitate comparison of related mechanisms; and (d) work in tandem with computational
modeling. These published diagrams illustrate various uses of shapes, arrows, and lines (Tversky’s
glyphs) as well as color, text, iconic symbols, graphic symbols, mathematical symbols, conven-
tional complexes, and space. These provide researchers a toolkit for representing their hypotheses
about mechanisms. However, the same resources also serve crucially as visual heuristics during the
research process. In section 6 we saw how researchers could exploit shape, color, and question
marks to aid cross-species comparisons. It is harder to access the many versions of draft diagrams
scientists produce for their own use during research or in preparing a published article, but for such
case studies see Burnston et al. (2014), Sheredos (2017), and Sheredos and Bechtel (in press-a).
252
When researchers craft diagrams in proposing or revising a mechanistic hypothesis, innova-

tion often takes the form of arranging glyphs and iconic symbols. The combinatorial potential
of these elements can be deployed to convey novel mechanistic hypotheses and to consider
components and activities in new contexts. Often the elements themselves need not change.
We have not highlighted novel discoveries or disagreements about mechanisms, but they most
definitely occurred over the 20 years covered here. Yet the basic types of glyphs in use have
been fairly stable, as have many iconic and graphic symbols.
Despite this stability in the basic toolkit of available resources, they were deployed in a vari-
ety of ways in the diagrams we discussed. Depending on the goals of the scientist producing the
diagram, different choices may be made regarding which parts and operations to show in detail,
which to merge into a single element, and which to omit. Also, stylistic preferences influence
whether parts are depicted by simple text labels, icons, or glyphs (labeled, color-coded, or not).
Depending on appetite for innovation, a particular researcher’s diagrams may tend to be con-
ventional and consistent over time or (as in Figure 18.4) offer new, pleasing solutions satisfying
ambitious goals. We would venture that just as no two scientists would write the same para-
graphs, neither would they produce the same diagrams. Yet each scientist finds his or her own
diagrams to be useful tools for developing, evaluating, and revising mechanistic explanations (see
Chapter 19). Some diagrams—such as Figures 18.1–18.6—will also, or instead, communicate
successfully with those who would not themselves produce exactly the same diagrams.
We have focused this chapter on mechanism diagrams, but conclude by noting that they
are not the only tool for developing or communicating mechanistic explanations, and mech-
anistic explanation is not the only function that diagrams or other visualizations can serve.
We have already mentioned that data graphics offer tools for displaying not only the phe-
nomenon to be explained but also data quantifying aspects of the mechanistic explanation.
We discussed how computationally annotated mechanism diagrams can serve as locality
aids for computational models, which generate predicted quantifications for comparison to
such data. We briefly discussed animation. Setting visual representations in motion has the
potential to enhance understanding of the dynamics of a mechanism, but there is much to
learn about how best to do this. For example, animation would fill in the gaps between the
four snapshot diagrams in Figure 18.4, but passive viewing may be less helpful than having
freeze-frame or speed-adjustment options. As a final example, visual tools providing access to
augmented reality and virtual reality are moving out of engineering labs toward everywhere
else, including science labs. The human drive to explore phenomena and to find, refine,
and communicate explanations will continue to foster a growing visual toolbox. We invite
our readers to go beyond our own work on diagrams to probe existing and future dynamic
visualizations of mechanistic explanations.
References
Bechtel, W., and Abrahamsen, A. (2012) “Thinking dynamically about biological mechanisms: Networks
of coupled oscillators,” Foundations of Science, 18, 707–723.
Bechtel, W., and Abrahamsen, A. (2013) “Roles of diagrams in computational modeling of mechanisms,”
Proceedings of the 35th Annual Conference of the Cognitive Science Society (pp. 1839–1844). Austin, TX:
Cognitive Science Society.
Bechtel, W., Burnston, D., Sheredos, B., and Abrahamsen, A. (2014) “Representing time in scientific
diagrams,” Proceedings of the 36th Annual Conference of the Cognitive Science Society (pp. 164–169). Austin,
TX: Cognitive Science Society.
Burnston, D. C., Sheredos, B., Abrahamsen, A., and Bechtel, W. (2014) “Scientists’ use of diagrams in
developing mechanistic explanations: A case study from chronobiology,” Pragmatics and Cognition, 22,
224–243.
253
Burnston, D. C. (2016) “Data graphs and mechanistic explanation,” Studies in History and Philosophy of
Biological and Biomedical Sciences, 57, 1–12.
Cheng, P. C.-H. (2011) “Probably good diagrams for learning: Representational epistemic recodification
of probability theory,” Topics in Cognitive Science, 3, 475–498.
Craver, C. F., and Darden, L. (2013) In Search of Mechanisms: Discoveries across the Life Sciences. Chicago, IL:
DeWoskin, D., Myung, J., Belle, M. D., Piggins, H. D., Takumi, T., and Forger, D. B. (2015) “Distinct
roles for GABA across multiple timescales in mammalian circadian timekeeping,” Proc Natl Acad Sci
USA, 112(29), E3911-9.
Glennan, S. (in press) The New Mechanical Philosophy. Oxford: Oxford University Press.
Goldbeter, A. (1995) “A model for circadian oscillations in the Drosophila period protein (PER),” Proceedings
of the Royal Society of London. B: Biological Sciences, 261, 319–324.
Hardin, P. E., Hall, J. C., and Rosbash, M. (1990) “Feedback of the Drosophila period gene product on
circadian cycling of its messenger RNA levels,” Nature, 343, 536–540.
Harmer, S. L., Panda, S., and Kay, S. A. (2001) “Molecular bases of circadian rhythms,” Annual Review of
Cell and Developmental Biology, 17, 215–253.
Hastings, M. H., Maywood, E. S., and O’Neill, J. S. (2008) “Cellular circadian pacemaking and the role
of cytosolic rhythms,” Current Biology, 18, R805–R815.
Hegarty, M. (1992) “Mental animation: Inferring motion from static displays of mechanical systems,”
Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 1084–1102.
Hegarty, M. (2011) “The cognitive science of visual-spatial displays: Implications for design,” Topics in
Cognitive Science, 3, 446–474.
Illari, P. M., and Williamson, J. (2012) “What is a mechanism? Thinking about mechanisms across the
sciences,” European Journal for Philosophy of Science, 2, 119–135.
Jones, N., and Wolkenhauer, O. (2012) “Diagrams as locality aids for explanation and model construction
in cell biology,” Biology and Philosophy, 27, 705–721.
Larkin, J. H., and Simon, H. A. (1987) “Why a diagram is (sometimes) worth ten thousand words,”
Cognitive Science, 11, 65–99.
Levy, A., and Bechtel, W. (forthcoming) “Toward mechanism 2.0. Expanding the scope of mechanistic
explanation.”
Nersessian, N. J. (2008) Creating Scientific Concepts. Cambridge, MA: MIT Press.
Perini, L. (2005) “Visual representations and confirmation,” Philosophy of Science, 72, 913–926.
Perini, L. (2013) “Diagrams in biology,” The Knowledge Engineering Review, 28, 273–286.
Reppert, S. M., and Weaver, D. R. (2001) “Molecular analyses of mammalian circadian rhythms,” Annual
Review of Physiology, 63, 647–676.
Sheredos, B. (2017) “Communicating with scientific graphics: A descriptive inquiry into non-ideal
normativity,” Studies in History and Philosophy of Science Part C: Studies in History and Philosophy of
Biological and Biomedical Sciences, 63, 32–44.
Sheredos, B., and Bechtel, W. (in press-a) “Sketching biological phenomena and mechanisms,” Topics in
Cognitive Science.
Sheredos, B., and Bechtel, W. (in press-b) “Imagining mechanisms,” in Godfrey-Smith, P. and Levy, A.
(eds.) The Scientific Imagination: Philosophical and Psychological Perspectives. Oxford: Oxford University
Press.
Tversky, B. (2005) “Prolegomenon to scientific visualizations in science education,” in J. Gilbert (ed.)
Visualization in Science Education (pp. 29–42), Dordrecht, Netherlands: Springer.
Tversky, B. (2011) “Visualizing thought,” Topics in Cognitive Science, 3, 499–535.
Woody, A. I. (2014) “Chemistry’s periodic law: Rethinking representation and explanation after the turn
to practice,” in L. Soler, S. Zwart, V. Israel-Jost, and M. Lynch (eds.) Science after the Practice Turn in
Philosophy, History, and the Social Studies of Biology. Oxford: Routledge.
254
19
STRATEGIES FOR
DISCOVERING MECHANISMS1
Lindley Darden
1. Introduction
The new mechanists have applied the mechanistic perspective to numerous philosophical topics,
such as explanation, reduction versus integration, function, and causation. Less discussed is the
topic of reasoning in discovering mechanisms. That is the topic here.
Traditional philosophers of science distinguished the logic of discovery from the logic of
justification (or falsification), arguing that there is no logic of discovery (e.g., Popper 1965).
“Friends of discovery” (Gutting 1980; Nickles 1980a, 1980b; Meheus and Nickles 2009) con-
ceded that deductive logic does not capture the nature of reasoning in discovery. Furthermore,
there is no algorithm that takes data as an input and yields an explanatory theory as an output,
as opposed to merely finding patterns in the data. Instead, the friends of discovery recast the
task; their goal is to find heuristic problem-solving strategies for scientific discovery. Their work
was influenced by Herbert Simon’s heuristic problem-solving approach (Simon 1977, 1996;
Nickles 1977; Wimsatt 2007). Darden (1991) argued that discovery of theories should be recast
as an iterative process of cycling through stages of construction, evaluation, and revision, with
heuristic reasoning strategies for each stage.
Norwood Russell Hanson was one of the few philosophers of science of the mid-twentieth
century who analyzed reasoning in hypothesis construction. Following Charles Sanders Peirce,
Hanson elaborated his own view of reasoning to a hypothesis. He might now be pleased that
reasoning in the discovery of mechanisms follows his pattern of reasoning in discovery, a very
abstract view of moving from puzzling phenomena to a hypothesis of a certain type:
Schematically, [retroductive reasoning] can be set out thus:

(1) Some surprising, astonishing phenomena p1, p2, p3 . . . are encountered.
(2) But phenomena p1, p2, p3 . . . would not be surprising were a hypothesis of H’s
type to obtain. They would follow as a matter of course from something like
H and would be explained by it.
(3) Therefore there is good reason for elaborating a hypothesis of the type of H;
for proposing it as a possible hypothesis from whose assumption phenomena
p1, p2, p3 . . . might be explained.
(Hanson 1961: 630; italics added)
255
Lindley Darden
In earlier versions, Hanson had simply discussed good reasons for elaborating a hypothesis H,
but in this 1961 paper, he instead used “type of H.” This was a step forward. Hypothesis gen-
eration and preliminary evaluation should be distinguished (Schaffner 1993: ch. 2). The ability
to connect a surprising phenomenon with a type of hypothesis is a step in generation, prior to
finding a specific plausible hypothesis. Using examples from physics, Hanson suggested a type
might be an inverse square type, in which something varies as the distance away increases. But
Hanson did not consider hypothesis types in biology and medicine. John Josephson expanded
retroduction (abduction) to include systematic generation of a search space of possible types of
hypotheses in medical diagnosis. He elaborated criteria for evaluating one as the most plausible
and instantiating it for a specific case (Josephson and Josephson 1994).
Often in biology, the type of hypothesis to be discovered is a mechanism schema, and the
search is guided and constrained by the characterization of a mechanism. The product shapes
the process of discovery. One can say, in general, what counts as an adequate description of a
mechanism; that will be the subject of the next section. Furthermore, heuristic reasoning strate-
gies can guide the construction stage; those strategies are the subjects of the following sections.
Because discovery consists of construction, evaluation, and revision, strategies for evaluation and
revision are also part of the story, but cannot be fully discussed here (see Craver and Darden
2013: chs. 6–9). Nonetheless, final sections here discuss recent work combining construction,
evaluation, and revision using computer simulations and biorobotics in mechanism discovery.
2. What is to be discovered: characterizing mechanisms

and mechanism schemas
The discovery of a mechanism typically begins with a puzzling phenomenon. (For more on
the nature of the phenomenon, see Garson, Chapter 8 in this volume.) Data provide evidence
for the phenomenon (Bogen and Woodward 1988). When the goal is to find what produces
the phenomenon, then one searches for a mechanistic type of hypothesis. That decision rules
out other parts of a large search space. One is not seeking merely a set of correlated variables.
One is not seeking an economical equation that describes the phenomenon, although such an
equation can provide a constraint in the search for a mechanism. One is not seeking a law from
which a description of the phenomenon can be derived. One is not merely seeking a rela-
tion between a cause and an effect, although such a relation provides clues about mechanism
components (Darden 2013; see Matthews and Tabery, Chapter 10 in this volume). Nor is one
merely seeking to find a pathway, characterized by nodes and unlabeled links, although that is
a very abstract way of representing some aspects of a mechanism. Rather, one is attempting to
construct a mechanism schema that describes how structural and active components are spatially
and temporally organized together to produce the phenomenon.
Employing a specific characterization of a mechanism provides guidance in discovery. One
oft-cited mechanism characterization is this: “Mechanisms are entities and activities organized
such that they are productive of regular changes from start or set up to finish or termination
conditions” (Machamer, Darden, and Craver 2000: 3). (For a more minimal characterization of
mechanisms, see Glennan and Illari, Chapter 1 in this volume.) The goal is to find the entities
and activities, to describe how they are organized together, and to show that when they are
organized together in a productively continuous way, they produce the phenomenon of inter-
est. This characterization directs one to ask: What are the setup and finish conditions? Is there
a specific, triggering start condition? What is spatially next to what? What is the temporal order
of the steps? What are the entities in the mechanism? What are their structures? What are the
activities that drive the mechanism? What are their range and their rate? What interacts with
256
Strategies for discovering mechanisms
what? How does each step of the mechanism give rise to the next? How was each step driven by
the previous one? What is the overall organization of the mechanism: does it proceed linearly or
is the mechanism perhaps cyclic (with no clear start and stop), or is it organized with feedback
loops, or does it have some other overall organizational motif? Where is it spatially located? In
what context does the mechanism operate and how is it integrated with other mechanisms?
These kinds of questions show how the product guides the process of its discovery.
Mechanism schemas are representations of mechanisms. What I call a “schema” others refer
to as mechanistic “models,” but “model” has many other uses, other than abstractly represent-
ing the structure of a target mechanism. (See Chapter 17.) This is an example of a very abstract
schema: DNARNAprotein. Such schemas are often depicted in diagrams. William Bechtel
and his collaborators (see Abrahamsen et al., Chapter 18 in this volume) discuss the many “visual
heuristics” that diagrammatic representations of mechanism play, including a gray box with
question marks to indicate gaps to be filled.
Schemas have several dimensions: sketchy to sufficiently complete, abstract to specific, small
to general scope of applicability, and possible to actual (Craver and Darden 2013: ch. 3). A goal
in discovering a mechanism is to convert an incomplete sketchy representation into an adequate
one for the purpose at hand. Incomplete sketches indicate where black (unknown components)
and gray (only functionally specified) boxes need to be filled to have a productively continu-
ous schema. During the construction phase of discovery, moving from a sketch to a sufficiently
complete schema allows one to work in a piecemeal fashion; one can work on one part of the
mechanism at a time while leaving other parts as black or gray boxes. Because one is attempting
to reveal the productive continuity of a mechanism from beginning to end, what one learns
about one step of the mechanism places constraints on what likely has come before or what
likely comes after a given step.
Abstraction comes in degrees and involves dropping details; specification involves adding
details all the way to instantiation, where an instantiation is still a representation, with sufficient
details to represent a productively continuous mechanism from beginning to end. A goal in dis-
covery is to find a schema at a given degree of abstraction, from a very abstract type of schema
with few specified components to a fully instantiated one for a particular case. The desired
degree depends on the purpose for which the mechanism is sought. Although degree of abstrac-
tion is an independent dimension from the scope of the domain to which the schema applies,
more abstract schemas (if they have any instances at all) may have a wider scope. Hence, when
the goal of the discovery process is to find a very generally applicable mechanism schema, it is
likely to be represented at a high degree of abstraction. The move from how possibly to how
plausibly to how actually is driven by applying strategies for evaluation, such as experimental
testing, and strategies for anomaly resolution, such as localizing faults and revising the schema.
Consider the example of mechanisms connecting gene to phenotype. One starts with the
beginning point, e.g., a gene or gene mutation, and some characterization of the phenotype
of interest. In Mendelian genetics, phenotypic characters are usually observable properties of
organisms, e.g., the color of peas or eyes in fruit flies. In molecular biology the phenotype might
be, e.g., the presence of a protein, the abundance of that protein, or an altered structure or
abundance of a protein (see, e.g., Nachtomy et al. 2007). For medical studies, a gene mutation
may be statistically associated with a disease phenotype. Between the gene or gene mutation and
the phenotypic character is a black box; the goal is to fill in sufficient details so as to represent
the mechanism connecting the two.
Having evidence of a beginning point (the gene or gene mutation) and the end point (the
wild type or disease phenotype), the discovery task is to fill in the black box to some degree
of detail. For example, if the goal is to knock out the gene or to replace the gene during
257
Lindley Darden
gene therapy, then it may be unnecessary to find all the intervening steps in the mechanism.
A highly abstract schema may be sufficient to guide the work to knock out or replace the gene.
However, if the goal is to design a therapy to alter an entity or activity in a mechanism step
downstream from the gene, then specific details become important: e.g., one may need to find
the three-dimensional structure of a protein and identification of its active site. Rarely is one
gene responsible for one phenotypic trait, so multiple beginning points may converge on the
way to the trait. Or one gene mutation may play a role in multiple forking downstream steps.
Or feedback loops may need to be represented. (See Abrahamsen et al., Chapter 18 in this
volume.) The purpose of the study will aid in deciding how many black boxes to fill and with
how much detail.
3. Stages in mechanism discovery

For convenience, we can divide the mechanism discovery process into at least four stages:
characterizing the phenomenon, constructing a schema, evaluating the schema, and revising the
schema (Darden 2006, ch. 12; Craver and Darden 2013, chs. 4–9). These stages are often pur-
sued in parallel and in interaction with one another. Mechanism discovery is frequently pursued
piecemeal. A scientist might work on a part of a mechanism, while leaving much else about the
mechanism inside a black box. Also, the different stages of discovery frequently interact with
one another. One is forced to recharacterize the phenomenon in the face of learning about
the mechanism (Bechtel and Richardson 1993). Or, one is forced to re-evaluate experimen-
tal findings because one recognizes a previously unrecognized region of the space of possible
mechanisms. Or, when one attempts to build a computer or biorobotic simulation, one finds
previously unidentified black boxes that need to be filled.
The first stage, characterizing the phenomenon, frames the problem to be solved in the
search for mechanisms. This stage involves developing a more or less precise description of
the puzzling phenomenon produced by the mechanism. The nature of the phenomenon often
provides clues about the kind of underlying mechanism that might possibly be responsible for it.
The second stage of mechanism discovery, constructing a schema, involves generating a space
of possible mechanisms for a given phenomenon. Construction strategies include employing a
schema type, assembling modules, and forward and backward chaining.
The third stage of discovery, evaluation, aims at revealing the empirical and conceptual
constraints on the space of possible mechanisms for a given phenomenon. The components of
a mechanism—that it is composed of parts with sizes, shapes, and structures, that it has a par-
ticular temporal sequence, that it is composed of one kind of entity and not another or makes
use of one kind of activity and not another—provide conceptual constraints by which scientists
evaluate mechanism schemas. For example, to what extent does a proposed entity have the
activity-enabling properties to carry out the following step? Empirical constraints come from
observations and experiments. An experiment involves intervening to change some part of
a mechanism or some background condition to learn something about how the mechanism
works. Other constraints come from fitting the proposed schema into a larger context so as to
cohere with other known mechanisms.
The final stage of mechanism discovery is revision. Revision is required when a favored
mechanism schema is challenged by an empirical anomaly, or does not fit coherently into
a larger context, or has some other kind of failure. Strategies for anomaly resolution result
from the range of choices that the scientist might consider in response to an anomaly for
this mechanism schema. Some of the choices concern whether the apparent challenge to the
schema can be barred from being a challenge because of, for example, a failed experiment or
258
being beyond the scope of the intended schema. When such barring fails, and the anomaly
is determined to directly challenge a proposed schema, there are a number of search strate-
gies that scientists may use to localize the site of fault within their schema and to correct it
to remove the fault.
Scientists use construction strategies to populate the space of possible mechanisms. They use
evaluation strategies to identify constraints that prune the space of possible mechanisms. They
use revision strategies to diagnose and localize the source of error in a hypothesized mechanism
and to provide clues as to how that hypothesis might fruitfully be amended.
Traditionally, “discovery” usually refers only to the construction stage, so that will be the
focus of the next sections.
4. Guidance from the phenomenon

A mechanism discovery episode typically begins with at least two sources of empirical knowl-
edge to guide the discovery process. The scientist comes to a problem with a store of knowledge
about the phenomenon under investigation and about the target system in question. First, one
begins knowing something, even something sketchy, about the phenomenon for which one
is hoping to discover a mechanism. Second, one typically begins with prior knowledge about
the kinds of entities and activities that might be involved in such phenomena in a given type of
organism or system.
Because the phenomenon provides guidance in schema construction, the more that is known
about it at the outset, the more clues it supplies. Work to characterize it is in order. One needs
to find the setup conditions under which it occurs, any triggering conditions needed to make it
come about, and inhibiting conditions that prevent its occurrence. One can investigate the phe-
nomenon by passive observation and interventions, such as staining to find relevant structures
(Kästner 2015), a technique often used to image cellular structures. Or scientists may use more
deliberate experimental manipulations to activate, enhance, or ablate hypothetical components
(Craver 2007), such as ablation in gene knockout experiments. If the phenomenon is the statisti-
cal association of a gene mutation with a disease phenotype, then all the standards for evaluating
adequate statistical sampling apply.
If the characterization of the phenomenon shows that it is of a familiar type, the field might
already have developed a library of mechanism-types through which one might search. For
example, in the search for mechanisms to connect a DNA mutation to a disease phenotype, the
many roles that DNA bases play provide a library of possible types of mechanisms, e.g., change
in a coding region producing a change in an amino acid; change in a regulatory region leading
to increased or decreased expression of the associated gene; change in a splicing region, leading
to a malformed protein; or change in a region producing a microRNA, with subsequent alter-
nation of its functional role. Standard methods exist for testing whether one of these possible
types applies to a particular phenomenon.
Raoul Gervais and Erik Weber (2015) discuss a type of experiment, which they dub an
“orientation experiment,” that aids in finding the abstract type of mechanism schema at an early
stage of the discovery process. “Orientation experiments . . . are useful in situations in which we
know (next to) nothing about the mechanism, which are often (but not exclusively) the crucial
early stages of an investigation into a phenomenon, namely the orientation-phase” (Gervais and
Weber 2015: 49). For example, suppose the puzzling phenomenon is how an ant finds its way
back to the nest. Two possible types of mechanisms are visual recognition or chemical signaling.
An orientation experiment would remove light, find that the ant still finds its way back, and
indicate the search should be for a chemical signal.
259
Lindley Darden
5. Localization
Localization is a second, often crucial, clue in the construction stage of the search for
mechanisms, as William Bechtel and Robert Richardson argue (1993, 2010). It is often far
from obvious at the beginning of a discovery episode where in the system under study the
phenomenon takes place. In some cases it is unclear whether the phenomenon even takes
place in the system under study at all, or whether, in contrast, it takes place in the interac-
tions between the system and its environment (or perhaps entirely in the environment).
For example, the phenomenon of inheritance came to be localized to hereditary materials
inside germ cells. Inheritance is the tendency of offspring to resemble their parents. In sexually
breeding species, inheritance is biparental; in other words, the child has a mixture of traits from
both parents. By the mid-nineteenth century (with the discovery of the mammalian egg), the
locus of transmission of hereditary traits was quite plausibly the sperm and the egg, because those
germ cells link the generations. However, some debate occurred as to whether influences from
a previous mating could linger in the mother’s womb.
Gregor Mendel’s work in the 1860s on patterns of inheritance, rediscovered around 1900,
provided evidence for unobservable differences among germ cells that might explain the inher-
itance of traits through several generations. Further evidence for localization of hypothesized
hereditary particles within the germ cells occurred after the discovery in the late nineteenth
century of chromosomes (thread-like bodies in the nuclei of cells) that were hypothesized to be
the hereditary material. Evidence for this localization was strengthened in the early twentieth
century when abnormalities in chromosomes were found to correspond to abnormalities in pat-
terns of inheritance of traits. Biochemists noted that chromosomes are composed of two types of
chemicals: proteins and deoxyribonucleic acid (DNA). Chemical analysis of DNA in the 1930s,
given the sensitivity of the chemical techniques at the time, seemed to indicate that it had a
simple repeating structure, whereas proteins had much more complexity. Hence in the 1930s
and 1940s, the most plausible location for the hypothesized hereditary particles, the genes, was
in the proteins of chromosomes. This view was dispelled with more accurate chemical analysis
of the DNA molecule and the discovery of its double helical structure in 1953. The idea that
the gene is a linear sequence of bases along a DNA chain localized part of the hereditary process
and, in so doing, offered an entry point into the mechanism from which one could conjecture
how genes could possibly produce hereditary traits (Darden 1991; Darden and Craver 2002).
Locating (and re-locating) the mechanism is often a crucial turning point in discovering a
mechanism. By choosing a locus for the search for mechanisms, one places the mechanism in
an appropriate context, one searches for entities and activities at that locus, one circumscribes
the experimental and observational strategies that will be useful in addressing the problem, and
one often has clues for choosing among an available set of mechanism schemas that might be
brought to bear in the given case.
6. Employ a schema type

Characterizing the phenomenon and finding a likely location for the mechanism may provide
clues for retrieving from a library one or more abstract schemas appropriate to phenomena of
that type. The abstract schema provides variables, black and gray boxes, that can be filled with
entities and activities to find the mechanism for the specific case (Darden 2002).
When the phenotype of interest is a disease phenotype, then the goal is to discover a
disease mechanism. Thagard discusses abstract schemas for representing defective molecular
260
mechanisms associated with diseases (e.g., Thagard 1998, 1999). He distinguishes abstract
schemas for diseases caused by external agents versus internal gene defects. One searches for
a different kind of mechanism if the disease is due to invading bacteria, such as pneumonia,
rather than diseases due to a single gene defect, such as cystic fibrosis. Identifying the defective
mechanism aids in designing therapies that target specific entities and activities to alleviate the
disease (Thagard 2003; Craver and Darden 2013: ch. 11).
7. Modular subassembly
Modular subassembly is the strategy of putting together known types of parts, namely
modules of entities and activities that carry out particular kinds of mechanism role func-
tion (see Chapter 8). One cobbles together different modules to construct a hypothesized
how-possibly mechanism, guided by the goal of finding modules to fill all the functionally
characterized gray boxes in a mechanism sketch. In doing so, scientists draw upon their
knowledge of types of modules that have a particular function (Darden 2002).
Evolution itself often works by copy and edit: copies of genes can be found in mutated form,
playing similar or different roles in the same or related organisms. Finding such recurrent motifs
has been a powerful tool in discovery in biology. There are various types of receptors, neuro-
transmitters, enzymes, and gene regulatory components (e.g., inducers, repressors, and other
types of transcription factors). In a given field, researchers develop a library of types of modules
that perform various functions. When a functional requirement can be specified in the mecha-
nism sketch, then appropriate types of modules can be sought to satisfy it. Experimentalists can
manipulate such modules to investigate their functional roles.
For example, developmental biologists discuss modularity as a principle in evolution: com-
ponents are individually modified or conserved in different lineages. One of the remarkable
cases of a functional module consists of the Pax6 gene, along with a group of genes related to
it. This module plays a role in the formation of eyes in invertebrates such as fruit flies, as well as
in vertebrates such as frogs and mice. This module recurs even though the eyes themselves have
quite different structures (compound eyes in flies and eyes with a single lens in vertebrates). An
amazing abnormality results from placing a mouse Pax6 gene in the leg of a fruit fly, in which
structures characteristic of the insect eye develop on the leg (Gilbert 2003: 413). Hence, if one
is seeking the mechanism of eye development in a new species, knowledge of such types of
conserved modules will aid in constructing a plausible target mechanism.
Finding a new type of module opens a new region of the space of possible mechanisms.
The 1970 discovery of the enzyme reverse transcriptase was just such a new type (Temin
and Mizutani 1970; Baltimore 1970; Darden 2006: ch. 10). This enzyme copies RNA back
into DNA:
RNA–via reverse transcriptase–>DNA
This reverse transcriptase module plays an important role in retroviral infection. Retroviruses,
such as the HIV-AIDS virus, carry this enzyme into the infected cell. Retroviruses use reverse
transcriptase to copy their own viral RNA into DNA. The DNA copy of the viral genome is
then integrated into the host genome and is copied as new cells are produced. Integration into
the host DNA makes such retroviruses very difficult to attack with drugs. The discovery of
reverse transcriptase opened up a space of possible mechanisms in which base sequences could
be copied and inserted into the DNA, a hitherto unknown activity.
261
Lindley Darden
The discovery of this new type of module expanded the space of possible mechanism
modules. Many sequences in the human genome appeared to result from reverse transcription
of viral DNA in past evolutionary history. Finding recurrent motifs in biological mechanisms
provides a growing store of modules of use in constructing how-possibly schemas.
8. Forward/backward chaining
Another strategy for constructing mechanism schemas involves first learning something about the
mechanism or one of its components and then using that knowledge to make inferences about
what came before it or what is likely to come after it (Darden 2002). In forward chaining, one
uses the early steps of a mechanism to reason about the types of entities and activities that are likely
to be found in later steps. In backward chaining, one reasons from the entities and activities in
later steps in a mechanism to find entities and activities appearing earlier. With cyclic or feedback
mechanisms, some separable step can serve as a relative starting point for reasoning about earlier or
later ones. Thus, the strategy of forward/backward chaining is likely available to scientists when
they know anything, or can conjecture anything, about the specific entities and activities anywhere
in the hypothesized mechanism, even if no libraries of schema or module types are available.
Consider some of the ways one might use forward/backward chaining. Entities engage in
activities by virtue of the fact that they have activity-enabling properties. One can often use general
knowledge about kinds of properties and their association with particular kinds of activities in
biological systems to reason forward about the activities in which the entities can or do engage.
Such activity-enabling properties include three-dimensional structure, size, location, orienta-
tion, and charge. Structures can promote or inhibit the push/pull of geometrico-mechanical
activities. Three-dimensional shapes can be open or closed, narrow or wide, exposing or con-
cealing. An open entity permits movement through it more or less as its opening is narrow
or wide. Entities may also have different kinds of charges, and molecules have valences, both
of which affect the kinds of bonding activities in which they engage. So, finding the activity-
enabling properties of entities gives clues to the kinds of activities in which they might engage
in the next step of the mechanism. Melinda Fagan (2012) calls one type of such activity-enabling
properties of entities “meshing” properties; they enable complexes of entities to form whenever
the appropriate spatio-temporal relations between them obtain.
Conversely, in backward chaining, activity signatures are a source of clues as to what came
before. When an activity operates, it produces an effect that changes the entities involved.
Noting the properties that may have been changed, the activity signatures, allows one to con-
jecture what happened in a previous step. For example, if one detects a hydrogen bond between
two entities in a complex in a later step of a mechanism, then one can reason that an earlier step
included polar molecules, with complementary weak charges that have been neutralized in the
formation of the bond. Another example is allosteric molecules. An allosteric molecule changes
shape (via stresses and strains of geometrico-mechanical activities) when it bonds to an effector
molecule. The new shape exposes a new active site, so that the allosteric molecule is enabled to
eject or bond to a third molecule. If one detects an allosteric molecule bound to an effector (an
activity signature), that indicates that effector bonding occurred in a prior step.
Forward and backward chaining both contributed to the discovery of the mechanism of
protein synthesis in the 1950s and 60s. Biochemists knew that the endpoint of the protein
synthesis mechanism was a string of amino acids held together by strong covalent bonds.
They thus reasoned back to an antecedent step in which the amino acids were free and
unbound. Because energy is required to form such strong bonds, reasoning backward suggests
the existence of a high-energy intermediate in the preceding step. Biochemists isolated such a
262
high-energy intermediate. Thus, the strong covalent bonds were activity signatures, indicating
the components needed in the preceding steps to form them. Surprisingly, the activated
amino acid was associated with RNA. A typical biochemical reaction schema to synthesize a
molecule has the separate chemical components on one side and the newly synthesized mol-
ecule on the other. Accordingly, biochemists expected amino acids and an energy source to
be required for synthesizing a protein (a chain of amino acids). However, such a schema had
no role for RNA because RNA is not a component of proteins. Reasoning backward from
protein to free amino acids did not suggest an RNA intermediate in the chemical reaction.
Meanwhile, molecular biologists were reasoning forward from the DNA double helix to
the next step in the protein synthesis mechanism. Biochemists and cell biologists had discov-
ered that RNA was involved in the mechanism. Molecular biologists suggested that RNA
acted as a template that would guide the assembly of the protein. The order of the bases in
the coding strand of DNA would be transcribed into similarly ordered bases in RNA. The
RNA then serves as the template that orders the amino acids in the protein. Biologists used
these tandem strategies to fill gaps in the productive continuity of the proposed mechanism.
The molecular biologists reasoned forward from the DNA, while the biochemists reasoned
backward from the finished protein. Their work met in the middle of the mechanism, with
the discovery of the various types of RNAs and their roles. One biochemist suggested that the
scientists were like workers building a tunnel: they started at opposite sides of a mountain, and
eventually they met in the middle.2
In sum, forward and backward chaining are reciprocal strategies for reasoning about one part
of a mechanism on the basis of what is known or conjectured about other parts in the mecha-
nism. They may be used independently of or in conjunction with the other two strategies of
using a schema type and assembling modules. The strategy of employing a schema is a top-down
strategy that provides a how-possibly overall organizational structure for the target mechanism.
Modular subassembly involves putting together functionally characterized, working subcom-
ponents of a schema. Finally, at a finer grain, one can construct a hypothesized mechanism by
reasoning forward or backward about the entities or activities themselves.
9. Computer simulations and biorobotics in mechanism discovery

William Bechtel (2016) uses case studies from circadian rhythm studies to show how computa-
tional modeling contributes to the discovery of complex mechanisms. Bacteria were found to
lack what had been considered a central component of the circadian clock mechanism in other
organisms and thereby indicated the nature of the crucial core of the clock mechanism in other
organisms. Both computational modeling and empirical investigations played roles in the dis-
covery of how the simplified mechanism works in bacteria. Researchers built a computational
model to test whether their proposed mechanism could generate the phenomenon (a typical
use of computational simulations). A state change in the model introduced a variable in need
of physical interpretation. The authors proposed a hitherto unknown change in the conforma-
tion of one of the proteins, thereby leading to a newly proposed module for this mechanism
requiring new experimental investigation. This case shows the interplay between constructing
a mechanistic hypothesis and building a computational model to test it. If computational trac-
tability involves adding black boxes (here, uninterpreted variables) so as to make the simulation
run, then researchers are guided in adding previously undetected components to the mechanis-
tic hypothesis. As Bechtel put it: “the modeling played a central role in the formulation of the
proposed mechanistic account; it did not merely serve to demonstrate that an already advanced
model could generate the phenomenon” (Bechtel 2016: 117).
263
Lindley Darden
Datteri and Tamburrini (2007) contrast the roles of implementing a proposed mechanism in
computational simulations versus in biorobots. Many of the same issues arise in mapping a bio-
logical mechanism description into a biorobotic one as in computational modeling: what are the
similarities and differences in the way the model mimics the biological activities? In biorobotics
too, requirements for implementing a working model can point to previously unrecognized
black boxes needed for productive continuity so that the robot works. However, a robot, placed
in an appropriate environment, can mimic aspects of the biological system for which there may
be no computational simulation.
Datteri and Tamburrini’s example is of the investigation of the chemotaxis in lob-
sters, namely their capability to track turbulent chemical plumes to their sources (Grasso
et al. 2000). The mechanism involves feedback analysis: lobster antennae can detect local
gradients in the chemical plume, and local gradients provide lobsters with directional infor-
mation to localize the plume source. Computer simulations appear to be unsuitable to test
mechanism schemas that are consistent with this broad picture of lobster chemotaxis, insofar
as accurate replication of plume turbulence in water is unfeasible (especially in view of the
difficulty of developing accurate fluid-dynamical models at the temporal and spatial scales
that are needed to simulate lobster sensorimotor loops in chemo-orientation tasks). In con-
trast, robots placed in real chemical plumes allow for direct comparison between robotic
and biological behaviors. Two previously proposed mechanism components did not enable
the robot lobster to adequately mimic the biological lobster’s behavior; the experiments
not only indicated which components of the mechanism needed to be changed but also
provided clues as to the needed repairs.
As Datteri and Tamburrini summed up:
Behavioral match evaluations of robotic and biological behaviors may enable one to
corroborate or falsify mechanism schemata and descriptions, to identify previously
overlooked interactions with the environment, to suggest the existence of entities
and activities implementing some role function in the target biological system, and
more generally to support the iterative specification of mechanism descriptions from
mechanism sketches.
(Datteri and Tamburrini 2007: 424)
From this computational and biorobotic work, we again see how construction, evaluation,
and revision are intimately tied in mechanism discovery. A mechanistic hypothesis or hypoth-
esis space is constructed, often in a sketchy way with numerous black boxes; evaluation of
hypothesized components via experimentation or simulation provides constraints and indicates
components in need of revision; and the outcome of those investigations serves to indicate the
kind of fix needed during anomaly resolution. This iterative process continues to remove black
boxes and find support for the hypothesized mechanism components, thereby removing incom-
pleteness and incorrectness.
10. Conclusion
In the discovery of mechanisms, the product shapes the process of discovery. The search
for mechanisms is guided by the very fact that it is a search for mechanisms rather than
something else (e.g., an entity, a mathematical law). Strategies for constructing mechanism
schemas include characterizing the phenomenon (to provide clues about the mechanism that
produces it), localization (to find where the mechanism operates), employing an abstract
264
schema (to provide the overall organization of the mechanism), modular subassembly (to
provide functionally characterized groupings of entities and activities), and forward/back-
ward chaining (to provide what comes next or what comes before any given step). Further
evaluation and revision strategies guide the iterative process of constructing, evaluating, and
revising mechanism schemas during mechanism discovery. More needs to be said about the
roles played by interlevel and interfield contexts of the target mechanism. Surely other strat-
egies will be found to add to this list as the new mechanistic philosophy of science works to
understand ways of discovering mechanisms.
Notes
1 Many thanks to Carl Craver for all the work we’ve done together on these ideas over the years. The DC/
Maryland History and Philosophy of Biology Discussion Group gave me very useful suggestions on an
earlier draft. Stuart Glennan and Phyllis Illari provided very helpful guidance for rewriting.
2 For more on Mendel to Watson and Crick, see Darden (1991), Watson and Crick (1953a). On forward
chaining in the discovery of the mechanism of protein synthesis, see Watson and Crick (1953a, 1953b);
Crick (1988). On backward chaining by biochemists, see Hoagland (1955, 1996), Hoagland et al. (1959),
and Zamecnik (1960). Zamecnik analogized their work to tunnel diggers (personal communication to
author). Their work is discussed in Rheinberger (1997), who emphasized experimental systems rather
than mechanisms, and in Darden and Craver (2002), who traced the discovery of the mechanism of
protein synthesis.
References
Baltimore, David (1970), “Viral RNA-dependent DNA Polymerase,” Nature 226: 1209–1211.
Bechtel, William (2016), “Using Computational Models to Discover and Understand Mechanisms,”
Studies in History and Philosophy of Science Part A 56: 113–121.
Bechtel, William and Robert C. Richardson (1993), Discovering Complexity: Decomposition and Localization
as Strategies in Scientific Research. Princeton, NJ: Princeton University Press.
Bechtel, William and Robert C. Richardson (2010), Discovering Complexity: Decomposition and Localization
as Strategies in Scientific Research. 2nd. ed. Cambridge, MA: MIT Press.
Bogen, James and James Woodward (1988), “Saving the Phenomena,” Philosophical Review 97: 303–352.
Craver, Carl F. (2007), Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience. New York:
Craver, Carl F. and Lindley Darden (2013), In Search of Mechanisms: Discoveries across the Life Sciences.
Chicago, IL: University of Chicago Press.
Crick, Francis (1988), What Mad Pursuit: A Personal View of Scientific Discovery. New York: Basic Books.
Darden, Lindley (1991), Theory Change in Science: Strategies from Mendelian Genetics. New York: Oxford
University Press.
Darden, Lindley (2002), “Strategies for Discovering Mechanisms: Schema Instantiation, Modular
Subassembly, Forward/Backward Chaining,” Philosophy of Science 69 (Proceedings): S354–S365.
Darden, Lindley (2006), Reasoning in Biological Discoveries: Mechanisms, Interfield Relations, and Anomaly
Resolution. New York: Cambridge University Press.
Darden, Lindley (2013), “Mechanisms versus Causes in Biology and Medicine,” in Hsiang-Ke Chao,
Szu-Ting Chen, and Roberta L. Millstein (eds.), Mechanism and Causality in Biology and Economics.
Dordrecht: Springer, pp. 19–34.
Darden, Lindley and Carl F. Craver (2002), “Strategies in the Interfield Discovery of the Mechanism
of Protein Synthesis,” Studies in History and Philosophy of Biological and Biomedical Sciences 33: 1–28.
Corrected and reprinted in Darden (2006, ch. 3).
Datteri, Edoardo and Guglielmo Tamburrini (2007), “Biorobotic Experiments for the Discovery of
Biological Mechanisms,” Philosophy of Science 74: 409–430.
Fagan, Melinda Bonnie (2012) “Jointness and Mechanistic Explanation,” Philosophy of Science 79: 448–472.
Gervais, Raoul and Erik Weber (2015), “The Role of Orientation Experiments in Discovering
Mechanisms,” Studies in History and Philosophy of Science Part A 54: 46–55.
265
Lindley Darden
Gilbert, Scott F. (2003), Developmental Biology, 7th ed. Sunderland, MA: Sinauer Associates.
Grasso, F. W., et al. (2000), “Biomimetic Robot Lobster Performs Chemo-Orientation in Turbulence
Using a Pair of Spatially Separated Sensors: Progress and Challenges,” Robotics and Autonomous Systems
30: 115–131.
Gutting, Gary (1980), “Science as Discovery,” Revue Internationale de Philosophie 34: 26–48.
Hanson, Norwood Russell ([1961] 1970), “Is There a Logic of Scientific Discovery?” in H. Feigl and
G. Maxwell (eds.), Current Issues in the Philosophy of Science. New York: Holt, Rinehart and Winston.
Reprinted in B. Brody (ed.), Readings in the Philosophy of Science. Englewood Cliffs, NJ: Prentice Hall,
pp. 620–633.
Hoagland, Mahlon B. (1955), “An Enzymic Mechanism for Amino Acid Activation in Animal Tissues,”
Biochimica et Biophysica Acta 16: 288–289.
Hoagland, Mahlon B. (1996), “Biochemistry or Molecular Biology? The Discovery of ‘Soluble RNA’,”
Trends in Biological Sciences Letters (TIBS) 21: 77–80.
Hoagland, Mahlon B., Paul Zamecnik, and Mary L. Stephenson (1959), “A Hypothesis Concerning the
Roles of Particulate and Soluble Ribonucleic Acids in Protein Synthesis,” in R. E. Zirkle (ed.), A
Symposium on Molecular Biology. Chicago, IL: Chicago University Press, pp. 105–114.
Josephson, John R. and Susan G. Josephson (eds.) (1994), Abductive Inference: Computation, Philosophy,
Technology. New York: Cambridge University Press.
Kästner, Lena (2015), “Learning about Constitutive Relations,” in U. Mäki et al. (eds.), Recent
Developments in the Philosophy of Science: EPSA13 Helsinki, European Studies in Philosophy of Science
1, pp. 155–167. Dordrecht: Springer. DOI 10.1007/978-3-319-23015-3_12.
Machamer, Peter, Lindley Darden, and Carl F. Craver (2000), “Thinking about Mechanisms,” Philosophy
Meheus, Joke and Thomas Nickles (eds.) (2009), Models of Discovery and Creativity. Series: Origins: Studies
in the Sources of Scientific Creativity, v. 3. Dordrecht: Springer.
Nachtomy, Ohad, A. Shavit, and A. Yakhimi (2007), “Gene Expression and the Concept of the
Phenotype,” Studies in History and Philosophy of Biological and Biomedical Sciences 38: 238–254.
Nickles, Thomas (1977), “Heuristics and Justification in Scientific Research: Comments on Shapere,” in
in F. Suppe (ed.), The Structure of Scientific Theories. Urbana: University of Illinois Press, pp. 571–589.
Nickles, Thomas (ed.) (1980a), Scientific Discovery, Logic and Rationality. Dordrecht: Reidel.
Nickles, Thomas (ed.) (1980b), Scientific Discovery: Case Studies. Dordrecht: Reidel.
Popper, Karl R. (1965), The Logic of Scientific Discovery. New York: Harper Torchbooks.
Rheinberger, Hans-Jörg (1997), Towards a History of Epistemic Things: Synthesizing Proteins in the Test Tube.
Stanford: Stanford University Press.
Schaffner, Kenneth (1993), Discovery and Explanation in Biology and Medicine. Chicago, IL: University of
Chicago Press.
Simon, Herbert A. (1977), Models of Discovery. Dordrecht: Reidel.
Simon, Herbert A. (1996), The Sciences of the Artificial. Third Edition. Cambridge, MA: MIT Press.
Thagard, Paul (1998), “Explaining Disease: Causes, Correlations, and Mechanisms,” Minds and Machines
8: 61–78.
Thagard, Paul (1999), How Scientists Explain Disease. Princeton, NJ: Princeton University Press.
Thagard, Paul (2003), “Pathways to Biomedical Discovery,” Philosophy of Science 70: 235–254.
Temin, Howard M. and Satoshi Mizutani (1970), “RNA-dependent DNA Polymerase in Virions of Rous
Sarcoma Virus,” Nature 226: 1211–1213.
Watson, James D. and F. H. C. Crick (1953a), “Molecular Structure of Nucleic Acids: A Structure for
Deoxyribose Nucleic Acid,” Nature 171: 737–738.
Watson, James D. and F. H. C. Crick (1953b), “Genetical Implications of the Structure of Deoxyribonucleic
Acid,” Nature 171: 964–967.
Wimsatt, William (2007), “Heuristics and Human Behavior,” in Re-Engineering Philosophy for Limited
Beings. Cambridge, MA: Harvard University Press, pp. 75–93.
Zamecnik, Paul C. (1960) “Historical and Current Aspects of the Problem of Protein Synthesis,” The
Harvey Lectures 1958–1959. Series 54. New York: Academic Press, pp. 256–281.
266
20
MECHANISMS AND
DYNAMICAL SYSTEMS
1. Introduction
The mathematical framework of dynamics, or more specifically dynamical systems theory (hereafter
DST), is having a major impact across many sectors of contemporary biology and neurosci-
ence. The DST toolkit (described in more detail shortly), which includes differential equations,
geometric state-space analyses, and other visualization techniques such as attractor landscapes
and bifurcation diagrams, is playing an increasingly important role in modeling and explaining
the time-varying activity of biological and neural systems ranging from single neurons and local
circuits to entire brain networks, biological populations, and complex ecosystems (for book-
length treatments, see Abraham and Shaw 1992; Amit 1992; Izhikevich 2005; Strogatz 1994).
The surge of scientific interest in dynamical modeling has attracted considerable philosophi-
cal attention. One recent topic of debate concerns the nature of the explanations these models
provide. Some argue that the powerful mathematical framework of dynamics signals the emer-
gence of a radically new explanatory paradigm that is importantly distinct from, and provides a
competing alternative to, the dominant framework of mechanistic explanation (e.g., Chemero
and Silberstein 2008; Stepp et al. 2011; Ross 2014; Silberstein and Chemero 2013). Others
more friendly to the mechanistic perspective grant that DST introduces something distinctive
and new—for example, a new set of mathematical techniques for describing patterns of change
over time in neural systems, or even a novel theoretical framework for characterizing cogni-
tion and brain activity in non-representational or non-computational terms (Van Gelder 1995,
1998)—but reject the idea that it offers a distinct and competing explanatory framework to that
of mechanism. Instead of viewing dynamical and mechanistic models as competitors, propo-
nents of the mechanistic perspective see dynamical models as complementary to the description
of mechanisms (e.g., Bechtel and Abrahamsen 2010; Kaplan and Bechtel 2011; Kaplan and
Craver 2011; Kaplan 2015; Zednik 2011). In what follows, I survey the different sides of the
debate over the nature of dynamical explanation with a specific focus on examples from cogni-
tive science and neuroscience. (For related discussion of dynamical modeling in other areas of
biological science including systems biology, see Chapter 27.) First, as background, I outline
some basic features of the DST framework. After exploring the main positions staked out in
the debate and identifying their primary strengths and limitations, I highlight the advantages of
267
embracing a mechanistic approach to dynamical explanation. I then address several major open
challenges facing the mechanistic approach and conclude by briefly discussing some important
future directions for the debate.
2. Dynamics: a primer
A dynamical system is, roughly speaking, a set of variables that interact over time. Changes in
these variables over time—their time series—can exhibit various patterns. Dynamics or DST
comprises a highly general set of mathematical techniques for modeling, analyzing, and visual-
izing these patterns in time series data. At the core of every dynamical model is one or more
differential equations (for the continuous time case) or difference equations (for the discrete
time case), which specifies how a system changes at any given time point as a function of its
state at that time. Although differential (and difference) equations are powerful modeling tools,
some do not admit of exact solutions—i.e., the function or set of functions satisfying a given
equation. When analytical techniques fail, numerical methods are often used to arrive at inexact
or approximate solutions with arbitrarily high levels of precision (e.g., Mascagni and Sherman
1989). In other circumstances, it is more appropriate or useful to model long-term system
behavior in terms of solution trajectories in a geometrically defined space.
Geometric methods and concepts are central to DST. A state space is defined as the set of all
possible values or states that a given dynamical system can take or be in over time. Each system
variable defines a corresponding dimension in this space and the number of dimensions of this
space reflects the total number of state variables (e.g., the dynamical system in Figure 1 is defined
by two variables, y1 and y2). Different possible trajectories of the system within that space are
determined by plugging in different values to the relevant differential equation. In this manner,
the differential equation serves to define a vector field—an assignment of an instantaneous direc-
tion and rate of change for each point in the state space (Figure 20.1a). An individual solution
trajectory is a particular temporal sequence of states through this vector field, from some initial
state of the system. The set of all possible solution trajectories is called a flow (Figure 20.1b).
Many dynamical systems will eventually converge on a small subregion of the state space, called
the limit set. If the system state passes through a limit set, the dynamics will operate to keep it
there and small external perturbations will only alter the system momentarily. Afterward, it will
converge back to the same state. Limit sets can either be single points (fixed or equilibrium points)
Figure 20.1 Basic constructs of DST. (a) A vector field for a two-dimensional dynamical system defined
by the two differential equations, dot[y1] = f1(y1, y2) and dot[y2] = f2(y1, y2). (b) A flow
diagram depicting representative solution trajectories for the same system. (c) A phase
portrait depicting the system’s limit sets, their stabilities, and their basins of attraction. Dots
indicate equilibrium points. Other limit sets are indicated by lines. Source: Beer 2000
268
Mechanisms and dynamical systems
Figure 20.2 HKB experiment and model. (a) In-phase finger oscillation with relative phase (ϕ) = 0°.
(b) Anti-phase finger oscillation with relative phase (ϕ) = 180°. (c) Three-dimensional
vector field diagram. Thick lines indicate stationary or fixed points of the system (i.e.,
where dϕ/dt = 0). Curved surface and grayscale gradients depict the overall attractor
landscape for the system. Source: Zednik 2011; adapted from Kelso 1995
or trajectories that loop back on themselves (limit cycles or oscillations) (Figure 20.1c). For a stable
limit set, also known as an attractor, all nearby trajectories will converge to it. The set of points
that converge to an attractor over time is termed a basin of attraction. The overall characteriza-
tion of all the limit sets and basins of attraction of a dynamical system is called a phase portrait
(Figure 20.1c). Finally, the flow and overall attractor landscape can depend on other independ-
ent parameters (called control parameters). Sometimes the flow field of a dynamical system changes
continuously when a control parameter is varied. Other times, a discontinuous change in the
flow pattern can result from small and continuous changes in the control parameter, called
a bifurcation. With this basic terminology in hand, we can move on to explore the nature of
dynamical explanation.
The well-known Haken–Kelso–Bunz (HKB) model of bimanual coordination (Haken et al.
1985) provides a useful illustration of DST in action. The HKB model was originally developed
to account for behavioral data collected in a bimanual coordination task. Subjects made repetitive
side-to-side (oscillating) movements of their index fingers in the transverse plane either in-phase
(simultaneous, mirror-symmetric movements toward the midline of the body; Figure 20.2a) or anti-
phase (simultaneous, parallel movements to the left or right of the body midline; Figure 20.2b).
Movements were completed in time with a pacing metronome, and metronome speed was an
independent variable under experimental control. Several interesting results emerged. First, subjects
could reliably perform both in-phase and anti-phase coordination patterns at low oscillation frequen-
cies. Second, in trials where movement frequency was increased beyond a certain critical threshold,
subjects could no longer maintain the anti-phase movement pattern and switched abruptly and
involuntarily into the in-phase pattern. Third, switches never occurred on trials initiated in-phase,
even when the critical frequency was exceeded.
The core of the HKB model is the differential equation capturing the rate of change over
time in relative phase between the fingers:
dφ (1)
= − a sin φ − 2b sin 2φ
dt
where ϕ represents relative phase (in-phase, ϕ 0°; anti-phase, ϕ ± 180°); a and b are empirically
fitted parameters such that the ratio b/a is inversely proportional to frequency. In the language
of DST, the ratio b/a is a control parameter. The corresponding DST analysis depicts system
269
behavior in terms of an attractor landscape that changes as a function of the control parameter
and time (Figure 20.2c). When oscillation frequencies are relatively low (b/a > 0.25), two stable
attractors exist corresponding to in-phase and anti-phase movement patterns. When oscillation
frequencies are relatively high (b/a < 0.25), only in-phase coordination has a stable attractor and
the anti-phase pattern becomes unstable. As the system moves in the direction of a decreasing
b/a ratio (movement from top to bottom along the y-axis in Fig. 20.2c), associated with higher
oscillation frequencies, the stable attractor for anti-phase coordination disappears and is subse-
quently transformed into an unstable equilibrium point (at b/a = 0.25). At this critical frequency,
the system undergoes a phase transition or bifurcation. Once the system crosses this threshold,
any small perturbation pushes it toward the stable fixed point attractor corresponding to the in-
phase movement pattern.
The HKB model is an exemplar of the dynamical approach. It remains one of the most influ-
ential quantitative models of human motor behavior (Beek et al. 2002; Fuchs 2013; Swinnen
2002) and serves as a central case in philosophical discussions concerning dynamical explanation
(Gervais and Weber 2011; Kaplan 2015; Kaplan and Craver 2011; Kelso 1995; Walmsley 2008).
3. Varieties of approaches to dynamical explanation

According to the once dominant covering law (CL) model, explanation involves (deductive or
inductive) subsumption of some event or phenomenon to be explained under a law or set of
laws (Hempel 1965). More precisely, a specification of the relevant laws and empirical condi-
tions under which the phenomenon obtains (initial and boundary conditions) is supposed to pro-
vide good evidential grounds for expecting the occurrence of the explanandum-phenomenon.
According to the CL approach to dynamical explanation (hereafter, DCL), explana-
tory dynamical models are just special cases of CL explanations (Bechtel 1998; Bechtel and
Abrahamsen 2002; Walmsley 2008). The DCL approach is relatively common in the literature,
and many dynamicist researchers explicitly describe their models as involving laws (e.g., Bressler
and Kelso 2001; Kelso 1995; Schöner and Kelso 1988). Zednik (2011) characterizes it as the
“received view” about dynamical explanation. Despite its prominence, proponents of DCL
have failed to address many of the standard challenges to the general nomological conception of
explanation upon which it is based (for additional discussion, see Kaplan 2015).
The main challenge facing DCL involves showing how explanatory dynamical models either
make explicit reference to a genuine law or implicitly convey information about the existence
of the relevant law. If dynamical models cannot satisfy this basic requirement on CL explana-
tions, DCL is a non-starter. Answering this challenge is difficult because it requires a sufficiently
precise notion of law that can successfully distinguish laws from accidental generalizations;
and consensus about general criteria for lawhood remains elusive in the philosophy of science
(Hempel 1965; Salmon 2006; Woodward 2000, 2003).
Defenders of DCL can avail themselves of one of two main options for characterizing the
nomological character of the generalizations appealed to in dynamical explanations. It is impor-
tant to note that these are the primary options available if the defender of DCL wishes to remain
within the covering law tradition. If, however, one is willing to abandon the central notion
of subsumption under law, then other options (including but not limited to the mechanistic
approach) come into view such as interventionist approaches to explanation (e.g., Woodward
2000, 2003). According to Woodward’s account, the explanatory import of a generalization
depends on the degree to which it is invariant under intervention rather than on how well it
satisfies the traditional criteria for lawhood. Exploring how dynamical explanation might fit
within these approaches goes beyond the scope of this chapter. Returning to the main thread,
270
the first option is to maintain that these generalizations do in fact satisfy the standard criteria for
lawhood and therefore qualify as full-blown laws, which in turn renders them suitable to feature
in CL explanations. The second option is to claim that, like many explanatory generalizations
found in the special sciences, dynamical generalizations are best construed as so-called non-strict,
qualified, or ceteris paribus laws (e.g., Fodor 1991; Pietroski and Rey 1995). Both options face
difficulties. (For further discussion of laws and their connection to mechanistic explanation, see
Glennan (1996) and Chapters 11 and 12 of this volume.)
The main problem with the first option is that the mathematical generalizations commonly
found in dynamical models do not seem to satisfy many of the standard criteria for lawhood. It is
widely assumed, for example, that whatever else a law may be, it must at least be an exception-
less generalization with wide scope. Scope describes the range of different individual systems (or
types of systems) over which a given generalization holds (e.g., the gravitational inverse square
law has wide scope in the sense that it applies to all massive bodies throughout the universe).
Satisfying these criteria is important because if generalizations admit of exceptions or have scope
restrictions, they will fail to play the role required of them by the CL framework. This is
because deductive inference of the explanandum in a CL explanation cannot occur unless the
law statement in the explanans takes the form of a universally quantified generalization (that
ranges over its specified domain without exception). Importantly, this challenge also applies to
inductive-statistical explanations (Hempel 1965) since exceptions in a statistical law featured in the
explanans can lower the probability conferred on the occurrence of the explanandum event and
thereby weaken the explanation.
Yet the mathematical generalizations typically employed in dynamical models, like most
generalizations in biology, appear to be both highly restricted in scope and exception-ridden.
The paradigmatic HKB model, for example, covers an impressive range of coordination patterns
involving two or more oscillating components—bimanual coordination (Haken et al. 1985),
coordinated movements across subjects (Schmidt et al. 1990), and even some forms of social
coordination (Oullier et al. 2008). Yet its scope is restricted in important respects. For example,
the model fails to apply to all rhythmic limb movements including those involved in walking
or running. No discontinuous shift from walking or running (anti-phase leg movements) to
hopping (in-phase leg movements) occurs when humans increase their movement speed above
some critical threshold (Rosenbaum 1998).
Along similar lines, many of the mathematical regularities captured by dynamical models
are exception-ridden in the sense that they only hold within a certain domain or regime of
changes and break down outside of these. The HKB model holds for movement frequen-
cies on the order of a couple of cycles per second (Haken et al. 1985), but it remains unclear
whether the relationships characterized in the model remain invariant across all changes in this
movement frequency. For example, exceptions might be expected at extreme speeds (e.g.,
1000 cycles/second) approaching or exceeding the limits imposed by the musculoskeletal and
nervous systems.
A number of philosophers have emphasized how the ability of dynamical generalizations
to support counterfactuals might justify their status as laws (e.g., Bechtel 1998; Clark 1997;
Van Gelder 1998). Clark, for instance, highlights the importance of “the way it [the system]
responds to new, not-yet-encountered circumstances” (Clark 1997, 119). Bechtel claims that
“DST accounts . . . are clearly designed to support counterfactuals [and] [t]his suggests that it
may be appropriate to construe these DST explanations as being in the covering law tradition”
(Bechtel 1998, 311).
Given the centrality of the DST notion of a state space embodying information about
all possible (and actual or observed) states and trajectories that a dynamical system can take,
271
philosophical emphasis on counterfactual support should be unsurprising. Possible state-space

trajectories constitute straightforward counterfactuals about what a given dynamical system
would have done if things had been different. Unfortunately, even though dynamical gener-
alizations satisfy this criterion in spades, this criterion fails to distinguish laws from accidental
generalizations. This is because many accidental generalizations support counterfactual predic-
tions. Borrowing an example from Woodward (2003, 280), the generalization “All the coins in
Clinton’s pocket are dimes” is both accidental and counterfactual-supporting. Supposing as a
background condition that Clinton permitted only dimes in his pocket, the above generalization
supports the counterfactual: “If c were in coin in Clinton’s pocket, then it would be a dime”
(Woodward 2003, 280). Examples of this kind illustrate how appeals to counterfactual support
appear inadequate to underwrite an account of dynamical laws.
This brings us to the second option, which involves construing dynamical generalizations as
non-strict laws. This option is also fraught with difficulties. According to the general strategy,
a generalization with exceptions can still play an explanatory role if it can be “completed” by
specifying some further set of conditions that, together with the conditions outlined in the orig-
inal generalization, are nomologically sufficient to generate the explanandum (Reutlinger and
Unterhuber 2014). When supplemented with the appropriate “completer,” the resulting gener-
alization qualifies as an exceptionless law because it restricts the scope to precisely the range of
circumstances where the regularity obtains without exception. The well-known problem with
this proposal involves filling out the completer clause without producing generalizations that are
either trivially true or false (Earman and Roberts 1999; Woodward 2002, 2003). Many philoso-
phers remain skeptical about whether this challenge can be met (Earman et al. 2002; Reutlinger
and Unterhuber 2014). Even if one assumes this problem can be handled, other difficulties arise.
For example, Woodward (2002) argues that the very notion of a non-strict law incorporating
qualifying clauses in this manner provides an exceedingly poor philosophical reconstruction of
the kind of causal generalizations they are supposed to pick out in the special sciences.
Proponents of DCL therefore find themselves in a highly undesirable position. They must
either face up to the difficult challenge of producing a satisfactory account of the nomological
status of dynamical generalizations or jettison the approach.
Another approach to dynamical explanation emphasizes the tight connection between pre-
diction and explanation (Chemero 2011; Chemero and Silberstein 2008; Stepp et al. 2011; Van
Gelder 1998). This view has been termed predictivism (Kaplan and Craver 2011). Even though
predictivism drops the problematic requirement that laws are needed for explanation, it is still
closely connected to the CL approach and therefore faces many of the same difficulties.
The first problem facing predictivism as an account of the explanatory power of dynamical
models is that well-known counterexamples to the CL model demonstrate how prediction is
insufficient for explanation (Salmon 2006). Similar examples can be constructed to illustrate
how dynamical models can be predictively adequate in the sense that they predict all the rele-
vant aspects of the phenomenon with the desired precision and accuracy, yet fail as explanations
because the model variables represent magnitudes that merely correlate with some other com-
mon cause for the phenomenon. Consider a dynamical model describing a set of three gears in
an automotive transmission system. The gears are organized such that only one gear (gear C)
directly connects to the motor. The other gears (A and B) rotate only as a consequence of C’s
rotational motion. Because the behavior of all three gears is time-locked to motor speed, their
dynamics will be coupled (i.e., correlated). One could build a dynamical model that incorpo-
rates information about the speed or acceleration of gear A and thereby predicts the behavior
of gear B (and vice versa). Yet it is problematic to characterize one as explaining the other
because a common cause—gear C—is fundamental to explaining the correlation between them.
272
Similarly, one could use temporal information about A (or B) to predict the behavior of C,
even though the rotational motion of C causes the motion of A (and B). This move is equally
problematic from the point of view of explanation.
Explanations must respect this fundamental asymmetry between causes and effects (and
mere correlates), even though effects and correlated variables can be extremely useful pre-
dictors of their causes. These and many other similar examples can be generated to illustrate
how prediction is insufficient for explanation, and why the predictive force of a given model
should not be confused with its explanatory force. Instead, what is needed is an account that
provides an understanding of precisely why the regularities that constitute the phenomenon
hold in the first place.
Because it assimilates explanatory and predictive power, another major problem for the
predictivist view is that it is incapable of capturing the explanatory gains among predictively
equivalent models where one describes the causal structure of the target system with increased
accuracy. According to predictivism, the quality of an explanation can be improved primarily
by increasing its predictive power. Yet there seem to be other ways—including building in
more causal-mechanical details—to improve an explanation, which the predictivist view cannot
accommodate. According to predictivism, unless these additional details improve the predictive
power of the model, they are explanatorily inert. But this runs counter to widespread views
about how scientific progress is achieved, and specifically about the kinds of refinements and
model-building activities that produce better explanations.
Given the problems facing predictivism (and DCL), what is the alternative? One appealing pos-
sibility is to treat dynamical explanation as a species of mechanistic explanation. Doing so allows us
to sidestep the problems outlined above. It also grounds dynamical explanations in a dominant and
well-understood form of explanation that is widespread across the biological sciences.
Instead of emphasizing the gap between dynamics and mechanism, as the previous approaches
do, another option is to show how the apparent gap is to be bridged. When biologists and neu-
roscientists put forward explanations, they frequently seek to identify the mechanism responsible
for maintaining, producing, or underlying the phenomenon of interest (Bechtel and Richardson
1993/2010; Craver 2007; Machamer et al. 2000). In other words, they seek to provide mech-
anistic explanations. Mechanistic explanations specify the component parts, the operations or
activities of these parts, and the (spatial and temporal) organization of the parts and their activi-
ties in the mechanism as a whole. All major accounts of mechanistic explanation identify a key
role for each of these core elements. Despite the fact that lip service is frequently paid to the
importance of organization, this remains the most underdeveloped aspect of many accounts of
mechanistic explanation.
Brief reflection nevertheless reveals how temporal organization is critical to mechanisms and
mechanistic explanations. Consider the internal combustion engine. Engines are composed of
structural parts including pistons, spark plugs, and intake valves that perform activities such as
sliding, sparking, and opening. Critically, these parts must interact in space and time in a highly
organized manner for the engine to function properly. Engine components must bear specific
spatial relationships to the other parts with which they have causal interactions. For example,
pistons must be located within the cylinders, which in turn must be spatially proximate and
mechanically linked to the crankshaft via connecting rods to produce torque in the axles. Precise
temporal organization of the activities performed by the engine parts is no less important to
engine performance. For example, spark plugs must emit their spark at the top of the compres-
sion stroke of the pistons. Intake and exhaust valves must open and close at precise times so that
they are sealed shut during compression and combustion and open during the exhaust stroke.
Together, these comprise the engine dynamics.
273
Although in this particular case it might not be especially illuminating, the state of each
of the components and/or the global state of the engine system as a whole could be quan-
tified (e.g., position, velocity, or acceleration for the parts; maximum power or torque for
the engine as a whole) and plotted as a function of time. Each could also be subjected to a
dynamical analysis according to which its evolving state is represented as a trajectory in a suitable
state space. Nevertheless, for many complex natural systems as opposed to artificial or designed
systems, dynamical analyses can help to reveal dynamic patterns or latent temporal structure
in system activity that would otherwise be exceedingly difficult if not impossible to discern
(e.g., Churchland et al. 2012).
Organization is a necessary part of most moderately complex mechanisms such that per-
turbing either the spatial organization or temporal dynamics of a mechanism, even while the
components and their individual activities remain unchanged, can have appreciable effects on
performance. Thinking about mechanistic explanation, then, it is clearly insufficient to describe
only the properties and activities of the component parts in a given mechanism without giv-
ing adequate weight or attention to the spatial and/or temporal organization of those parts and
activities. Often this point is underappreciated or lost when considering the nature of mechanis-
tic explanation (for a notable exception, see Bechtel and Abrahamsen 2010).
This discussion highlights how even relatively simple mechanisms, such as internal combus-
tion engines, can exhibit rather complex temporal organization. Consequently, understanding
the dynamical “structure” of a mechanism can sometimes be just as important as understanding
its physical structure. This last point rings especially true in the context of neuroscience and
biology, where mechanisms often exhibit a wide range of complex dynamic patterns that are
essential to their proper functioning (for additional discussion, see Kaplan 2015).
A full development and defense of the mechanistic approach to dynamical explanation
sketched would ideally involve walking through some case studies in which both the frame-
work of dynamics and that of mechanism were jointly employed in the service of building
adequate explanations. Nevertheless, the view about the relationship between the frameworks
of dynamics and mechanism that falls out of the above considerations is one of subsumption,
not competition. In particular, dynamical models are usefully employed to reveal the tempo-
ral organization of activity in mechanisms. Because dynamical models are subsumed within
the broader toolkit for describing mechanisms, their explanatory value can be seen as clearly
depending on the presence of an associated account (however incomplete) of the parts in the
mechanism (and their interactions) that support, maintain, or underlie these activity patterns.
On this view, dynamical modeling approaches do not signal the emergence of a new explana-
tory framework that is distinct from mechanistic explanation. Instead, dynamical models with
explanatory force are to be understood as playing a vital role within the mechanistic framework.
4. Open challenges for the mechanistic approach to

dynamical explanation
The mechanistic approach to dynamical explanation faces challenges along several fronts. In this
section, I focus on two related challenges raised by the purported non-decomposability of many
dynamical systems.
It is widely acknowledged that the heuristic strategies of decomposition and localization are
vitally important for discovering mechanisms and building mechanistic explanations (Bechtel
and Richardson 1993/2010). Roughly, decomposition involves the identification of the com-
ponent processes or activities that contribute to the overall performance of a mechanism and
are recruited when the mechanism is engaged. Localization involves the identification of these
274
different component activities with the behavior or capacities of specific physical parts of the
mechanism. Silberstein and Chemero (2013) recently argue that sometimes the nature of
dynamical systems prevents the application of these strategies, and that in such cases mechanistic
explanations will be unavailable in principle. They identify several specific types of circum-
stances in which failures of decomposability or localizability will occur. I will argue that these
challenges can be rebutted from a mechanistic perspective.
The first type of problematic situation Silberstein and Chemero identify is “when there are
no component parts or operations that can be distinguished (such as a connectionist network),
in which case one can only talk about organizational features” (Silberstein and Chemero 2013,
961–2). For ease of reference, call this the challenge from functionally indistinguishable component
parts. The challenge seems to be that if the component parts in a system are doing function-
ally indistinguishable things, then the mechanistic strategies of functional decomposition and
localization will fail to gain traction. Consequently, mechanistic analyses will either provide
relatively thin explanations or no explanations at all. Surprisingly, arch-mechanists Bechtel and
Richardson make a similar point much earlier when they claim that:
Network models do account for the cognitive performance, but often they do so
without providing an explanation of component operations that is intelligible in
terms of the overall task being performed. The network is a cognitive system; the
components are not. The result is that we do not explain how the overall system
achieves its performance by decomposing the overall task into subtasks, or by
localizing cognitive subtasks.
(Bechtel and Richardson 1993/2010, 214)
The common idea here seems to be that when systems include parts that are functionally indis-
tinguishable (or unintelligible)—connectionist networks being an exemplary case—mechanistic
explanation cannot get off the ground. Although there are good reasons to reject the claim that
mechanisms in which all the parts are essentially performing the same function are not suitable
candidates for mechanistic explanation, let us grant this assumption for the sake of argument.
Even under this assumption, there are problems for the view.
At a relatively coarse-grained level, it is certainly true that nodes or units in artificial neural
networks do perform the same computational operations—they integrate signals and compute
some function of their inputs. More specifically, the activation or transfer function (e.g., step
or sigmoidal), which defines the output of a unit given some input or set of inputs, is typi-
cally implemented across all units within a layer (e.g., the hidden layer). So, there is a sense
in which each network “component” is doing the same thing (i.e., has the same functional
profile) and thus stands poised to block the strategies of decomposition and localization at the
heart of the mechanistic approach. Nevertheless, this challenge ignores a crucial feature of
such networks—their weights.
All connectionist networks contain sets of adaptive weights, i.e. numerical parameters that
are precisely tuned by a learning algorithm (e.g., backpropagation) during training. It is in virtue
of the systematic adjustment of weights across a network that it comes to exhibit the overall
performance that it does. In the untrained state, before weight adjustment (i.e., when weight
assignments are essentially random), network performance for any desired input-output func-
tion is typically extremely poor. Consequently, it might be reasonable to view untrained neural
networks as lacking functionally distinguishable parts, since the weights are randomly distributed
and they do not yet play any determinate functional (representational or information-processing)
role. Nevertheless, once trained, this assertion is clearly false. Suitably trained networks have
275
functionally distinguishable parts that make them susceptible to mechanistic decomposition and
localization. Reinforcing this point is the fact that numerous techniques are now available to
“open the hood” and quantitatively (and qualitatively) assess the diverse functional (e.g., rep-
resentational or information processing) roles that individual units play in support of overall
network performance. These methods include Hinton diagrams (e.g., Hinton 1986); hierarchical
clustering analysis (e.g., Sejnowski and Rosenberg 1987); and even receptive-field techniques
(e.g., Blohm et al. 2009; Zipser and Andersen 1988). Hence, this challenge seems to rely on an
overly simplistic view of neural network function and structure.
Since artificial neural networks are simulated on digital computers, the response becomes
even more forceful in the context of real biological neural networks. This is because biological
circuits, in contrast to their artificial counterparts, are typically composed of a diversity of neu-
ronal cell types (the “units”) that have different functional and structural properties. Moreover,
functionally distinguishable synaptic “weights” on axonal connections in real biological circuits
reflect structurally distinguishable underlying parts. Differences in axonal “weights” can reflect
underlying structural differences in myelination patterns that influence how well the input signal
propagates through the axon, or differences in the number and/or density of presynaptic vesicles
and AMPA and NMDA receptors on the cell membrane that influences the amount of neuro-
transmitter released into the synapse and absorbed by the postsynaptic neuron.
In contrast to the claims of Silberstein and Chemero (2013), artificial and biological neural
networks possess functionally and/or structurally distinguishable parts. This opens the door to
the powerful mechanistic strategies of decomposition and localization.
The second challenge amounts to the claim that the mechanistic approach and more specifi-
cally, decomposition and localization, will fail “when there are component parts and operations
but their individual behaviors systematically and continuously affect one another in a nonlinear
fashion” (Silberstein and Chemero 2013, 961–2). For ease of reference, call this the challenge from
nonlinearity. Before this challenge can be addressed or rebutted, however, one must clarify what
kind of nonlinearity is being invoked. It turns out that the notion of nonlinearity at the heart of
Silberstein and Chemero’s challenge is ambiguous and can be understood in (at least) two differ-
ent ways. Consequently, there are actually two distinct challenges for the mechanist to address.
In what follows, I show how both of these challenges can be answered.
One kind of nonlinearity, which has received considerably more attention in the mechanistic
literature than other kinds, involves the non-sequential temporal organization of component
activities or processes in a mechanism (e.g., Bechtel and Richardson 1993/2010, 202). Call this
process nonlinearity. Bechtel and Richardson, for example, discuss “explanations in which the
component tasks can be thought of as following a linear, sequential order” and go on to contrast
these with explanations that invoke “cyclic” organization (1993/2010, 202). They contrast the
two as follows: “[c]yclic rather than linear organization occurs when the activity of any given
component is dependent on a variety of other components that, in turn, depend on it” (Bechtel
and Richardson 1993/2010, 202).
Process nonlinearity poses no threat for the prospects of mechanistic explanation of dynam-
ical systems. Many neural dynamical systems, for instance, exhibit highly non-sequential or
“cyclic” temporal organization between mechanism components including parallel process-
ing, recurrent connections, and feedback loops. Yet neural systems exhibiting these properties
routinely yield to mechanistic explanation. A hallmark example is the mechanistic explana-
tion of the action potential. The underlying neural mechanism involves a complex interaction
or feedback loop between the activities of voltage-gated ion channels spanning the neuronal
membrane. Briefly, depolarization of the membrane potential from its resting level initiates the
rapid opening of sodium channels and a correspondingly sharp increase in sodium current into
276
the neuron (the so-called fast positive cycle). In parallel, the initial depolarization event also
induces the opening of slower potassium channels leading to a correspondingly slower increase
in potassium current out of the neuron and subsequent repolarization of the membrane potential
(the so-called slow negative cycle). In the case of the action potential, non-sequential temporal
organization of ion channel activities (process nonlinearities) are a fundamental aspect of the
underlying mechanism (for additional discussion, see Kaplan 2015). Nonlinearity in the form
of non-sequential temporal organization is an unproblematic feature of (some) mechanisms and
poses no problem for mechanistic explanation.
It remains possible that the anti-mechanistic challenge issued by Silberstein and Chemero
(2013) involves an entirely different kind of nonlinearity and so the previous response misses
its mark. In fact, there is another important kind of nonlinearity widely discussed across many
sciences (especially in neuroscience) that concerns the nature of a system’s input-output profile
and more specifically whether a system’s output is a linear function of its inputs. Call this response
nonlinearity. Perhaps this is the kind of nonlinearity they have in mind. If dynamical neural sys-
tems are nonlinear in this sense, perhaps this is why they will evade mechanistic explanation?
While it is true that many dynamical systems (including many neural systems) are nonlinear in
this sense, the anti-mechanistic conclusion does not follow.
Response nonlinearities are certainly commonplace in complex dynamical systems.
Computational neuroscientists Eliasmith and Anderson (2003) indicate why response non-
linearities should be expected in complex information-processing systems in the brain: “the
impressively complex behavior exhibited by neural systems is unlikely to be fully explained by
linear transformations alone. Rather, there is abundant evidence that nonlinear operations are
central to neural transformations” (2003, 153). But this does not imply that the prospects for
mechanistic explanation are dim. Understanding why requires a closer look at how this kind of
nonlinearity is defined in mathematics, engineering, and signal processing theory.
Response nonlinearity is standardly defined in terms of failing to satisfy the assumptions of
linearity. In mathematics, a system is said to be linear if and only if it exhibits both homogeneity
and additivity. Homogeneity entails that a change in the size of the input signal to a given system
produces a corresponding or proportional change in the output signal. More formally, if input
signal x[n] results in output signal y[n], then an input of kx[n] results in an output of ky[n], for
any input signal and constant, k. A system is said to be additive if added signals pass through
it without interacting; signals added at the input produce signals that are added at the output.
Again, more formally, if an input of x1[n] results in an output of y1[n], and if x2[n] results in
y2[n], then x1[n] + x2[n] results in y1[n] + y2[n]. Any system exhibiting both homogeneity and
additivity is linear. Any system failing to have both of these properties is nonlinear. This is a
suitably precise characterization of response nonlinearity. Now the question is whether this kind
of nonlinearity raises insurmountable problems for mechanistic explanation.
Evidence from neuroscience strongly suggests a negative answer to this question. All neural
dynamical systems exhibiting thresholding (e.g., spiking) behavior are nonlinear in this sense
since they fail to satisfy homogeneity and additivity. Yet these systems routinely succumb to
mechanistic explanation. The already discussed case of the action potential provides one clear
example. Other more exotic cases involve divisive normalization in neural systems (Carandini
and Heeger 2012). Normalization operations also fail to meet the strictures of homogeneity and
additivity, yet neuroscientists have proceeded to develop highly refined mechanistic models for
divisive normalization computation in the fly olfactory system (Olsen et al. 2010) and have pro-
duced partially worked out mechanistic explanations of comparable normalization operations
in mammalian V1, among others (Carandini and Heeger 2012). Hence, response nonlinearity
seems to pose no real threat to the prospects for mechanistic explanation.
277
5. Conclusions and future directions

I have argued that the frameworks of dynamical systems and that of mechanistic explanation
are closely related and complementary. Contrary to law- and prediction-based accounts of
dynamical explanation, there is no legitimate sense in which dynamical models explain phe-
nomena independently of describing mechanisms, either by subsumption under general laws
or by appealing to their predictive force alone. Although importantly distinct in many ways,
these two approaches to modeling are related in terms of subsumption—dynamical models
provide one important set of resources among many for discovering and describing temporal
features of a mechanism. The DST framework can often reveal or predict dynamic patterns
or latent temporal structure in the activity of mechanisms (and their components) that might
otherwise be difficult to discern using more conventional modeling techniques. Nevertheless,
the real explanatory weight of dynamical models depends on the presence of an associated
description (however incomplete) of the mechanisms that support, maintain, or underlie these
activity patterns. The frameworks of dynamics and mechanism are natural allies—each plays
an essential part in the common enterprise of describing the parts, activities, and organization
of underlying mechanisms.
In addition to the challenges for the mechanistic approach to dynamical explanation iden-
tified earlier, more hard work remains for those wishing to defend a mechanistic perspective.
For instance, dynamists have also appealed to notions of emergence or downward causation
in complex dynamical systems to argue for the limitations of the mechanistic perspective
(e.g., Chemero and Silberstein 2008; Kelso 1995; Silberstein and Chemero 2013). To date,
defenders of the mechanistic approach have not satisfactorily addressed this challenge.
Another direction for the debate over dynamical systems and mechanistic explanation
involves clarifying the roles of abstraction and idealization in dynamical and mechanistic mod-
els. Although it is widely recognized that strategies of abstraction and idealization are critically
important to many forms of scientific modeling (see Chapter 17), it remains unclear how these
are related to mechanistic and dynamical explanation. Mathematical abstraction is clearly impor-
tantly to DST, and some see this as indicating the emergence of a distinctive non-mechanistic
approach to explanation (e.g., Batterman and Rice 2014; Ross 2014). But abstraction also plays
an important role in more traditional modeling of mechanisms (e.g., Levy and Bechtel 2013).
This cluster of topics represents an undeniable growth area in contemporary philosophy of sci-
ence that stands to shed new and valuable light on the nature of the relationship between the
frameworks of dynamics and mechanism.
References
Abraham, R., Shaw, C.D. 1992. Dynamics: The Geometry of Behavior. Redwood City, CA: Addison-
Wesley.
Ahrens, M.B., Li, J.M., Orger, M.B., Robson, D.N., Schier, A.F., Engert, F., Portugues, R. 2012. “Brain-
wide neuronal dynamics during motor adaptation in zebrafish.” Nature 485(7399):471–7.
Amit, D.J. 1992. Modeling Brain Function: The World of Attractor Neural Networks. Cambridge: Cambridge
University Press.
Batterman, R., Rice, C. 2014. “Minimal model explanations.” Philosophy of Science 81(3):349–76.
Bechtel, W. 1998. “Representations and Cognitive Explanations: Assessing the Dynamicist’s Challenge in
Cognitive Science.” Cognitive Science 22(3):295–318.
Bechtel, W., Abrahamsen, A. 2002. Connectionism and the Mind. Parallel Processing, Dynamics, and Evolution
in Networks. Oxford: Blackwell.
——. 2010. “Dynamic Mechanistic Explanation: Computational Modeling of Circadian Rhythms as an
Exemplar for Cognitive Science.” Studies in History and Philosophy of Science Part A 41(3):321–33.
278
Bechtel, W., Richardson, R.C. 1993/2010. Discovering Complexity: Decomposition and Localization as
Strategies in Scientific Research. Cambridge, MA: MIT Press.
Beek, P.J., Peper, C.E., Daffertshofer, A. 2002. “Modeling Rhythmic Interlimb Coordination: Beyond
the Haken–Kelso–Bunz Model.” Brain and Cognition 48(1):149–65.
Beer, R.D. 2000. “Dynamical Approaches to Cognitive Science.” Trends in Cognitive Sciences 4(3):91–9.
Blohm, G., Keith, G.P., Crawford, J.D. 2009. “Decoding the Cortical Transformations for Visually
Guided Reaching in 3D Space.” Cerebral Cortex 19(6): 1372–93.
Bressler, S.L., Kelso, J.A.S. 2001. “Cortical Coordination Dynamics and Cognition.” Trends in Cognitive
Sciences 5(1):26–36.
Carandini, M., Heeger, D.J. 2012. “Normalization as a Canonical Neural Computation.” Nature Reviews
Neuroscience 13(1):51–62.
Chemero, A. 2011. Radical Embodied Cognitive Science. Cambridge, MA: MIT Press.
Chemero, A., Silberstein, M. 2008. “After the Philosophy of Mind: Replacing Scholasticism with Science.”
Philosophy of Science 75(1):1–27.
Churchland, M.M., Cunningham, J.P., Kaufman, M.T., Foster, J.D., Nuyujukian, P., Ryu, S.I., Shenoy,
K.V. 2012. “Neural Population Dynamics during Reaching.” Nature 487(7405):51–6.
Clark, A. 1997. Being There: Putting Brain, Body, and World Together Again. Cambridge, MA: MIT Press.
Craver, C.F. 2007. Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience. New York: Oxford
University Press.
Earman, J., Roberts, J. 1999. “Ceteris Paribus, There is no Problem of Provisos.” Synthese 118(3):439–78.
Earman, J., Roberts, J.T., Smith, S. 2002. “Ceteris Paribus Lost.” Erkenntnis 57(3):281–301.
Eliasmith, C., Anderson, C.H. 2003. Neural Engineering. Cambridge, MA: MIT Press.
Fodor, J.A. 1991. “You Can Fool Some of the People All of the Time, Everything Else Being Equal;
Hedged Laws and Psychological Explanations.” Mind 100(397):19–34.
Fuchs, A., Jirsa, V.K. 2008. Coordination: Neural, Behavioral and Social Dynamics. Vol. 1. Dordrecht: Springer.
Fuchs, A. 2013. “Haken-Kelso-Bunz (HKB) Model: Nonlinear Dynamics.” In Complex Systems: Theory
and Applications for the Life-, Neuro- and Natural Sciences. Berlin: Springer-Verlag, pp. 133–45.
Gervais, R., Weber, E. 2011. “The Covering Law Model Applied to Dynamical Cognitive Science:
A Comment on Joel Walmsley.” Minds and Machines 21(1):33–9.
Glennan, S.S. 1996. “Mechanisms and the Nature of Causation.” Erkenntnis 44(1):50–71.
Haken, H., Kelso, J.A.S., Bunz. H. 1985. “A Theoretical Model of Phase Transitions in Human Hand
Movements.” Biological Cybernetics 51(5):347–56.
Hempel, C.G. 1965. Aspects of Scientific Explanation. New York: The Free Press.
Hinton, G., McClelland, J.L., Rumelhart, D.E. 1986. “Distributed Representations.” In Parallel
Distributed Representations: Explorations in the Microstructure of Cognition, edited by E. Rumelhart and
J.L. McClelland. Cambridge, MA: MIT Press, pp. 77–109.
Izhikevich, E.M. 2005. Dynamical Systems in Neuroscience. Cambridge, MA: MIT Press.
Kaplan, D.M. 2015. “Moving Parts: The Natural Alliance between Dynamical and Mechanistic Modeling
Approaches.” Biology and Philosophy 30:757–86.
Kaplan, D.M., Bechtel, W. 2011. “Dynamical Models: An Alternative or Complement to Mechanistic
Explanations?” Topics in Cognitive Science 3(2):438–44.
Kaplan, D.M, Craver, C.F. 2011. “The Explanatory Force of Dynamical and Mathematical Models in
Neuroscience: A Mechanistic Perspective.” Philosophy of Science 78(4):601–27.
Kelso, J.A.S. 1995. Dynamic Patterns: The Self-Organization of Brain and Behavior. Cambridge, MA: MIT
Press.
Levy, A., Bechtel, W. 2013. “Abstraction and the Organization of Mechanisms.” Philosophy of Science
80(2):241–61.
Machamer, P., Darden, L, Craver, C.F. 2000. “Thinking about Mechanisms.” Philosophy of Science
67(1):1–25.
Mascagni, M.V., Sherman, A.S. 1989. “Numerical Methods for Neuronal Modeling.” Methods in Neuronal
Modeling. Cambridge, MA: MIT Press.
Olsen, S.R., Bhandawat, V., Wilson, R.I. 2010. “Divisive Normalization in Olfactory Population Codes.”
Neuron 66(2):287–99.
Oullier, O., de Guzman, G.C., Jantzen, K.J., Lagarde, J., Kelso, J.A.S. 2008. “Social Coordination
Dynamics: Measuring Human Bonding.” Social Neuroscience 3(2):178–92.
Pietroski, P., Rey, G. 1995. “When Other Things Aren’t Equal: Saving Ceteris Paribus Laws from
Vacuity.” British Journal for the Philosophy of Science 46(1):81–110.
279
Reutlinger, A., Unterhuber, M. 2014. “Thinking about Non-Universal Laws.” Erkenntnis 79 (10):
1703–13.
Rosenbaum, D.A. 1998. “Is Dynamical Systems Modeling Just Curve Fitting?” Motor Control 2(2):101–4.
Ross, L.N. (2014). “Dynamical Models and Explanation in Neuroscience.” Philosophy of Science 81(1):32–54.
Salmon, W.C. 2006. Four Decades of Scientific Explanation. Pittsburgh, PA: University of Pittsburgh Press.
Schmidt, R.C., Carello, C., Turvey, M.T. 1990. “Phase Transitions and Critical Fluctuations in the Visual
Coordination of Rhythmic Movements between People.” Journal of Experimental Psychology: Human
Perception and Performance 16(2):227–47.
Schöner, G., Kelso, J.A.S. 1988. “Dynamic Pattern Generation in Behavioral and Neural Systems.” Science
239(4847):1513–20.
Sejnowski, T.J., Rosenberg, C.R. 1987. “Parallel Networks that Learn to Pronounce English Text.”
Complex Systems: 145–68.
Silberstein, M., Chemero, A. 2013. “Constraints on Localization and Decomposition as Explanatory
Strategies in the Biological Sciences” Philosophy of Science 80(5):958–70.
Stepp, N., Chemero, A., Turvey, M.T. 2011. “Philosophy for the Rest of Cognitive Science.” Topics in
Cognitive Science 3(2):425–37
Strogatz, S.H. 1994. Nonlinear Dynamics and Chaos: With Application to Physics, Biology, Chemistry, and
Engineering. Boulder, CO: Westview Press.
Swinnen, S.P. 2002. “Intermanual Coordination: From Behavioural Principles to Neural-Network
Interactions.” Nature Reviews Neuroscience 3(5):348–59.
Van Gelder, T. 1995. “What Might Cognition Be, If Not Computation?” The Journal of Philosophy
92(7):345–81.
——. 1998. “The Dynamical Hypothesis in Cognitive Science.” Behavioral and Brain Sciences 21(5):615–28.
Walmsley, J. 2008. “Explanation in Dynamical Cognitive Science.” Minds and Machines 18(3):331–48.
Woodward, J. 2000. “Explanation and Invariance in the Special Sciences.” British Journal for the Philosophy
of Science 51(2):197–254.
——. 2002. “There Is No Such Thing as a Ceteris Paribus Law.” Erkenntnis 57(3):303–28.
——. 2003. Making Things Happen: A Theory of Causal Explanation. Oxford: Oxford University Press.
Zednik, C. 2011. “The Nature of Dynamical Explanation.” Philosophy of Science 78(2):238–63.
Zipser, D., Andersen, R.A. 1988. “A Back-Propagation Programmed Network that Simulates Response
Properties of a Subset of Posterior Parietal Neurons.” Nature 331(6158): 679–84.
280
PART IV
Disciplinary perspectives on
mechanisms
21
MECHANISMS IN PHYSICS1
Meinard Kuhlmann
1. Introduction
Since mechanisms became a central topic in contemporary philosophy of science some 15 years
ago, physics has never been in the focus of larger debates. Prima facie, this is understandable in
the face of one major motivation for today’s mechanistic program: The traditional twentieth-
century philosophy of science with its roots in logical empiricism was dominated by physics as
the paradigmatic science. In physics, theories with universal laws seem to be center stage. They
are the basis for predictions, explanations, and our general understanding of nature. However,
this law-focused philosophy of science seems much less suited when it comes to special sci-
ences such as biology, neuro-science, and psychology—leaving aside the question of whether
it is appropriate for physics itself. In the special sciences, mechanisms play a far greater role than
universal laws, if indeed there are any such laws at all. With the growing attention paid to spe-
cial sciences in the philosophy of science, mechanisms have accordingly become a central issue.
Today, with the ascent of the new mechanistic program in philosophy of science, the question
has almost turned: How does physics fit into the picture?
When thinking about mechanisms in physics, the most interesting question is different
than for other fields. The primary concern is not how the notion of mechanisms can be cap-
tured and how exactly mechanistic explanations work. Rather, the main question is to what
extent physics, in particular fundamental physics, deals with mechanisms in the first place.
A second question concerns whether the character of the physical processes that underlie all
natural and social phenomena may even endanger the tenability of mechanistic reasoning in
the special sciences.
Physics can be important for mechanisms on higher levels, like those in chemistry or biology.
However, is physics itself concerned with mechanisms? Here are some reasons why physics or
parts of physics may not be dealing with mechanisms:
a) One may argue that fundamental physics rests on the brute fact that one observes certain
regularities, which physical theories organize in a systematic way. Thus no explanation
of the observed regularities is given, and thus a fortiori no mechanistic explanation. If
this should be true then it seems that there is no room for mechanistic explanations in
fundamental physics.
283
Meinard Kuhlmann
b) At least some explanations in physics are non-causal and thereby arguably not mechanistic.
Examples are explanations based on energy conservation considerations or dimensional
analysis. These cases may not be representative, however. Moreover, arguably these often
very elegant explanations are only convincing provided that one also has (more compli-
cated) micro explanations of the same type of phenomena.
c) Only very recently, a number of philosophers of science have argued that the range of
non-causal explanations may be greater than has been thought previously. In particular, it
may also include important classes of micro explanations in physics, such as explanations
in statistical physics and specifically so-called renormalization group explanations. In our
context this debate is of particular importance because it concerns explanations which seem
mechanistic. If true, this line of thought could mean that mechanistic explanations are not
only absent in fundamental physics but also in higher levels.
d) To make things worse for the new mechanists, one could even argue that the specific nature
of quantum physics undermines the mechanical philosophy concerning all higher levels, i.e.
even regarding phenomena that are not studied in physics but in the special sciences.
These potential limitations to mechanistic reasoning in physics bring us to three interesting

questions:
1) Do mechanisms play any role in fundamental physics?

2) To what extent are explanations in the non-fundamental part of physics mechanistic?
3) Does quantum physics undermine mechanistic reasoning on higher levels?
As we will see, none of these three questions has an easy, straightforward answer. Since the
answers depend crucially on which parts of physics one is looking at, the structure of this chapter
largely follows the structure of physics as a discipline.
2. Fundamental physics: general reflections

Presupposing for now that there is a fundamental physics,2 it is natural to assume that it com-
prises those theories that state what the fundamental physical entities and the fundamental laws
are. The laws reflect how the entities behave, in isolation as well as when they interact with
each other. Advanced accounts usually describe the fundamental entities in terms of perma-
nent properties, which are given by the invariants under certain symmetry transformations, i.e.
changes that only affect how we describe the world but not what is described. Which symmetry
transformations are actually supposed to be the ones that leave the physical reality unchanged
is itself a vital part of our theories: The symmetries that characterize Classical Mechanics differ
from the ones for Special Relativity Theory.
We do not know the true fundamental physics yet, at least not in full, and we may never do
so. The considerations discussed in this section are, however, general enough that they should
not depend much on the state of the art in current fundamental physics. This changes when
we consider specific theories, as we will do in the subsequent sections. There, one must keep
in mind that we are just dealing with previous or current candidates for fundamental physics.
Although these theories are provisional or even proven wrong, it is still worth exploring how
they fit into the mechanistic program. First, these theories have shaped our worldview, and
unveiling the role of mechanisms in them helps us to clarify this picture. Second, exploring the
occurrence or non-occurrence of mechanisms in today’s physics may contribute to finding or
evaluating candidates for future fundamental physics.
284
Mechanisms in physics
Within the modern debate on mechanisms, the special status of fundamental physics for
the mechanistic program was put on record early on. Among the advocates of this pro-
gram, Glennan is arguably the one who has most persistently pointed to this issue. Glennan
(1996: 50) notes:
[That] two bodies . . . gravitationally attract each other . . . is just a “brute fact” about
the world in which we live. . . . [W]e cannot explain the interaction by reference to
any mechanism. . . . I suggest that there should be a dichotomy in our understand-
ing of causation between the case of fundamental physics and that of other sciences
(including much of physics itself).
Glennan (2002: S348) adds:
While most laws are mechanically explicable, inevitably there must be some laws that
are not. For instance . . . Maxwell’s equations are not mechanically explicable. There is
not, for instance, a mechanical ether consisting of particles whose interactions could
explain the propagation of electromagnetic waves. Laws such as these, which I call
fundamental laws, represent brute nomological facts of our universe.
(My emphases, MK)
This view is close to the one I listed as argument a) above.3 The lack of mechanistic explana-
tions for fundamental interactions has nothing to do with the nature of the specific theories in
place. If there were a mechanistic explanation for a supposedly fundamental law, then the laws
governing the interacting parts of this mechanism would be the truly fundamental laws and
not the one we started with. These laws in turn have no lower mechanistic basis, for other-
wise one could invoke the same argument again. The upshot is that there are no mechanisms
in fundamental physics, but only in higher-level sciences, including the non-fundamental
parts of physics.4
Recently, Glennan has adopted a somewhat different take on this issue:
Our discussion of this problem will be facilitated by some new terminology. Let us call
a mechanism (that is, a set of entities engaging in activities and interactions) that does
not depend upon other mechanisms, a fundamental (or basic) mechanism. All other
mechanisms are mechanism-dependent or compound mechanisms.
(Glennan forthcoming, sec. 7.5)
While in the original view, mechanisms played no role in fundamental physics, now it abounds
with mechanisms. Glennan comments on this change as follows:
In past work (Glennan, 1996, 2002, 2011) I have simply spoken of fundamental
interactions rather than saying they are interactions produced by fundamental mecha-
nisms, so this language represents a terminological shift. This shift does not affect
the main substantive claim, which is that if there are such fundamental mechanisms,
there is a base case of mechanisms whose productive capacities are not themselves
mechanism-dependent. . . . The advantage of saying these interactions are produced
by fundamental mechanisms is that then every causally related pair of events will be
connected by a mechanism.
(Glennan forthcoming, chap. 7, footnote 8)
285
Meinard Kuhlmann
One may question whether this is really just a “terminological shift”; it seems at least a shift in
emphasis from epistemology toward ontology. But, in any case, it gives us an attractive coher-
ent picture: The interactions described by the laws of fundamental physics cannot be explained
mechanistically. However, this does not mean that there is no room for mechanisms in fun-
damental physics, because the fundamental interactions are themselves mechanisms. The only
aspect that distinguishes them from mechanisms in higher-level sciences is that they are basic
mechanisms, since they do not rest on further mechanisms on a lower level.
If we accept this line of reasoning, one crucial question then is which basic mechanisms there
are in fundamental physics. Let us start with a straightforward, somewhat naïve list of—old or
current—candidates of basic mechanisms in fundamental physics:
1) Newtonian physics: gravitational attraction as action-at-a-distance

2) Special Relativity Theory (SRT): length contraction and time dilation
3) General Relativity Theory (GRT): gravitational attraction in terms of the curvature of
space-time
4) Classical Electrodynamics: electromagnetic interaction
5) Fermi theory: weak interaction
6) Quantum Electrodynamics (QED): quantum version of electromagnetic interaction
7) “Standard Model” of elementary particle physics:5
(7.1) Salam-Glashow-Weinberg theory: electroweak interaction (unified weak and elec-
tromagnetic interaction)
(7.2) Quantum Chromodynamics (QCD): strong interaction
8) Quantum Gravity (String theory or Loop Quantum Gravity): all four forces in a unified way.
That some interactions appear repeatedly in the above list is mainly because they are dealt with
in different ways in subsequent theories, such as Newtonian physics and Relativity Theory, or
classical and quantum theories. Also in the background is the concept of “effective field theo-
ries,” which I will discuss in more detail below.
We have seen above that there can be no mechanistic explanations for fundamental interac-
tions. However, for a number of independent reasons, mechanistic reasoning may nevertheless
play a role in specific parts of fundamental physics. To see why this possibility is not in contra-
diction with the above general line of argument, I propose to distinguish four sorts of entities
that physical theory seeks to characterize:
i) Fundamental laws (at their time): Newton’s law of gravitation, Maxwell’s equations,
Einstein’s field equations in General Relativity Theory, the Schrödinger equation, the field
equation in Quantum Electrodynamics.
ii) Phenomena that are immediate consequences of fundamental laws: Electromagnetic waves (e.g. visible
light) in Maxwell’s Theory, Heisenberg’s uncertainty relations in Quantum Mechanics, black
holes in General Relativity Theory, pair creation in Quantum Field Theory.
iii) Basic composite systems: Composite systems whose immediate constituents are fundamental.
iv) Higher-order composite systems: Composite systems whose constituents are themselves com-
posite systems.
In actual cases, the distinction between (i) and (ii) as well as that between (iii) and (iv) may be
controversial. Nevertheless, I think in principle it is comprehensible and justified and it has
an important bearing on our present question, since immediate consequences of fundamental
286
laws, i.e. (ii), make up large parts of fundamental physics, and the treatment of higher-order
composite systems (iv) is also partly a topic in fundamental physics. In the following I will
consider to what extent mechanistic reasoning is in fact relevant in the different parts of
3. Fundamental physics: relativity theory and quantum physics

In Classical Mechanics, forces remain largely unspecified. Gravitation is the only force for
which we have a specific fundamental law. We argued above that gravitational attraction is—
ontologically—a basic mechanism, but it cannot be explained mechanistically. The same applies
to electromagnetic forces. In Special and General Relativity the situation is less straightforward.
Length contraction and time dilation of speeding objects are two closely connected phenom-
ena, which are crucial for the development of twentieth-century physics. What is the nature of
the explanation of these phenomena? Is there a mechanism that leads to the decreased length
of a speeding spaceship? The received view today is that the correct explanation in terms of
Einstein’s Special Theory of Relativity from 1905 is not mechanistic. However, historically this
was not the only answer. The Dutch physicist and 1902 Nobel Prize winner Hendrik Lorentz
famously proposed a very different theory. Moreover, reconsidering this alternative proposal,
one may question the orthodox view of Special Relativity, too (see Brown 2005). This turns the
question of whether the correct explanation of length contraction and time dilation is mechanis-
tic into a controversial one. Even if there should not be any doubt about the correct answer in
the end, the discussion is still worth pursuing because it sheds an interesting light on the general
question concerning the scope of mechanistic explanations in physics.
The crucial aspect about the Lorentz theory in our context is that it supplies a dynamical
account of length contraction. The length of a speeding spaceship has really decreased. So why
does it not get tight in the spaceship?—Because everything in the spaceship got smaller, too.
In particular, all measuring devices like rods will also be smaller. It is like a big conspiracy:
Everything has changed but we do not notice this change because everything we could use
to detect this change has itself changed exactly such that measurements will not tell us of any
change. The conspiracy flavor is one reason why most physicists have rejected the Lorentz
theory. Nevertheless, it has its attraction and that is probably because it tells a dynamical and
thereby potentially a mechanistic story of why relativistic effects occur. Spaceships and rods shrink
because of how their micro constituents behave and of course interact.
In Einstein’s Special Theory of Relativity the explanation of length contraction and time
dilation is of a completely different nature. It is kinematical, i.e. it has to do with the very geometry
of space-time and not with anything specific about the behavior of objects due to their proper-
ties and their interaction with physical fields. As Janssen (2009: 28) puts it:
Special relativity is completely agnostic about what inhabits . . . Minkowski space-time.

All the theory has to say about systems inhabiting/carrying Minkowski space-time is
that their spatio-temporal behavior must be in accordance with the rules it encodes.
Special relativity thus imposes the kinematical constraint that all dynamical laws must
be Lorentz invariant. This property transcends the individual laws.
Janssen nicely compares the geometry of space-time with the rectangularity of paintings,
which puts certain restrictions on paintings and which is realized in paintings, but which has
nothing to do with the nature of the individual paintings. The upshot is that there are certain
kinds of explanations in fundamental physics that genuinely show why certain phenomena
287
Meinard Kuhlmann
occur (as opposed to just describing what occurs), but which are not in any way mechanistic.
Special Relativity really explains length contraction and time dilation but the explanation
does not rest on any physical processes in space and time but instead on the geometrical
properties of space-time itself, regardless of any specific objects in space-time. Felline (2015)
proposes to rate this as a “structural explanation” because its rests on the geometrical structure
of space-time rather than on any mechanisms in space-time. Moreover, Felline argues, this
shows that a lack of mechanistic explanations in fundamental physics does not entail a lack
of explanations tout court.
Examples such as the one we just discussed, and many others, are of paramount importance
in a recent debate on non-causal explanations because it is surprisingly hard to draw a clear
line between causal and non-causal explanations, and the way they work, in a general way
(see Baker 2005; Batterman 2010; Lange 2013; Skow 2014; Pincock 2015). Counting only
explanations as causal that explicitly cite causes would arguably be too restrictive. Therefore,
Lange (2013: 493) proposes the broader notion that an explanation is causal if “it explains
by virtue of describing contextually relevant features of the result’s causal history or, more
broadly, of the world’s network of causal relations.” In the light of this characterization, the
way in which SRT explains why the Lorentz transformations hold, or more specifically, the
occurrence of length contraction and time dilation is clearly non-causal. The explanation
rests on spatiotemporal symmetries and the principle of relativity, and these do not describe
“the world’s network of causal relations,” but rather impose constraints on the (causal) laws
of nature (Lange 2013: 494). This is exactly what Janssen metaphorically compared with the
rectangularity of paintings.
The situation seems different when we consider gravitation in Einstein’s General Theory of
Relativity (GRT): Masses affect the curvature of space-time. This curvature in turn causes bodies
to move in a certain way. This behavior appears to be due to gravitational forces between bodies
but in reality they just follow their geodesics, i.e. the shortest way from A to B, given the curved
space-time in which they live. Thus it seems reasonable to say that GRT explains the apparently
gravitational behavior of massive bodies in a mechanistic fashion, namely by the interaction
between large masses, the bendable space-time and massive bodies in question.
Now let us turn our attention to quantum physics. The most serious obstacle for the inter-
pretation of quantum mechanics is the fact that the standard formalism is completely silent about
the dynamical process that leads to an observable measurement outcome. Apparently, such a
process would have to be one that leads from an initial state of the composite system of measure-
ment apparatus and object to be measured via their interaction to a final determinate state of (at
least) the pointer of the measurement apparatus. Therefore, one could rephrase our first state-
ment by saying it is the lack of a mechanism leading to determinate measurement outcomes that
blocks the interpretation of standard quantum mechanics. Note that physically there is nothing
specific about measurements in a laboratory. The problem is generic. In general the dynami-
cal law of QM does not lead to determinate values, in contrast to the apparent occurrence of
classical properties everywhere in nature, be it in the laboratory or in the wild. However, it is
customary to phrase the problem in terms of measurements. Alternatively, one could say that
quantum measurements happen everywhere and all the time, i.e. not only in the laboratory.
Using the famous Schrödinger cat example: The cat itself is the measurement apparatus already.
There is no need for an experimenter who observes the result. In particular, it has nothing to
do with a conscious observer.6
Unfortunately, whatever the sought-after mechanism that leads to determinate “measure-
ment” outcomes would look like, it seems inevitable that it is in conflict with the standard
assumption that the Schrödinger equation specifies the universally valid dynamical law for the
288
time evolution of quantum systems. The reason is that there is usually a whole spectrum of
different possible measurement outcomes and the linear Schrödinger time evolution cannot
accommodate the fact that in a measurement process the spectrum of possibilities seems to col-
lapse into one actual outcome.
Thus in the context of the standard Hilbert space formalism of quantum mechanics, the
prediction of certain measurement outcomes does not have any mechanistic explanation.
However, this is one of the issues in modern physics that count as not completely understood.
And every approach that is a candidate for solving the measurement problem either supplies
a mechanistic explanation of definite measurement outcomes—where mechanism is under-
stood in the sense of, e.g., Machamer, Darden, and Craver (2000) and Glennan (1996, 2002,
forthcoming)—or shows that there is some physical process that leads to it—which fits with
Dowe’s (2000) notion of mechanisms.
The oldest proposal for solving this conflict is to simply postulate that there is a second kind
of dynamical law that is only valid for the apparent collapse in a measurement process. The
disadvantage of this move is that it remains unclear why measurement processes should differ
so fundamentally from other physical processes. Other proposals are more creative, but they all
agree in that it is inevitable to give up some dear belief (see Maudlin 1995). Ghirardi, Rimini,
and Weber give up the standard linear Schrödinger equation and replace it with a stochastic
version—paying the price of introducing ad hoc two new natural constants, which cannot be
measured independently. Bohm gives up the assumption that the quantum state is complete and
uses the always determinate particle positions as the additional hidden parameters—but in doing
so uses a preferred frame, which causes a conflict with Special Relativity. The many-worlds
interpretation takes the plurality of possible measurement outcomes as reality, but only in dif-
ferent causally separated worlds. Thus it gives up the assumption that a measurement process
actually leads to one determinate measurement outcome. The many-worlds interpretation, too,
has big downsides, however: the way it tries to account for the probabilities for different meas-
urement outcomes in terms of decision theory is highly controversial. Moreover, it severely
violates the principle of ontological parsimony.
What all these proposals have in common is that they try to give—in very different ways—a
mechanistic explanation for the occurrence of determinate measurement outcomes. Since each
of these proposals has its merits and advocates but also serious drawbacks, we currently have a
certain stalemate. However, there is at least something that most people agree upon as being at
least a part of the solution for the conflict between quantum mechanics and the classical appear-
ance of our world: namely, decoherence. In Kuhlmann and Glennan (2014) we argue that
decoherence is of crucial importance for showing that quantum mechanics does not generally
undermine the “new mechanical philosophy.”
One of the outstanding features of decoherence and arguably one main reason why
it attracts attention from so many very different camps is that it is a real physical process.
Schlosshauer (2007) offers a comprehensive exposition of the theoretical and experimental
aspects of decoherence. It is not a postulate, a story, or an ad hoc addition but an observable
and analyzable mechanism. Moreover, one can fully describe this mechanism within the limits
of QM. To be sure, decoherence does not solve the quantum measurement problem, but it
is discernible how it may contribute at least something to its solution—in fact, whereas it is
essential for the many-worlds interpretation, the role of decoherence is not so straightforward
for other interpretations. The main idea of decoherence is that, because of a suitable interac-
tion of a quantum system with a macroscopic environment, the troublesome superpositions
(e.g. of a dead and a live cat) are effectively no longer observable even though they do
not in fact disappear. This non-observability is not just an illusion; it is grounded in solid
289
Meinard Kuhlmann
physical circumstances. In principle, the superpositions are not only still there but they are
even enhanced by decoherence. However, they are spread outside of the system in question
such that it is virtually impossible to recover them in a detectable manner.
4. Fundamental physics of today

The list of candidates for basic mechanisms in fundamental physics (in section 1) contained
numerous duplications. As we saw, many of these duplications are due to the transitions either
from Newtonian physics to Relativity Theory or from classical to quantum theories. However,
there is also another explanation of these duplications, which is less known to non-physicists but
which is crucial for contemporary physics. Most of the above theories are regarded as so-called
“effective field theories” today. Hartmann (2001) and Castellani (2002) discuss effective field
theories from a philosophical point of view, in particular in the context of reduction. Very gen-
erally, physicists call some quantity X “effective” if it is what we actually observe, while there is
some underlying quantity, let’s call it Y, that can have completely different characteristics. For
instance, the underlying contributions to a potential can be of a very different kind than the
overall potential we effectively measure.
One particularly important “mechanism”7 that brings us from the underlying hidden Y to
the observable X is so-called “spontaneous symmetry breaking”—a notion that was explicitly
introduced only in the 1960s and turned out to be crucial in a huge spectrum of otherwise
largely unrelated contexts. The key idea of spontaneous symmetry breaking is that we can have
underlying laws (our “Y”) that are symmetric, whereas the ground state, which we observe (i.e.
the “X”), is not symmetric.
In effective field theories the X we “observe”8 is one particular field theory, e.g. Quantum
Electrodynamics (QED) of the Standard Model, and the underlying Y can be something that
is not even a field theory at all, like string theory. So why do we “see” the electromagnetic
quantum field and not the supposedly underlying theory? The reason is the energy regime that
is relevant in the world we are dealing with today. The conditions in our world are quite dif-
ferent from the ones after the big bang because the universe has cooled down considerably. In
this cooling spontaneous symmetry breaking occurred repeatedly, so that we do not directly
see the “true physics” any more. So the idea is that the laws as such have not changed, but we
effectively see different laws for different temperatures.
Obviously these considerations have an immediate bearing for our search of basic mecha-
nisms. The effective field theory lesson would be that there are no basic mechanisms simpliciter
but only basic mechanisms for a given energy regime. For anyone but physicists working on
fundamental theories, this is good enough, of course, because we have no handle on the energy
regime we live in today anyway.
Another important—although yet speculative—case of an effective field is General Relativity
Theory (GRT). Famously GRT is a field theory, i.e. space and time are at the root of the
theory. However, space and time may prove to be only “effective” or emergent entities: in a
future theory of Quantum Gravity, space and time may no longer be basic notions but rather
macroscopic notions that emerge in an approximate way from the underlying microscopic
theory. If this speculation should turn out to be true and if we think that mechanisms are
spatiotemporal entities, then it seems possible that there is no room for basic mechanisms in
Thus, effective field theories may pose two threats to the mechanistic program. The first
threat is that there is no most fundamental theory and therefore no basic mechanisms. In reaction
290
to this threat one may simply welcome the infinite tower of effective field theories and say that
every theory has a mechanistic grounding, namely in terms of the respective underlying theory on
another energy regime.
The second threat is that space and time no longer appear in fundamental physics but
emerge as macroscopic notions. Therefore, fundamental physics would not be dealing with any
mechanisms—provided one understands mechanisms as spatiotemporal entities. Two possible
reactions by mechanists come to mind. One possibility would be to reconceptualize mecha-
nisms in a way that they do not need to be spatiotemporal. An alternative reaction would be
that emergent space-time is not really a threat for the very notion of basic mechanisms because
basic mechanisms simply appear only on one level higher where space-time is already manifest.
5. Physics of composite systems

Two things may block mechanistic explanations for composite systems. First, the composition
laws of quantum systems could preclude mechanistic explanations for the behavior of composite
systems. Second, complex systems dynamics may be incompatible with the micro-reductive
nature of mechanistic explanations, roughly to the same extent that complex systems exhibit
emergent behavior. I will argue that both concerns should disappear on closer scrutiny.
Let us turn to the first concern. Although classical and not quantum physics is used for
building cars and refrigerators, the universal validity of quantum physics casts doubts on the
theoretical legitimacy of mechanistic explanations even on macroscopic scales. Of course, any
consideration on this issue ultimately hinges on the pending solution of the quantum meas-
urement problem. However, even before this is accomplished, one may partially alleviate the
concern that QM undermines mechanistic reasoning. In Kuhlmann and Glennan (2014) we
identify three features of quantum mechanics that could conflict with the mechanistic approach,
namely (A) indeterminacy of properties, (B) non-localizability of quantum objects, and
(C) quantum holism. Although (B) is just a special case of (A), it deserves separate attention, first,
because well-localized parts in particular may appear prerequisite for mechanisms and, second,
because one can dissolve worry (B) even without a general solution to (A). The main point is
that causal and not spatiotemporal organization is crucial for mechanisms.9
Quantum holism (C) is another issue that may endanger a mechanistic approach to com-
posite systems. Since the mechanistic conception of explanation rests on the reductionist
idea that the behavior of a composite system is reducible to the activities and interactions of
its parts, it may seem that the failure of reductionism due to quantum holism may infect the
mechanistic program too. However, this is not the case because mechanistic explanations
are concerned with the dynamics of compound systems and not with the question whether
the states of the subsystems determine the state of the compound system at a given time
(cf. Hüttemann 2005). To determine how a compound quantum system evolves in time, all
we need to know are the relevant terms for the subsystems and their respective interactions,
and these terms are simply added up.
We have argued in section 3 that as we approach macroscopic scales, mechanistic explana-
tions usually become less problematic. However, here too, things are not as straightforward as
one might expect. Arguably the most interesting issue is dynamical complexity, which occurs
in two flavors (see Strevens 2016). First, simple macro behavior can result from convoluted micro
dynamics. Second, simple micro dynamics can produce surprisingly intricate macro behavior.10
Both cases do not appear to be amenable to mechanistic reasoning because micro dynamics
and macro behavior do not fit with each other: In the first case, the interactions between the
291
Meinard Kuhlmann
system’s parts seem to be too chaotic to explain the simple behavior of the macro variables.
In the second case, the reverse obtains, since the interactions between the system’s parts seem to
be too simple to explain the non-trivial macro behavior.
Let us take a closer look at both cases. The first case characterizes quite generally the way
in which statistical mechanics underlies thermodynamics: thermodynamics deals with the rela-
tion of various macroscopic quantities such as pressure, density, and temperature. One crucial
difference between (classical) thermodynamics and statistical mechanics is that only the latter
deals with probabilities. This fact not only represents a qualitative difference between micro
dynamics (statistical mechanics) and macro behavior (thermodynamics), but also opens a venue
for bridging the gap: convoluted micro dynamics can lead to simple macro behavior because
only certain probabilistic features of the micro dynamics matter.11 Whether statistical mechan-
ics (micro dynamics) actually does the trick of explaining thermodynamical macro behavior
depends on whether thermodynamics can be reduced to statistical mechanics.12 Strevens (2016,
section 2.3) proposes a so-called enion probability analysis which shows how “probability distri-
butions over microlevel properties . . . can be used to explain stabilities or simplicities in a gas’s
macrolevel properties.”
But is this a mechanistic explanation? While Strevens is silent on this matter, Weber and
Lefevere (2014) propose a negative answer. They rate the above explanation (not specifically
Strevens’ account of it) as an “aggregation explanation,” which they see as a different type of
micro-explanation than mechanistic explanation:
We use the term aggregation explanation to refer to explanations which explain the
behaviour of the macro-systems . . . by means of the component parts, the activities of
these parts, organisation and random distribution of behaviour of components. This means
that, compared to mechanistic explanations, aggregation explanations have an extra
ingredient [so that they should be rated as] non-mechanistic micro-explanations.
(2014: 43f)
However, if randomness is merely an extra ingredient, then why should aggregation expla-
nations no longer be mechanistic? It seems more natural to say that they are a sub-class of
mechanistic explanations (see Chapters 7 and 13).
Finally, let us consider complex systems that have non-trivial macro behavior despite their
simple micro dynamics (see Chapter 20 for biological examples). One example is the formation
of macroscopic convection cells in a container with a viscous fluid that is uniformly heated from
below. Not a few rate such self-organized coordination as emergent behavior, and then argue
that this makes mechanistic reasoning inapplicable.13 For complex systems, the occurrence of
emergent behavior has been claimed in various overlapping contexts. The most important ones
involve the non-linear connection between micro causes and macro behavior (resulting in
chaotic behavior), feedback with top-down causation, “enslavement” in Synergetics, the dynami-
cal role of the so-called order parameter (e.g. in superconductivity), and universal behavior in
phase transitions across diverse materials.14 In claims that these phenomena in complex systems
are emergent and therefore defy mechanistic explanation, the general idea is roughly this: It is
not the arrangement of the micro details that accounts for the behavior of the whole system but
rather something that only emerges when the composite system evolves, and which acquires a
certain life of its own. To show what this can mean in more concrete terms, I want to pick out
two issues, namely universal behavior (or short “universality”) and enslavement.
Universal behavior arises in phase transitions, for instance from a liquid to a vaporous state (or
phase) or from an unmagnetized to a ferromagnetic state of a piece of metal. Phase transitions in
292
radically diverse materials can exhibit remarkably similar or even identical properties. Numerically,
universality shows up in (almost) identical “critical exponents,” which characterize a system’s
behavior in phase transitions. Whereas similar behavior of materially diverse tables is easy to explain
by reference to their geometry, universal behavior in phase transitions is more intricate. A satis-
factory explanation was only given in the 1970s by Kenneth Wilson’s “renormalization group
method.” Batterman (2002, 2010) argues that renormalization group explanations are a genuine
new type of explanation (involving so-called “infinite idealizations”), which are not mechanistic (at
least not in the traditional sense) because their essence consists in “systematic methods for throwing
details away” (Batterman 2002: 18), i.e. exactly those things that seem essential for a mechanistic
approach. This claim prompted a number of critical responses (Butterfield 2011; Norton 2012) as
well as largely sympathetic ones (Morrison 2012). In any case, it initiated a fruitful debate that is
by no means finished. My own view is that renormalization group explanations should be rated as
causal and, more specifically, as mechanistic because showing that most micro details are irrelevant
does not mean that none are relevant—in accordance with Lange’s (2013: 493) above-mentioned
broader notion of causal explanation. The procedure shows us exactly which (structural) aspects
of the micro details do matter. In Kuhlmann (2014) I then argue—in a related context—that one
needs a notion of structural mechanisms to account for this fact. In this account, the explanatory
weight rests on certain structural features of the dynamical interactive organization and not on the
details of the initial set-up.
6. Summing up
One naturally assumes that physics gives us the ultimate basis for mechanistic reasoning in the
special sciences. However, it is not evident to what extent physics itself deals with mechanisms.
One reason for the potential lack of mechanisms is that on the fundamental level mechanistic
explanations come to a natural end. Arguably the most coherent way to deal with this fact is to
say that fundamental physics gives us the basic mechanisms, which have no further mechanistic
grounding. One recent development in physics that could endanger this view is the notion of
effective field theories, according to which there may be no most fundamental level. Moreover,
space and time, which seem required for mechanisms to operate in, may be emergent rather
than basic features of the universe. Apart from these more recent developments, conventional
quantum physics already may undermine mechanistic reasoning for a number of reasons. Finally,
the mechanistic approach seems to be at odds with the behavior of complex systems, because
either the convolution of micro details prevents the explanation of simple macro behavior, or
the simple organization on the micro level leaves the intricate macro behavior inexplicable. I
have shown that one can dispel or at least alleviate each of these worries.
Notes
1 I wish to thank Elena Castellani, Roman Frigg, and Carl Hoefer, as well as the editors Stuart Glennan
and Phyllis Illari for many valuable comments on an earlier draft.
2 As we will see below, this is controversial. The threat are so-called “effective field theories” which could
result in an infinite cascade of ever more fundamental theories.
3 The difference is that Glennan explicitly rejects the regularity account of laws and causation. In our
context, the inexplicability of fundamental laws, this difference does not matter.
4 Felline (forthcoming) argues that the lack of a mechanistic underpinning for fundamental phenomena
does not necessitate an ancillary account of causation for the fundamental level because according to
her—contrary to widespread opinion—the causal anti-fundamentalism in Glennan’s approach is in fact
compatible with the idea that there is no causality at the fundamental level.
293
Meinard Kuhlmann
5 The so-called standard model of particle physics from the 1980s is still the valid theory of elementary
particles and interactions. In this model the “Higgs mechanism” plays a crucial parts because—
according to common gloss—it explains why particles have mass. However, on closer scrutiny the
“Higgs mechanism” does not fulfill the requirements for being a mechanism in several respects.
Instead, the “Higgs mechanism” rather seems to be the core ingredient of a non-causal explanation
(see Kuhlmann forthcoming).
6 In the early period of thinking about the quantum measurement problem it was quite common to see
the act of a conscious observer as an indispensable ingredient. Partly because of the formulation of various
potential solutions that do “Quantum Theory Without Observers” (BI), there is hardly anyone today in
the relevant scientific community who assigns any physical significance to a conscious observer.
7 Below I will address the question of whether—or under which conditions—this is really a mechanism.
8 One cannot observe a theory, of course. What I mean by this way of talking is that the world appears to
us as if this was the correct underlying theory.
9 In Kuhlmann (2015) I show in detail what this means for the quantum-mechanical explanation of
laser light.
10 Note that often only this second case goes under the title “complex systems.”
11 Note that the role of probabilities depends crucially on the framework. While they are essential in the
Gibbsian approach, they are something like epiphenomena in the Boltzmannian approach. See Frigg
(2008) for details.
12 See Sklar (2015, section 6) for a non-technical introduction, and Dizadji-Bahmani, Frigg, and Hartmann
(2010) for a defense of the reductive account.
13 To my knowledge Broad (1925, chapter 2) introduced the distinction of mechanistic versus emergent
explanations. In any case, his discussion was arguably the most influential one.
14 Batterman (2002, 2010) and Morrison (2012, 2014) are among the most active contributors to this
debate.
References
Baker, A. (2005) “Are There Genuine Mathematical Explanations of Physical Phenomena?,” Mind 114:
223–38.
Batterman, R. W. (2002) The Devil in the Details, Oxford: Oxford University Press.
Batterman, R. W. (2010) “On the Explanatory Role of Mathematics in Empirical Science,” British Journal
for the Philosophy of Science 61: 1–225.
Broad, C. D. (1925) Mind and Its Place in Nature, London: Kegan Paul.
Brown, H. (2005) Physical Relativity: Space-Time Structure from a Dynamical Perspective, Oxford: Oxford
University Press.
Butterfield, J. (2011) “Less Is Different: Emergence and Reduction Reconciled,” Foundations of Physics 41:
1065–135.
Castellani, E. (2002) “Reductionism, Emergence, and Effective Field Theories,” Studies in History and
Philosophy of Modern Physics 33: 251–67.
Dizadji-Bahmani, F., R. Frigg and S. Hartmann (2010) “Who’s Afraid of Nagelian Reduction?,”
Erkenntnis 73: 393–412.
Dorato, M., and L. Felline (2011) “Scientific Explanation and Scientific Structuralism,” in A. Bokulich
and P. Bokulich (eds): Scientific Structuralism, Dordrecht: Springer, pp. 161–76.
Dowe, P. (2000) Physical Causation, Cambridge: Cambridge University Press.
Falkenburg, B., and M. Morrison (eds.) (2015) Why More Is Different. Philosophical Issues in Condensed Matter
Physics and Complex Systems, Berlin: Springer.
Felline, L. (2015) “Mechanisms Meet Structural Explanation,” Synthese, First online: 30 May 2015: 1–16.
Felline, L. (forthcoming) “Mechanistic Causality and the Bottoming-out Problem,” in Felline et al. (eds.)
New Developments in Logic and Philosophy of Science, London: College Publications.
Frigg, R. (2008) “A Field Guide to Recent Work on the Foundations of Statistical Mechanics,” in
D. Rickles (ed.): The Ashgate Companion to Contemporary Philosophy of Physics, London: Ashgate,
pp. 99–196.
Glennan, S. S. (1996) “Mechanisms and the Nature of Causation,” Erkenntnis 44: 49–71.
Glennan, S. S. (2002) “Rethinking Mechanistic Explanation,” Philosophy of Science 69: S342–S353.
Glennan, S. S. (forthcoming) The New Mechanical Philosophy, Oxford: Oxford University Press.
294
Hartmann, S. (2001) “Effective Field Theories, Reductionism, and Explanation,” Studies in History and
Philosophy of Modern Physics 32: 267–304.
Hüttemann, A. (2005) “Explanation, Emergence and Quantum-Entanglement,” Philosophy of Science 72:
114–27.
Janssen, M. (2009) “Drawing the Line Between Kinematics and Dynamics in Special Relativity,” Studies
In History and Philosophy of Modern Physics 40: 26–52.
Kuhlmann, M. (2014) “Mechanisms in Dynamically Complex Systems,” in P. McKay Illari, F. Russo,
and J. Williamson (eds.): Causality in the Sciences, Oxford: Oxford University Press, 2011, pp. 880–906.
Kuhlmann, M. (2015) “A Mechanistic Reading of Quantum Laser Theory,” in B. Falkenburg and
M. Morrison (2015), pp. 251–71.
Kuhlmann (forthcoming) “Is the Higgs Mechanism a Mechanism?”
Kuhlmann, M. and S. Glennan (2014) “On the Relation Between Quantum Mechanical and Neo-
Mechanistic Ontologies and Explanatory Strategies,” European Journal for Philosophy of Science 4: 337–59.
Lange, M. (2013) “What Makes a Scientific Explanation Distinctively Mathematical?,” British Journal for the
Machamer, P., L. Darden, and C. F. Craver (2000) “Thinking about Mechanisms,” Philosophy of Science
67: 1–25.
Maudlin, T. (1995) “Three Measurement Problems,” Topoi 14: 7–15.
Morrison, M. (2012) “Emergent Physics and Micro-Ontology,” Philosophy of Science 79: 141–66.
Morrison, M. (2014) “Complex Systems and Renormalization Group Explanations,” Philosophy of Science
81: 1144–156.
Norton, J. (2012) “Approximation and Idealization: Why the Difference Matters,” Philosophy of Science
79: 207–32.
Pincock, C. (2015) “Abstract Explanations in Science,” The British Journal for the Philosophy of Science 66:
857–82.
Schlosshauer, M. (2007) Decoherence and the Quantum-to-Classical Transition, Heidelberg: Springer.
Sklar, L. (2015) “Philosophy of Statistical Mechanics,” in E. N. Zalta (ed.): Stanford Encyclopedia of Philosophy
(Fall 2015 Edition), available at https://plato.stanford.edu/entries/statphys-statmech.
Skow, B. (2014) “Are There Non-Causal Explanations (of Particular Events)?,” British Journal for the
Strevens, M. (2016) “Complexity Theory”, in P. Humphreys (ed.): Oxford Handbook of the Philosophy of
Science, doi: 10.1093/oxfordhb/9780199368815.013.35.
Weber, E., and M. Lefevere (2014) “The Role of Unification in Micro-Explanations of Physical Laws,”
Theoria 79: 41–56.
295
22
MECHANISMS IN
EVOLUTIONARY BIOLOGY
Lane DesAutels
1. Introduction
In 1961, renowned Harvard biologist Ernst Mayr asked himself why the warbler he had been
observing on the grounds of his New Hampshire summer home began its southward migration
on the night of the 25th of August. The answer, he came to realize, was not a simple one. There
were ecological causes: being an insect-eater, the warbler would starve in the New Hampshire
winter. There were physiological causes: the warbler had an intrinsic capacity to sense the dwin-
dling number of daylight hours. And there were genetic causes: the warbler’s special sensitivity
to environmental stimuli indicating the approach of colder climate was programmed into its
very DNA. But on Mayr’s view, these myriad reasons for the warbler’s migration were really
just of two kinds: the proximate causes dealing with the physiology of the warbler as it related to
the photoperiodicity and air conditions in its environment, and the ultimate causes dealing with
the bird’s evolutionary history, the way its genetic constitution had been molded by natural
selection over many thousands of generations (Mayr 1961).
Well beyond understanding the migration habits of the common warbler, Mayr’s distinction
between proximate and ultimate causes was to serve as grounds for demarcating the distinct
explanatory magisteria of two different kinds of biology. On the one hand, functional biology
seeks to understand proximate causes: “the functional biologist is vitally concerned with the
operation and interaction of structural elements, from molecules up to organs and whole indi-
viduals. His ever-repeated question is ‘How?’” (Mayr 1961, 1502). Evolutionary biology, on the
other hand, lives squarely in the domain of ultimate causes:
The animal or plant or micro-organism . . . [the evolutionary biologist] is working

with is but a link in an evolutionary chain of changing forms, none of which has any
permanent validity. There is hardly any structure or function in an organism that can
be fully understood unless it is studied against this historical background.
(Mayr 1961, 1502)
Rather than asking “How?”, the evolutionary biologist’s perpetual question is “Why?”
Mayr’s ideas regarding functional vs. evolutionary biology have been the subject of renewed
interest in the philosophy of biology literature. Some have argued that they have significant
flaws (Laland et al. 2011, 2013; Calcott 2013); while others have attempted to support the spirit,
296
Mechanisms in evolutionary biology
if not the letter, of Mayr’s conclusions (Ariew 2003; Scholl and Pigliucci 2014). I will not weigh
in on the specifics of this ongoing debate, but I bring up Mayr’s framework because it repre-
sents an influential and pervasive way of thinking about the explanatory scope of evolutionary
biology—one that, if sound, has important implications for the role of mechanistic philosophy
of science in evolutionary biology. Consider, for example, the following recent characterization
of evolutionary biology from Futuyma’s foundational text, Evolution:
Evolutionary biology is a more historical science than most other biological dis-
ciplines, for it seeks to determine what the history of life has been and what has
caused those historical events. It complements studies of the PROXIMATE
CAUSES (immediate, mechanical causes) of biological phenomena—the subject of
cell biology, neurobiology, and many other biological disciplines—with an analysis
of the ULTIMATE CAUSES (the historical causes, especially the action of natural
selection) of those phenomena.
(Futuyma 2013, 13)
As seen here, evolutionary biology is still today conceived of as an essentially historical science
devoted to understanding ultimate causes. And it is still conceived of as distinct from fields like
cell biology, molecular biology, developmental biology, or genetics whose job is to delineate
immediate, mechanical, proximate causes.
What does any of this have to do with the philosophy of mechanisms and its role in evolu-
tionary biology? Here is the rub. Implicit in both Mayr’s and Futuyma’s ideas are the following
two claims:
i) Evolutionary biology takes ultimate rather than proximate causes as its subject matter.
and
ii) Mechanistic causes are proximate rather than ultimate.
But if we accept (i) and (ii), it would seem to follow that
iii) Mechanistic causes are not the subject matter of evolutionary biology.
Ergo, we might extrapolate,
iv) The mechanistic approach to philosophy of science has nothing (at least directly) to offer the study
of evolution.
The purpose of this chapter is to explore whether these conclusions are warranted—that is, to
explore whether, and to what extent, the mechanistic philosophical framework has any purchase
in evolutionary biology.
There are a number of ways we might go about this. One way would be to question whether
there are any immediate mechanistic causes driving evolution. Is natural selection aptly under-
stood as a mechanism? Is drift? Mutation? Let us refer to this as the metaphysical question (MQ):
(MQ) Are any of the primary causal drivers of evolution aptly understood as mechanisms?
If any of the primary causal drivers of evolution are aptly understood as mechanisms, and the
study of the primary causes of evolution is within the subject matter of evolutionary biology,
then it would seem that—contra (iii)—there is a straightforward sense in which evolutionary
biology should take mechanistic causes as (at least partly constitutive of ) its subject matter.
297
Lane DesAutels
Beyond the metaphysical question, however, there is also an important epistemological question (EQ):
(EQ) When applied to evolutionary biology, does the mechanistic philosophical

framework help add to our understanding of evolution?
The answer to this question, I suggest, may be quite independent from the first.1 Regardless of
whether there are any mechanisms of evolution in the robust metaphysical sense, it might still
be the case that the mechanistic philosophical framework can offer some important explanatory,
pragmatic, or otherwise strategic resources that, when applied to the field of evolutionary biol-
ogy, help illuminate the phenomenon of evolution. If so, then—contra (iv)—there may indeed
be a valuable role for the philosophy of mechanisms in the study of evolution.
In what follows, I will consider responses to both the metaphysical and epistemological ques-
tions. In section 2, I briefly lay out some of the central metaphysical features of mechanisms as
they have been characterized in recent literature. In section 3, I examine whether these features
of mechanisms are aptly understood as being instantiated by three of the primary causal drivers
of evolution: natural selection, drift, and mutation. In section 4, I expound a few of the strategic
roles played by mechanistic thinking and hint at the ways they might apply to the field of evo-
lutionary biology regardless of whether there turn out to be any mechanisms driving evolution.
I conclude, in section 5, that—on at least some philosophical characterizations of mechanisms,
and with regard to at least some of the central processes of evolutionary biology—there is room
for affirmative answers to both the metaphysical and epistemological questions. And thus, there
may indeed be an important place for mechanisms in evolutionary biology.
2. Central metaphysical features of mechanisms

In this section, I highlight five central metaphysical features of mechanisms: components, opera-
tions, organization, function, and regularity.2 Many of these aspects of mechanisms have been
discussed by other authors in this volume (see especially Chapters 7, 8, 9, 10, 12), so I will not
spend much time here. The point is to say just enough about what makes mechanisms the kinds
of things that they are to be able to arbitrate the question of whether natural selection, drift, or
mutation might qualify.
F1: components
Mechanisms have components or constitutive parts. Machamer, Darden, and Craver (MDC) call
these “entities” and claim that, along with their “activities,” they form the two aspects of the
“dualist ontology” of mechanisms (MDC 2000, 3). Stuart Glennan refers to these as “parts”
(Glennan 1996, 52, 2002, S344) and Bechtel and Abrahamsen as “component parts” (Bechtel
and Abrahamsen 2005, 47) (see Chapter 9 for further discussion about mechanism components).
F2: operations
Mechanisms’ component parts perform operations; they do things. In MDC’s terms, these are
the “activities” performed by the entities—the second aspect of mechanisms’ dualist ontology
(MDC 2000, 3). Glennan characterizes this aspect of mechanisms as the “interactions” between
parts (Glennan 1996, 52, 2002, S344) and Bechtel and Abrahamsen as “component operations”
(Bechtel and Abrahamsen 2005, 47).
298
F3: organization
Mechanisms must be—in some sense—organized. For MDC, this means that mechanisms have
“start-up” and “termination” conditions, and that mechanisms’ entities must be (1) located,
(2) structured, (3) oriented; and a mechanism’s activities must have (4) temporal order, (5) rate,
(6) duration (MDC 2000, 3). Bechtel and Abrahamsen speak of a mechanism as a “structure”
performing a task by virtue of its component parts, operations, and their “organization” (Bechtel
and Abrahamsen 2005, 47). And Glennan speaks of mechanisms as “complex systems” underlying
a given phenomenon (Glennan 1996, 52, 2002, S344).
Unlike the merely nominal disagreements mentioned in (F1) and (F2), however, there appear
to be substantive differences in the organizational requirements placed on mechanisms between
contemporary mechanists. Bechtel and Abrahamsen (2009, 2010, 2013) argue that the MDC
organizational requirements are untenable if we wish to characterize feedback mechanisms (e.g.,
the mechanisms responsible for circadian rhythms) or for mechanisms for the maintenance of
equilibrium (e.g., regulatory mechanisms)—as these phenomena have no clear start or termina-
tion conditions. To accommodate single causal chains as mechanisms (e.g., the 1980 death of
French literary critic Roland Barthes when he was struck by a laundry truck while crossing a
Paris street), Glennan goes even further toward loosening the organizational requirement by
arguing for a notion of “ephemeral mechanism” according to which the configuration of a
mechanism’s parts may be “short-lived” and “non-stable” (Glennan 2010, 260).
F4: function
Mechanisms have to be set up to do something; they must carry out a function. Glennan (1996)
points out that “one cannot even identify a mechanism without saying what it is that the mecha-
nism does” (Glennan 1996, 52). For MDC, mechanisms have a function “to the extent that the
activity of a mechanism as a whole contributes to something in the context that is taken to be
antecedently important” (MDC 2000, 6). And Bechtel and Abrahamsen state explicitly that “a
mechanism is a structure performing a function” (Bechtel and Abrahamsen 2005, 47).
Once again, however, there appear to be substantive disagreements about how to understand
mechanistic function. As in the more general debate about functions in biology, there seem to
be roughly two camps: proponents of normative notions of function (e.g., Millikan’s (1989) and
Neander’s (1991) notion of “proper function”—notably defended in the context of mechanisms
by Garson (2011)) vs. non-normative causal notions of function (e.g., Cummins’ (1975) idea of
“causal-role function”—notably applied to mechanisms by Craver (2001)). (See Chapter 8 for a
detailed discussion of mechanistic function.)
F5: regularity
Mechanisms carry out their function in a regular fashion. For MDC, the entities and activities
constitutive of a given mechanism “work always or for the most part in the same way under
the same conditions” (MDC 2000, 3). Bechtel and Abrahamsen (2005) say that mechanisms are
structures responsible for a given “phenomenon”—where “phenomenon” is here understood
in the regularist sense advocated by Bogen and Woodward (1988).
As with (F3) and (F4), there is again substantive disagreement between mechanists about
the degree to which regularity should be conceived of as a metaphysical prerequisite for
mechanisms. Machamer (2004) as well as Bogen (2005) have argued that regularity should
be struck from the MDC characterization of mechanism on the grounds that many of the
299
Lane DesAutels
phenomena targeted for mechanistic explanation (e.g., especially synaptic transmission) fail
to achieve termination conditions much more often than they succeed. DesAutels (2011,
2015) argues that neither a fully regularist nor a fully irregularist characterization of mecha-
nism is tenable, and a notion of stochastic mechanism must be developed. Holly Andersen
(2012) defends a version of the MDC regularity requirement but develops a helpful tax-
onomy according to which mechanistic regularity comes in a variety of strengths and may
be located in a variety of loci within a given mechanism. (See Chapters 11, 12, and 13 for
further exposition of some of the issues surrounding mechanistic regularity.)
3. Do the primary causal drivers of evolution instantiate (F1)–(F5)?

Having now laid out the above central metaphysical features of mechanisms (F1)–(F5) as well
as having drawn attention to a few places where contemporary mechanists disagree on the
nature and strength of these various requirements, we can get on with exploring whether
they are instantiated by some of the primary causal drivers of evolution: natural selection,
drift, and mutation.
Natural selection
Provided there is heritable variation of fitness-relative traits among a population as well as
competition for limited resources, nature tends to preserve those characteristics that afford their
possessors the greatest chance to survive and reproduce, and it tends to reject those that do
not. The result is that, over time, species become increasingly well matched to their respective
environments—or else they go extinct. In its basic form, this is natural selection. Understood in
this way, does natural selection instantiate (F1)–(F5)?
Regarding (F1) and (F2), I suggest that natural selection fares pretty well. Consider Darwin’s
finches as a paradigmatic case. In this example, natural selection seems clearly made up of
components: finches with differing beak-lengths and seeds that are harder or easier to forage
depending on beak-shape. And these componential entities undertake operations: foraging for
seeds, differential reproduction, and predator avoidance among fitter finches. Barring general
skepticism about the nature of part–whole relations—regarding which mechanistic mereology
would hardly seem to be a special case—it seems entirely reasonable to conceive of natural selec-
tion as being composed of active entities.
Regarding (F3)–(F5), the story is more complicated. Whether natural selection is taken to
meet the organization criterion set forth in (F3) depends on how strict we take the requirement
to be. On the original MDC characterization of mechanism according to which mechanisms
are to have definitive set-up/start and termination conditions, natural selection seems to fall
short.3 When would we say of Darwin’s finches that the mechanism of natural selection begins
or ends? It seems that any attempt to delineate such temporal boundaries would be arbitrary or
ad hoc. Furthermore, unlike human-made mechanisms like clocks or toasters, there is not an
obvious physical boundary around the mechanism of natural selection—which makes it hard
to determine what its internal structure might be. That said, on more lenient versions of the
organizational requirement, there do not seem to be any obvious organizational impediments
for natural selection. If, like Bechtel and Abrahamson (2009, 2010, 2013), we allow for feedback
mechanisms or mechanisms for the maintenance of equilibrium, neither of which have obvi-
ous start or termination conditions, then it seems we should not disallow natural selection from
being a mechanism on the grounds that it lacks such conditions. Or if, like Illari and Williamson
(2010), we cash out mechanistic organization in terms of functional individuation, then natural
300
selection meets the organizational requirement just as well as paradigmatic mechanisms from
molecular biology (e.g., protein synthesis). Or if, like Glennan (2010), we allow for any single
causal chain to be a short-lived, non-stable, ephemeral mechanism, then it is even easier to see
that natural selection passes muster.
What about (F4)? Does natural selection have a function? Is it set up to do something?
Once again, these questions do not have straightforward answers. Whether there is sense to
be made of teleology in evolution is an issue with a long and complicated history.4 That said,
even if we agree (as most do) that there is not teleological directionality to evolution—no
end-goal it sets out to achieve—there may still be a sense in which natural selection has a
function. Namely, it is that which brings about adaptation. And if we take on something like the
causal-role understanding of biological function (à la Cummins (1975) and Craver (2001)),
and grant that natural selection is that which plays the causal role of bringing about adaptation,
then natural selection might well satisfy (F4).5
So how about (F5)? Does natural selection evince the regularity that is required for being
a mechanism? Once again, whether we answer this in the affirmative depends on how one
understands mechanistic regularity. If one requires fully deterministic output conditions from
a process for it to count as a mechanism, then natural selection surely does not pass. For as
Gould (1990) famously argued, if the tape of evolution were played back again and again, it
would never turn out the same way twice. In this way, natural selection should be understood
as probabilistic rather than deterministic. And Skipper and Millstein (2005) have argued that this
probabilistic character of natural selection precludes it from meeting the regularity requirements
set forth by MDC (2000). On the other hand, failure to behave fully deterministically does not
preclude all notions of mechanistic regularity from being met. Barros (2008) argues that natural
selection is a biased stochastic mechanism: one that is regular in that it succeeds in producing
predicted outcomes over 50 percent of the time. DesAutels (2016) argues that, when we dis-
tinguish between process vs. product, internal vs. external, and abstract vs. concrete regularity,
then natural selection escapes the Skipper and Millstein regularity critique just fine.
Drift
Drift happens when a population changes over time—where these changes are not the result
of natural selection.6 Consider, for example, a population of snails, half of which are pink and
half are which are yellow.7 Imagine further that the yellow snails are twice as fit as the pink
ones because of their greater resistance to the sun. Scientists observing these snails expect,
therefore, that the population of yellow to pink snails should increase from one half of the
population to two thirds in the subsequent generation. However, suppose that after observing
these snails for one generation, they find that the population of yellow snails actually decreases
from one half to two fifths of the snail population. Because this change in the population is
not due to natural selection (selection would have increased the relative proportion of yellow
snails), they attribute this unexpected result to drift. Understood in this way, how does drift
fare with regard to (F1)–(F5)?
Regarding (F1) and (F2), I suggest it remains reasonable to consider drift as comprising both
components and operations. In our above example, there are pink and yellow snails engaged in
usual survival and reproductive activities. There are also whatever entities and activities were
responsible for the unexpected decrease in the population of yellow snails—perhaps a dangerous
fungus to which only the yellow snails were vulnerable. So once again, modulo general mereo-
logical skepticism, there seems to be little problem with conceiving of this instance of drift as
being composed of active entities.
301
Lane DesAutels
Unfortunately, the story regarding (F3)–(F5) is even more difficult than it was with natural
selection. For instance, it seems particularly difficult to conceive of a sense in which drift is
organized—or that it has any particular structure. At least with natural selection, there are some
commonalities between the sorts of relationships that occur among its participants (e.g., competi-
tion for limited resources or differential reproduction favoring fitter members of the population),
and the participants must meet certain preconditions for natural selection to occur (e.g., varia-
tion in the population and heritability of these variations). On the other hand, because drift is an
umbrella term catching all non-selective instances of population change over time, it seems much
more difficult to delineate any common characteristics shared among all of its instances. And this,
I suggest, makes it exceedingly difficult to understand drift as meeting any kind of organizational
constraint. That said, there may still be some sense to be made of drift as a causal process.8 (The
decrease in population size of the yellow snails, though not caused by natural selection, was surely
caused.) And if we follow Glennan (2010) in allowing for a conception of ephemeral mechanism
that covers all instances of single causal chains, then it may well be that we could conceive of
instances of drift as ephemeral mechanisms: i.e., mechanisms that are “short-lived” and “non-
stable” like the one responsible for the death of Roland Barthes.
But does drift have a function? Is there something that drift is set up to do? Once again, this
is tough to answer affirmatively. Unlike with natural selection which (at least on a causal-role
notion of function) can be understood as that which brings about adaptation, drift does not seem
to do any specific thing; it does not serve any particular unified causal role. Unless we allow
for the function of drift to be the fixation of a given trait absent selection (a function that risks
vacuity given the very definition of drift), it seems drift cannot meet (F4).
Which leaves (F5): regularity. Is there any sense to be made of drift operating regularly? Here,
I suggest that the prospects are once again somewhat dim. Because of the aforementioned nature
of drift as a catch-all for every instance of non-selective population change, it is near impossible
to conceive of it as operating in a regular fashion. A population might drift because of disease,
human encroachment, natural disasters, or just plain bad luck. And if mechanistic regularity
requires, as MDC suggest, that mechanisms operate “always or for the most part in the same way
under the same conditions,” then we might think that drift appears to fall woefully short. That
said, there may be more space for understanding drift as operating regularly than all of this might
portend. For one thing, drift (at least according to many biologists) tends to occur more often among
smaller populations. So in this sense, there are at least some conditions we can specify under which
drift is more likely to occur. And if such conditions can be specified, then there may at least be
a sense in which drift operates more regularly in some circumstances than others.
Mutation
Deoxyribonucleic acid (DNA) is composed of two polymers made up of nucleotides and a
backbone of sugars and phosphate groups—all organized into the shape of a double helix. DNA
replication begins when the parent molecule gets unzipped as the hydrogen bonds between the
base pairs are broken. Once separated, the sequence of bases on each of the unzipped strands
becomes a template for the insertion of a complementary set of bases. Deoxynucleoside triphos-
phates assemble the new strands in the order that complements the order of bases on the strand
serving as the template. When the process is complete, two DNA molecules have been formed
identical to each other and to the parent molecule. However, there are several ways that the
process of DNA replication can (and does) go wrong—resulting in mutation. One such way is
when a base is changed by the repositioning of a hydrogen atom, altering the hydrogen bonding
302
pattern of that base, and resulting in incorrect base pairing during replication. Another way is
when there is a loss of a purine base (A or G) to form an apurinic site (AP site). There can also
be denaturation of the new strand from the template during replication, followed by renatura-
tion in a different spot. This can lead to insertions or deletions.
Despite the fact that mutations are essentially mistakes in DNA replication, it is a good thing
they occur. Without them, there would be no evolution. Mutations are the source of the
variation on which natural selection acts. As such, mutation is an indisputable causal driver of
evolution. So how does mutation fare with regard to (F1)–(F5)?
Here, again, there are no serious obstacles to meeting (F1) or (F2): we have component
entities: DNA, a purine base, individual molecules, etc., and we have operations: unzipping,
separating, inserting, etc.
Regarding (F3)–(F5), the situation with mutation seems a good measure improved from
that of both natural selection and drift. Mutation appears to have plausible start and set-up
conditions: the presence of DNA in a living organism and the initiation of hydrogen bond
separation. And it has reasonable termination conditions: the existence of a mutated strand
of DNA. Additionally, unlike either natural selection or drift, the process responsible for
mutation has a readily discernible molecular structure common among all of its instances.
Furthermore, much like natural selection, mutation can easily be seen to carry out a unified
causal-role function: it is that which brings about genetic variation at the population level
as a result of copying errors at the individual level. Finally, more so than either selection or
drift, mutation seems to occur in a regular fashion. It is well known, for example, that errors
during nucleotide substitution are biased and regularly occur more often at certain locations
on the chromosome.
So how should we sum up the metaphysical question? To what extent is it the case that these
primary causal drivers of evolution (natural selection, drift, and mutation) are aptly understood
as mechanisms? What we have just seen is that the answer to the metaphysical question is com-
plicated. Among the primary causal drivers of evolution, whether they are aptly understood as
meeting (F1)–(F5) depends crucially on how liberal we understand these requirements to be.
With strict enough constraints on organization, function, and regularity, it may well be that
none of these processes end up counting as mechanisms. On highly liberal versions of these con-
straints, they all may well qualify. So the painfully residual question, then, is: how liberal should we
understand these metaphysical requirements to be? The answer to this question, I am afraid I cannot
offer in this short chapter. But I will say this. Liberalized metaphysical requirements on mecha-
nisms have the advantage of allowing more physical processes to be understood as mechanisms,
and this may have several pragmatic benefits (to be briefly expounded in the following section).
However, highly liberalized metaphysical requirements have drawbacks as well. Namely, the
more physical processes we allow to count as mechanisms, the less interesting and distinctive
mechanistic explanations will end up being. As with any tradeoff between theoretic virtues, the
right balance can only be arbitrated by extra-theoretic values.
4. Does the mechanistic framework add to our

understanding of evolution?
Let us turn our attention to the epistemological question. That is, let us consider whether,
regardless of whether any of the primary causal drivers of evolution are aptly understood as
mechanisms, the mechanistic philosophical framework might help add to our understanding
of evolution.
303
Lane DesAutels
But how could this possibly be? How could mechanistic philosophy of science help illuminate
the study of evolution even if there are no mechanisms driving it? In pondering this question,
it will be helpful to consider the notion of “strategic mechanism” recently developed by Levy
(2013). According to Levy, one way of approaching the philosophy of mechanisms is to conceive
of it as constituting a strategy: “a way of doing science, a framework for representing and reasoning
about complex systems” (Levy 2013, 105). Regardless of the metaphysical status of mechanisms
out there in the world, the idea is that the mechanist “sees mechanistic methods as having particu-
lar cognitive and epistemic features” (ibid., 105) and maintains that “certain phenomena are best
handled mechanistically” (ibid., 100).
In keeping with Levy, I suggest that there are at least three significant strategic roles that
mechanisms might play in our understanding of science regardless of their metaphysical status:
reductionist explanation, manipulation, and discovery. And when applied to the field of evolution-
ary biology, it seems to me that each of these strategic roles has the potential for adding to our
understanding of evolution.
The strategic role of reductionist explanation is emphasized most convincingly by Bechtel
and Richardson (1993) in their discussion of “complex localization.” They write,
Complex localization requires a decomposition of systemic tasks into subtasks, local-

izing each of these in a distinct component. Showing how systemic functions are, or
at least could be, a consequence of these subtasks is an important element in a fully
mechanistic explanation.
(Bechtel and Richardson 1993, 125)
Thinking about nature mechanistically leads naturally to explanation by reduction and decom-
position. When thinking mechanistically, we can more readily come to an understanding of
why some phenomenon occurs by reducing it to its component parts and operations and by
showing how the phenomenon in question is brought about by these parts and operations.
Evolution is no exception. To understand why some trait (or individual with a given trait) was
selected for, it seems a perfectly good strategy to decompose the larger ecological system in
which the individual is a member into its component parts and operations—and in doing so,
gain insight as to the advantages the trait confers. A similar story can be told of the strategic role
of mechanisms for manipulation. When we model an instance of selection, drift, or mutation as
a mechanism (often this is done using mechanism schemata made up of boxes and arrows rep-
resenting causal relationships), it may become more salient where to intervene on the system
in question to test various hypotheses about the nature of these causal interactions.9 Finally, as
emphasized in their recent book, Craver and Darden (2013) convincingly show that mecha-
nistic thinking has tremendous benefits viz. scientific discovery: by progressively filling in the
black boxes in our incomplete mechanism schemata, scientists are supplied with an invaluable
heuristic for moving from how-possibly explanations to how-actually explanations of the natural
world. This benefit seems especially apropos in the study of evolution where so many of our
discoveries (e.g., fossils of intermediate forms like Archaeopteryx or Tiktaalik) originated as the
product of speculative hypotheses.
Of course, it would require much more argument to establish that these strategic advantages
of mechanistic thinking actually do pay off in the field of evolutionary biology. However, since I
cannot undertake this work here in any detail, I am content for the time being to establish that
an affirmative answer to the epistemological question has a reasonably high degree of plausibil-
ity. And this plausibility remains even if no affirmative answer can be given to the metaphysical
question. Levy’s notion of strategic mechanism helps to do just that.
304
5. Conclusion
In this chapter, I have explored answers to both the metaphysical and epistemological
questions regarding the status of mechanisms in evolutionary biology. Regarding the meta-
physical question, I briefly laid out some of the central metaphysical features of mechanisms
and examined whether they are aptly understood as being instantiated by three of the pri-
mary causal drivers of evolution: natural selection, drift, and mutation. Provided we adopt
relatively lenient versions of these constraints, I suggest the answer is yes. Regarding the
epistemological question, I briefly expounded Levy’s notion of strategic mechanism and
hinted at ways the strategic benefits of mechanistic thinking might apply to the field of
evolutionary biology regardless of whether there turn out to be any mechanisms driving
evolution. I conclude, therefore, that Mayr and Futuyma may have been a bit hasty in their
implicit claim that evolutionary biology be confined to the realm of ultimate causes at the
exclusion of mechanistic ones. Contrary to this view, I have shown that—on at least some
philosophical characterizations of mechanisms, and with regard to at least some of the central
processes of evolutionary biology—there is room for affirmative answers to both the meta-
physical and epistemological questions. And thus, there may indeed be an important place
for mechanisms in evolutionary biology.
Notes
1 The distinction between metaphysical and epistemological questions regarding the appropriateness of
mechanistic thinking as applied to a given domain echoes one made by Glennan and Illari in Chapter 7
of this volume.
2 See Illari and Williamson (2012) and Chapter 1 for arguments that these requirements can be expressed
by a common conception of mechanism that applies widely across the sciences.
3 See Skipper and Millstein (2005) for an argument to this effect.
4 See Ariew (2007) for a nice survey of some of these issues.
5 Not everyone agrees that natural selection is a causal process (cf. especially Matthen and Ariew 2002).
Following Millstein (2006) and more recently Ramsey (forthcoming), I presume that natural selection
can be given a causalist interpretation.
6 More detailed philosophical analyses of drift vary greatly. Some argue that drift cannot be distinguished
from selection and so cannot be conceived of as a distinct evolutionary process (for instance, Matthen
and Ariew 2002; Matthen 2009). And there are some (for example, Millstein 2002, 2008) who argue
that drift is distinct from selection by virtue of the kind of process it is (indiscriminate as opposed to
discriminate sampling of a population).
7 Example comes from Roberta Millstein (1996, S15).
8 As with natural selection, there are some who deny that drift is causal. See especially Lange (2013) as an
example.
9 See especially Craver (2007) for a detailed exposition of this advantage.
References
Andersen, H. (2012) “The Case for Regularity in Mechanistic Causal Explanation,” Synthese 189 (3):
415–432.
Ariew, André (2003) “Ernst Mayr’s ‘Ultimate/Proximate’ Distinction Reconsidered and Reconstructed,”
Biology and Philosophy 18 (4): 553–565.
Ariew, André (2007) “Teleology,” in D. L. Hull and M. Ruse (eds.), The Cambridge Companion to the
Philosophy of Biology, Cambridge: Cambridge University Press, 160–181.
Barros, Benjamin (2008) “Natural Selection as a Mechanism,” Philosophy of Science 74: 306–322.
Bechtel, W. and Abrahamsen, A. (2005) “Explanation: A Mechanist Alternative,” in C. F. Craver and
L. Darden (eds.), Special Issue: “Mechanisms in Biology,” Studies in History and Philosophy of Biological
and Biomedical Sciences 36: 421–441.
305
Lane DesAutels
Bechtel, W. and Abrahamsen, A. (2009) “Decomposing, Recomposing, and Situating Circadian

Mechanisms: Three Tasks in Developing Mechanistic Explanations,” in H. Leitgeb and A. Hieke
(eds.), “Reduction: Between the Mind and the Brain,” Ontos 12–177.
Bechtel, W. and Abrahamsen, A. (2010) “Dynamic Mechanistic Explanation: Computational Modeling of
Circadian Rhythms as an Exemplar for Cognitive Science,” Studies in History and Philosophy of Science
Part A 41 (3): 321–333.
Bechtel, W. and Abrahamsen, A. (2013) “Thinking Dynamically about Biological Mechanisms: Networks
of Coupled Oscillators,” Foundations of Science 18 (4): 707–723.
Bechtel, W. and Richardson, R. C. (1993) Discovering Complexity: Decomposition and Localization as Strategies
Bogen, J. (2005) “Regularities and Causality; Generalizations and Causal Explanations,” in Special Issue:
“Mechanisms in Biology,” in C. Craver and L. Darden (eds.), Studies in History and Philosophy of
Biological and Biomedical Sciences 36 (2): 397–420.
Bogen, J. and Woodward, J. (1988) “Saving the Phenomena,” The Philosophical Review 97 (3): 305–352.
Calcott, B. (2013) “Why How and Why Aren’t Enough: More Problems with Mayr’s Proximate-Ultimate
Distinction,” Biology and Philosophy 28 (5): 767–780.
Craver, C. F. (2001) “Role Functions, Mechanisms, and Hierarchy,” Philosophy of Science 68: 53–74.
Craver, C. F. (2007) Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience, New York:
Craver, C. F. and Darden, L. (2013) In Search of Mechanisms: Discoveries across the Life Sciences, Chicago, IL:
Cummins, R. (1975) “Functional Analysis,” Journal of Philosophy 72: 741–764.
DesAutels, L. (2011) “Against Regular and Irregular Characterizations of Mechanisms,” Philosophy of
Science 78 (5): 914–925.
DesAutels, L. (2015) “Toward a Propensity Interpretation of Stochastic Mechanism for the Life Sciences”
Synthese 192: 2921–2953.
DesAutels, L. (2016) “Natural Selection and Mechanistic Regularity,” Studies in History and Philosophy of
Biological and Biomedical Sciences 57: 13–23.
Futuyma, D. J. (2013) Evolution (3rd Edition). Sunderland, MA: Sinauer Associates.
Garson, J. (2011) “Selected Effects and Causal Role Functions in the Brain: The Case for an Etiological
Approach to Neuroscience,” Biology and Philosophy 26 (4): 547–565.
Glennan, S. (1996) “Mechanisms and the Nature of Causation,” Erkenntnis 44 (1): 49–71.
Glennan, S. (2002) “Rethinking Mechanistic Explanation,” Proceedings of the Philosophy of Science Association
3: S342–S353.
Glennan, S. (2010) “Ephemeral Mechanisms and Historical Explanation,” Erkenntnis 72 (2): 251–266.
Gould, S. (1990) Wonderful Life: The Burgess Shale and the Nature of History, New York: W. W. Norton
and Company.
Illari, P. and Williamson, J. (2010) “Functional Organization: Comparing the Mechanisms of Protein
Synthesis and Natural Selection,” Studies in History and Philosophy of Biological and Biomedical Sciences 41:
279–291.
Illari, P. and Williamson, J. (2012) “What Is a Mechanism: Thinking about Mechanisms across the
Sciences,” European Journal for Philosophy of Science 2: 119–135.
Laland, K. N., Sterelny, K., Odling-Smee, J., Hoppitt, W., and Uller T. (2011) “Cause and Effect in
Biology Revisited: Is Mayr’s Proximate–Ultimate Dichotomy Still Useful?” Science 334: 1512–1516.
Laland, K. N., Odling-Smee, J., Hoppitt, W., and Uller, T. (2013) “More on How and Why: Cause and
Effect in Biology Revisited,” Biology and Philosophy 28 (5): 719–745.
Lange, M. (2013) “Really Statistical Explanations and Genetic Drift,” Philosophy of Science 80 (2): 169–188.
Levy, A. (2013) “Three Kinds of New Mechanism,” Biology and Philosophy 28: 99–114.
Machamer, P., Darden, L., and Craver, C. (2000) “Thinking about Mechanisms,” Philosophy of Science 67
(1): 1–25.
International Studies in the Philosophy of Science 18 (1): 27–39.
Matthen, M. (2009) “Drift and Statistically Abstractive Explanation,” Philosophy of Science 76: 464–487.
Matthen, M. and Ariew, A. (2002) “Two Ways of Thinking about Fitness and Natural Selection,” Journal
of Philosophy 49: 55–83.
Mayr, E. (1961) “Cause and Effect in Biology,” Science 134: 1501–1506.
Millikan, R. G. (1989) “In Defense of Proper Functions,” Philosophy of Science 56: 288–302.
306
Millstein, R. L. (1996) “Random Drift and the Omniscient Viewpoint,” Philosophy of Science 63 (3):
S10–S18.
Millstein, R. L. (2002) “Are Random Drift and Natural Selection Conceptually Distinct?” Biology and
Philosophy 17: 33–53.
Millstein, R. L. (2006) “Natural Selection as a Population-Level Causal Process,” British Journal for the
Millstein, R. L. (2008) “Distinguishing Drift and Selection Empirically: ‘The Great Snail Debate’ of the
1950s,” Journal of the History of Biology 41 (2): 339–367.
Neander, K. (1991) “Functions as Selected Effects: The Conceptual Analyst’s Defense,” Philosophy of
Science 58: 168–184.
Ramsey, G. (forthcoming) “The Causal Structure of Evolutionary Theory,” Australasian Journal of
Philosophy: DOI: 10.1080/00048402.2015.1111398.
Scholl, R. and Pigliucci, M. (2014) “The Proximate–Ultimate Distinction and Evolutionary Developmental
Biology: Causal Irrelevance Versus Explanatory Abstraction,” Biology and Philosophy 30: 653–670.
Skipper, R. A., Jr. and Millstein, R. A. (2005) “Thinking about Evolutionary Mechanisms: Natural
Selection,” in C. F. Craver and L. Darden (eds.), Special Issue: “Mechanisms in Biology,” Studies in
History and Philosophy of Biological and Biomedical Sciences 36: 327–347.
307
23
MECHANISMS IN
MOLECULAR BIOLOGY
Tudor M. Baetu
1. Introduction
While mechanistic explanations are by no means a novelty in biology (see Part I on historical
perspectives on mechanisms), their appearance dating back to Harvey’s discovery of the blood’s
circulation and Descartes’ mechanistic manifesto touted in The Treatise of Man, it is only in the
second half of the twentieth century, with the rise of molecular biology, that mechanistic think-
ing overtakes the whole of biology, becoming the predominant type of explanation. Explaining
biological phenomena as the effects of molecular mechanisms turned out to have a marked
influence on philosophy of science on two accounts. First, it motivated a renewed interest in
mechanistic explanation, providing philosophers with a wealth of examples that didn’t quite
fit the deductive-nomological approach promoted by the logical positivists (Wimsatt 1976).
Second, molecular biology motivated a shift in the way we think about mechanisms, fostering
a “new” mechanistic philosophy. Unlike the “old” mechanistic philosophy, which was closely
linked to the theory of classical mechanics and the clockwork view of the world, the “new”
mechanistic philosophy had to come to terms with the notion that most biological mechanisms
do not look and behave like eighteenth-century automata, but are much more complex and
“noisier” systems composed of hundreds, thousands, and even millions of non-fixed, non-rigid
parts whose behavior is nevertheless sufficiently constrained both by the properties of the parts
and by the spatio-temporal organizational features of the system as to reliably produce and
sustain biological phenomena and, ultimately, life itself.
2. What is molecular biology?

Molecular biology is the field of scientific investigation concerned with the molecular basis of
biological activity. Its most significant achievements are the elucidation of the mechanisms of
replication, transcription, and translation in the 1950s and 60s, providing an explanation of how
cells replicate their genetic material and how genes are expressed as proteins that contribute to
the phenotype of the organism. Molecular biologists quickly extended their inquiries to other
biological phenomena, most notably gene expression regulation, the cell cycle, and cellular sign-
aling. By the 1970s, molecular biology gained a firm footing in many other fields of biology, with
developmental biology, immunology, neurology, and microbiology “going molecular.” In this
308
Mechanisms in molecular biology
respect, molecular biology can be viewed as providing the guidelines of a general explanatory
approach which I shall explore in more detail in the subsequent sections of this chapter.
Historically, molecular biology was born from the convergence of work in genetics, bio-
chemistry, and physical chemistry in an attempt to figure out how genes determine phenotypes
(Carlson 1967; Darden 1991, 2006a; Fox Keller 2000; Kay 1993; Morange 1998; Schaffner
1993).1 Such questions fall outside the immediate explanatory scope of classical genetics, which
is mainly concerned with the transmission of inherited traits (Morgan 1935; Moss 2003; Waters
1994). They also fall outside the immediate scope of biochemistry, which focuses predomi-
nantly on metabolic activities (e.g., the chemical reactions taking place during glycolysis or
protein synthesis), as well as that of physical chemistry, which is concerned with the inter- and
intramolecular forces shaping the tridimensional structure and the physicochemical properties of
macromolecules (Baetu 2012a; Darden 2006b; Morange 2002; Olby 1994).
It is customary to distinguish molecular biology from related fields on the basis of its specific
techniques of investigation revolving around the “cloning” of genetic material (creating copies
of DNA fragments), along with techniques required to detect (e.g., electrophoresis), sequence
(e.g., chain-termination sequencing), amplify (e.g., polymerase chain reaction, or PCR), and
manipulate (e.g., site-directed mutagenesis) genetic material, many of which rely on the under-
standing of the mechanisms of replication, transcription, and translation which are at the core of
molecular biology (Astbury 1961; Waters 2008). These techniques involve experimental inter-
ventions at the resolution of individual nucleotides—and, via the alteration of codon sequences,
of individual amino-acids in a peptide sequence—thus tracking the flow of information within
the cell, revealing the functional role of sequence motifs such as promoters, codons, and zinc-
fingers relative to the operation of molecular mechanisms, and providing an understanding of
how changes in the sequence of various molecular components result in changes in the opera-
tion of mechanisms and their ability to produce or sustain biological phenomena.
It should be noted, however, that molecular biology was and continues to be part of a highly
integrated cluster of fields. Molecular biologists routinely rely on data, theoretical assumptions,
and formal and experimental techniques from genetics, biochemistry, physical chemistry, cell
biology, microbiology, statistics, epidemiology, systems biology, systematics, and many others.
At the same time, scientists in other fields, from evolutionary biology to psychiatry, rely on
explanatory strategies and experimental techniques inspired from molecular biology. As a result,
our current understanding of the molecular basis of biological activity integrates findings from a
wide variety of fields within and outside biology.
3. The nature of molecular mechanisms

In his celebrated essay What Is Life?, Erwin Schrödinger tackles the difficult question of the
origin of order in biological systems. He argues that biological systems escape entropy because
there is information in the system telling it how to assemble itself in an organized fashion.
On this account, information is a source of order in what would have otherwise been a ther-
modynamically disordered system. Genes would be the repositories of information, which is
propagated throughout the cell by a series of deterministic mechanisms that preserve order. This
view, predating the major discoveries of molecular biology, endures until today, to the point
that some philosophers define molecular biology as the study of the mechanisms of information
propagation within cells (Darden 2006a, 2006b; Morange 1998).
Schrödinger’s order-generating information was eventually identified with genetic infor-
mation, while its propagation throughout the cell by means of deterministic mechanisms
came to be known as genetic determinism. The best illustration is the “genetic program”
309
Tudor M. Baetu
view popularized by François Jacob (1976) and Jacques Monod (1972); for a philosophical
discussion, see Baetu (2012b). According to this view, information is unidirectionally propa-
gated via the mechanisms of transcription and translation to peptide sequences, which in turn
determine the tridimensional shape of proteins via the folding of α-helices, ß-sheets, and other
secondary structures, itself determining the specificity of binding to other molecules accord-
ing to a lock-and-key or induced fit model, thus directing the flow of chemical reactions
underpinning metabolic and signaling pathways as well as the self-assembly of the various
supramolecular structures that constitute the living cell.
While immensely popular in the 1970s and still echoing today, this view is in fact a
special case of a more general concept in molecular biology, namely that of specific bind-
ing (Kupiec 2009; Morange 1998). Specificity of binding is assumed to determine not only
the preferential pairing of nucleotides in complementary strands of nucleic acids during
replication and transcription or between codons and aminoacyl-tRNAs during translation,
but also virtually any single activity related to the operation of molecular mechanisms,
including enzyme-substrate interactions, the recognition of extracellular ligands by cell-
surface receptors, the binding of transcriptional factors to particular DNA sequences, the
self-assembly of microtubules and ribosomes, etc. Specificity is determined by chemical
affinity, which is typically measured as the average life span of a supramolecular complex in
a given chemical environment. When molecules form stable, long-lasting complexes likely
to have a marked impact on biological activity, their binding is said to be specific; when the
complexes are short-lived, their binding is said to be non-specific. Thus, order-generating
information is not restricted to the genome, but is in fact manifest in every single specific
molecular interaction taking place in the cell.
Less appreciated by the general public is the fact that binding specificity is an “analog,” or
stochastic concept (Rao et al. 2002). Any given molecule always interacts with many other
molecules, for some with stronger and for others with weaker affinity, such that specific bind-
ing invariably occurs against the “noisy” background of myriad non-specific interactions.
Furthermore, it is not uncommon that the same macromolecule can bind with relatively high
specificity not only one, but many other molecules (e.g., most transcriptional factors can bind
with variable, but relatively high specificity several related DNA sequences).2 By contrast,
genetic determinism assumes an idealized limit case, a “noise-free” or “digital” specificity prop-
agating itself throughout the cell without significant distortion as the inevitable result of the
expression of genomic sequences.
“Digital” specificity idealizations are still ubiquitous in the qualitative descriptions of
molecular biology. While there are epistemic benefits associated with such idealizations, most
notably increased intelligibility, the truth is that, until very recently, molecular biologists had
no other choice but to idealize. Traditionally, molecular techniques rely on the amplification
of the properties, states, and activities of molecular components up to an unequivocal threshold
of detection. This is achieved by studying millions of cells in bulk over (chemically speaking)
long periods of time, which means ignoring both time fluctuations, as well as differences from
one cell to another. Yet the underlying biochemistry dictates that most molecular activities
are chemical reactions controlled by the concentrations, states, and locations of the various
molecules involved. If these parameters are not identical in all cells, there will be fluctuations
from one cell to another with respect to quantitative and dynamic aspects of the operation and
output of mechanisms. Furthermore, chemical reactions are intrinsically stochastic since they
rely on random microscopic events that govern how fast reactions occur and in what order.
For years, it was assumed that although stochastic fluctuations are bound to occur, they are
biologically irrelevant “background noise” cells keep to a minimum. Thus, it was and often still
310
is customary to assume that what is true of a large population of cells can be safely extrapolated
to individual cells within that population. In turn, this justifies the belief that, on average, the
same mechanism, following the sequence of chemical events, is synchronously operating in all
cells of the same type subjected to the same conditions.
In contrast with this view, recent single-cell experiments revealed that stochastic fluctua-
tions are non-negligible, meaning that cells must rely on noise-suppressing mechanisms, such
as negative feedback and DNA “proof-reading” (Elowitz et al. 2002). It also turned out that
stochasticity itself can be biologically relevant and is in fact used as a means for generating
diversity and histological-level patterns. For example, probabilistic biases in the distribution
of adhesion proteins suffice to generate the right amount of twisting in the developing gut,
while stochastic gene expression is responsible for generating blue- to ultraviolet-sensitive
cells ratios in the Drosophila eye; for discussion, see Baetu (2015b), Heams (2011), and Merlin
(2011). The growing realization that many mechanisms in biology are stochastic motivated
a more careful investigation of the extent to which regularity is a distinctive characteristic of
mechanisms (Andersen 2012; Darden 2008; DesAutels 2011), with some authors emphasizing
the highly irregular (Bogen 2005) and even the singular nature of some mechanisms (Glennan
2010, 2011). The issues of stochasticity, regularity, and singular causation are discussed in
more detail in Chapters 10–13. More general implications for theories of causation can be
found in Hall (2004) and Psillos (2004).
Specific binding is clearly responsible for generating some supramolecular structures, from
the recruitment of polymerase complexes to the self-assembly of proteasomes and microtubules;
furthermore, techniques like in vitro translation or PCR would be impossible without spe-
cific binding. Nevertheless, specificity is sensitive to the effective concentrations of molecular
components, with the “background noise” of non-specific interactions increasing for low copy
numbers of molecules. This observation led to a questioning about whether specific binding is
the exclusive origin of order in biological systems.
The traditional model, inherited from early twentieth-century biochemistry, is the cell
as a “bag of enzymes.” According to this admittedly idealized model, the spatio-temporal
organization of molecular mechanisms, from the dynamics of activities to the assembly of
supramolecular structures, is driven by specificity of binding in the context of a free diffusion
solution chemistry. In short, were it not for differential binding specificities, the molecules
inside a cell would display the same amount of order as sugar dissolved in a glass of water.
This model turned out to be unsatisfactory. Molecular mechanisms operate within an intra-
cellular environment filled with other molecules. Macromolecular crowding functions as
an excluded volume effect, favoring the aggregation of macromolecules while drastically
decreasing the rate of any diffusion-dependent molecular activity (Ellis 2001). This strongly
suggests that the intracellular environment cannot consist of a disordered collection of mol-
ecules, as this would result in the proliferation of noise-generating non-specific interactions
at the expense of order-generating specific ones. Instead, it must be structured in such a
way as to bring in close proximity proteins and their ligands, thus favoring the specific
chemical interactions required for the operation of molecular mechanisms (Hochachka 1999;
Mathews 1993). However, if this is the case—and recent evidence supports this conclusion,
e.g. the effects of nuclear architecture on transcription (Cremer and Cremer 2001)—then
the tridimensional structure of the cell and tissue organization are also a source of order,
functioning as a scaffold constraining the behavior of molecular mechanisms by favoring
some activities while suppressing others. This “return to holism” in molecular biology did
not go unnoticed in philosophy of biology, motivating a renewed attack on genetic deter-
minism on the grounds that gene expression is a stochastic process (Kupiec 2009), that
311
Tudor M. Baetu
information is distributed throughout the cell ( Jablonka 2001), and that genes alone—or for
that matter the properties of individual molecules, such as their binding affinities—cannot
fully account for phenotype and development (Dupré 2010; Griffiths and Stotz 2006, 2013;
Oyama et al. 2001).
One final twist in the story of the molecular basis of biological activity is the “systemic turn”
in biology. Its impact on the way we think about mechanisms is discussed in Chapter 27. The
present discussion will focus on one of the many findings that prompted this turn, namely the
realization that molecular mechanisms operate in a less modular fashion than initially thought.
Molecular biologists often work under the assumption that mechanisms amount to discrete
functional modules organized hierarchically, serially or in parallel. While useful as a heuristic of
discovery, the modularity assumption came to be questioned by a growing body of evidence
revealing that distinct mechanisms responsible for distinct phenomena share mechanistic compo-
nents. Notoriously, intracellular signaling pathways and gene regulatory mechanisms invariably
intersect at some point, forming widespread molecular networks (Davidson and Levine 2005).
It has therefore been proposed that many molecular mechanisms are in fact inextricably inter-
locked into vast networks of partially overlapping mechanisms, where the sharing of mechanistic
components is thought to play a role in the fine-tuning of quantitative-dynamic aspects of the
phenomena produced by these mechanisms (Barabási and Oltvai 2004). In some cases, math-
ematical models were able to account for minute discrepancies between the predicted and
observed outcomes by taking into consideration interference from overlapping mechanisms,
thus providing some evidence for the biological significance of non-modular modes of organi-
zation; for philosophical discussion and references to the original scientific literature, see Baetu
(2015b); Bechtel and Abrahamsen (2010, 2011).
In recent philosophical debates, modularity is often understood along the lines of “independ-
ent disruptability” stating that the overall effect of a causal system can be decomposed into a set
of independent causal contributions attributable to each of the constituents of the system (Steel
2007; Woodward 2002). The systemic turn in biology prompted some authors to reject causal
modularity (Bogen 2004; Mitchell 2009), and even interpret it as a vindication of a more holistic
approach defended by the old organicist school of thought (Nicholson 2013); for a historical
discussion of mechanism vs. organicism, see Chapter 5. It is interesting, however, to note that if
molecular mechanisms are constrained by higher-level cellular and histological structures, then
significant forms of modularity must prevail in virtue of the physical partitioning of biochemi-
cal reactions (Callebaut and Rasskin-Gutman 2005; Hartwell et al. 1999). If so, the relevant
issue is not whether cells and organisms are organized in a modular fashion, but the extent to
which molecular mechanisms, as described in biology textbooks, amount to functional or causal
modules (Baetu 2016; Woodward 2010).
4. Mechanistic explanations in molecular biology

One of the explicit aims of molecular biology is to provide a reductive explanation of biological
phenomena in terms of molecular mechanisms. It is therefore crucial to investigate more closely
the notion of mechanistic explanation: how do mechanisms explain?
According to the ontic view, mechanistic explanations are objective features of the world.
To explain a phenomenon is to fit it into the causal structure of the world (Craver 2007).
The advantages of this view is that it eliminates the subjective notion of understanding while
emphasizing the tight link between explanation and the ability to control effects by interven-
ing on their causes. Notwithstanding, most philosophers prefer an epistemic view, according to
which mechanistic explanations are step-by-step descriptions of how mechanisms produce the
312
phenomena for which they are responsible, although some authors point out that both ontic and
epistemic considerations contribute to a successful mechanistic explanation (Illari 2013). (For
treatment of ontic vs. epistemic conceptions, see Chapter 16.)
In one version of the epistemic view, descriptions of mechanisms, in words or by means of
diagrams, convey an intuitive understanding consisting in simulating the working of mecha-
nisms in our imagination (Bechtel and Abrahamsen 2005) or by analogy with more common
types of activities (Machamer 2004). While imagination is indispensable for our ability to learn
about and understand molecular mechanisms, equating mechanistic explanation with intuitive
understanding can be problematic. For one thing, the simulations we perform in our minds are
heavily idealized. Detailed quantitative aspects are absent, while known facts are distorted to
outline a deterministic sequence of events tracking the fate of single molecules (a more accurate
biochemical description would be that of a series of back-and-forth equilibria involving popula-
tions of molecules bumping into each other), each assumed to be rigidly structured (ignoring
the fact that molecules “breathe,” vibrating and cycling through various configurations), as they
modify one another to bring about a change from an initial to a final state of affairs (thus mask-
ing an underlying variety of chemical pathways that contribute to the same final state, as well as
ignoring alternate pathways where mechanisms fail to contribute to the output).
Such idealizations are due to the fact that intuitive understanding relies on analogies with
macroscopic mechanisms more familiar to us. In contrast, examples of mechanistic explana-
tions from the sciences show that the working entities of a mechanism—that is, the parts of the
mechanism engaged in the operation of the mechanism (Machamer et al. 2000)—span multiple
levels of composition, thus supporting the notion that biological mechanisms combine a wide
variety of activities associated with different levels of composition, from push-pull mechani-
cal interactions to chemical reactions and thermodynamic processes (Craver 2007; Craver and
Darden 2013; Darden 2006b; also Chapter 14 in the present volume). There is even evidence
that some molecular mechanisms harness the quirks of quantum mechanics (Ball 2011); for a
philosophical discussion, see Barwich (2015). Arguably, an intuitive understanding of chemi-
cal equilibria, thermodynamic processes, and quantum mechanics by means of analogies with
macroscopic mechanisms is problematic, as many of these concepts defy our imagination and are
best captured by a rather complex mathematical formalism.
This brings us to a third view meant to reflect a relatively recent quantitative turn in molecu-
lar biology, namely an understanding mediated by evidence from the testing of mathematical
models aimed at demonstrating that mechanisms can in principle produce the phenomena for
which they are responsible in close approximation of detailed quantitative measurements (Baetu
2015b; Bechtel 2012; Bechtel and Abrahamsen 2011; Braillard 2010; Brigandt 2013; Gross
2015; also Chapters 20 and 27). This view also qualifies as epistemic, albeit the understanding
is more of a theoretical than an intuitive-imaginative nature, as it involves showing how phe-
nomena are consequences of explicit rules and assumptions about the operation of mechanisms.
(A more comprehensive treatment of representations of mechanisms in general can be found in
Chapter 18. Models of mechanisms are discussed in more detail in Chapter 17.)
It is interesting to note that mechanistic explanations do not automatically preclude a role
for generalizations and regularities in biology. The ability of some mechanisms to regularly pro-
duce a phenomenon accounts for some of the generalizations observable in the biological world
(Glennan 2002; Illari and Williamson 2012; Machamer et al. 2000). For example, mechanistic
constraints can explain why some evolutionary outcomes are more probable than others, and
allow for predictions in specific lineages (Baetu 2012c), while the reliance on exemplar organism
models supports the notion that general patterns shared by a large number of species exist and play
an important role in guiding research (Ankeny 2001; Bolker 1995; Schaffner 2001; Weber 2005).
313
Tudor M. Baetu
Conversely, the behavior of mechanistic parts is often characterized by invariant

change-relating generalizations (Glennan 1996, 2002; Woodward 2002, 2010). Likewise,
mathematical modeling presupposes a set of rules specifying how mechanisms change from
one state to another, where some of these rules are laws borrowed from chemistry and
physics (Schaffner 1993; Weber 2005). The current tendency to complement experimental
research with mathematical modeling is also responsible for reviving the notion that mathe-
matical or, more often, computational derivation contributes to the mechanistic explanation
by demonstrating that certain aspects of a phenomenon are the consequences of the rules
governing the operation of the mechanism. These considerations suggest a rather com-
plicated relationship between mechanisms and regularities in molecular biology, whereby
mechanisms both rely on regularities governing the behavior of their components and are
themselves responsible for generating novel regularities, typically under the form of recur-
rent or reproducible phenomena.
Another issue that requires clarification concerns the completeness of mechanistic expla-
nations. This issue is closely linked to the more general problem of how mechanisms are
discovered (Chapter 19). In this chapter, I will focus on questions about the levels of com-
position at which mechanistic explanations bottom-out and top-off—that is, the optimal
resolution of detail at which mechanisms should be described and the extent to which mecha-
nisms act as independent modules that can produce (and therefore explain) the phenomena
for which they are responsible when separated from the systems in which they are embedded.3
The decomposition of biological systems reveals a hierarchical organization, with lower-level
components organized as mechanisms underlying higher-level components (Bechtel 2006;
Bechtel and Richardson 2010; Craver 2007). It might therefore be tempting to conclude
that a more complete explanation can be achieved by investigating the components of a
mechanism to reveal the finer-grained sub-mechanisms responsible for their properties and
activities, as well as understanding how the mechanism fits in the context of progressively
more comprehensive systems of mechanisms. Nonetheless, even though such investiga-
tions are bound to generate new knowledge—for instance, by explaining why mechanistic
components have the properties they have or by providing the structural details necessary
to intervene on these components and their properties—they cannot guarantee explana-
tory completeness. Assuming that the primary objective of a mechanistic explanation is to
demonstrate causation—that is, show how an organized system of parts produces a specific
phenomenon—then whether a mechanistic explanation is complete is a matter of providing
evidence that parts having the properties and organization specified in the explanation can and
do produce the phenomenon of interest.4 This kind of evidence is generated by attempts to
build the mechanism in vitro using reconstitution experiments, in vivo using the techniques
of genetic engineering and synthetic biology, and in silico by testing mathematical models,
as well as more indirectly, by assessing the ability to correctly foresee and correct side-effects
of treatments and other technological applications based on mechanistic explanations (Baetu
2015a; Craver and Darden 2013; Morange 2009; Weber 2005).
Another way of looking at the issue of explanatory completeness is from a pragmatic stand-
point. There are many ways in which a mechanistic explanation is useful to the scientist, and
the purpose will determine when the explanation is deemed to be complete (Craver and Darden
2013). The main advantage of a pragmatic stance is that it can accommodate seemingly conflict-
ing evaluations. For instance, if one seeks to gain control over the phenomenon of interest, then
the explanation should focus on the manipulable components of the actual mechanism respon-
sible for the phenomenon, such as gene and protein sequences (Craver 2006; Waters 2007); if
prediction is the main goal, then the focus will fall on showing how changes in the components
314
of a mechanism result in changes in the phenomenon produced by the mechanism (Cartwright

2002; Woodward 2010); if one seeks intelligibility, be it for didactic purposes or to reveal
general patterns, then abstracting or even idealizing may be necessary (Levy 2014).
5. Conclusion
Molecular biology was undoubtedly the most influential field of biological research in the second
half of the twentieth century, with hardly any branch of biology left untouched by the long-
reaching arm of the molecular revolution. At the same time, it is equally important to realize that
molecular biology itself changed over time as a result of interactions with other fields, from its ori-
gins as interdisciplinary research in genetics, biochemistry, and physical chemistry, to its subsequent
integration of ideas from cell and developmental biology, to its current interactions with synthetic,
systems biology, and nanoscience. From a philosophical point of view, advances in molecular biol-
ogy prompted an inquiry about scientific explanation and the nature of biological mechanisms.
Under the impetus of the elucidation of the mechanisms of genome replication and expression,
molecular mechanisms were initially idealized as thoroughly deterministic devices propagating
genetic information. This deterministic conception with a strong reductionistic flavor gave way to
a more realistic interpretation endorsed by most molecular biologists today, namely that of molec-
ular mechanisms as deterministic systems with noise, where the spatio-temporal organization of
mechanisms is generated not only by specificity of binding (an important part of which is genetic
information), but also by cellular and other supramolecular organization constraints.
Notes
1 Among these authors, some take the official date of birth of molecular biology to coincide with the
coining of the term by Warren Weaver in 1938 (Kay 1993; Olby 1994), while others postpone it until
the elucidation of the structure of DNA in 1953 and the emergence of the modern concept of genetic
information (Darden 2006a, 2006b).
2 There are many reasons why molecular interactions are “noisy.” Organic molecules can assume multiple
stable configurations, each characterized by its distinct affinities; for instance, spontaneous mutations can
occur via the tautomerization of nucleotides. Such variability is expected to be even more pronounced
in macromolecules such as proteins, which can fold in a variety of configurations. Binding itself deforms
molecules, thus altering their affinities and generating rather hard-to-predict phenomena such as alloste-
rism and cooperative binding. Finally, new discoveries suggest that some proteins are unstructured and
assume different folding configurations depending on the partners with which they interact. For refer-
ences and philosophical discussion, see Kupiec (2009).
3 In many respects, these issues are a continuation of an earlier debate about reductionism in biology.
While some philosophers have argued that the best explanations in biology are those bottoming out at a
molecular level (Rosenberg 2006; Waters 1994), others replied that descriptions at higher levels of com-
position are needed to account for organizational features of biological systems (Laubichler and Wagner
2001) or to enhance intelligibility (Kitcher 1984).
4 Some physicists proposed that biological explanations are likely to bottom-out at a molecular level because at
this size scale the mechanical, electrostatic, chemical, and thermal energies of objects have similar magnitudes.
The convergence of energy values means that mechanical, electrical, and chemical energy can be converted
into one another, which can explain some of the most fundamental properties of living things, such as their
ability to convert food (chemical energy) into motion (mechanical energy). Furthermore, thermal energy
(random molecular collisions) may quite literally “kick start” molecular mechanisms, thus explaining their
ability to work autonomously and spontaneously. In contrast, mechanical forces at macroscopic scales or
binding forces stabilizing sub-molecular structures such as atoms are largely insensitive to thermal fluctua-
tions at room/body temperature, and therefore phenomena such as spontaneous activation, self-assembly, or
change of shape are impossible (Hoffmann 2012; Philips and Quake 2006).This suggests that there is some-
thing objectively special about the molecular level in the sense that only molecules, as opposed to atoms or
macroscopic objects, have the kind of properties required by mechanistic explanations in biology.
315
Tudor M. Baetu
References
Andersen, H. 2012. “The Case for Regularity in Mechanistic Causal Explanation.” Synthese 189
(3):415–32.
Ankeny, R. 2001. “Model Organisms as Models: Understanding the ‘Lingua Franca’ of the Human
Genome Project.” Philosophy of Science 68:S251–S61.
Astbury, W. 1961. “Molecular Biology or Ultrastructural Biology?” Nature 190 (4781):1124.
Baetu, T. M. 2012a. “Emergence, Therefore Antireductionism? A Critique of Emergent Antireductionism.”
Biology and Philosophy 27 (3):433–48.
——. 2012b. “Genomic Programs as Mechanism Schemas: A Non-Reductionist Interpretation.” British
Journal for the Philosophy of Science 63 (3):649–71.
——. 2012c. “Mechanistic Constraints on Evolutionary Outcomes.” Philosophy of Science 79 (2):276–94.
——. 2015a. “The Completeness of Mechanistic Explanations.” Philosophy of Science 82 (5):775–86.
——. 2015b. “From Mechanisms to Mathematical Models and Back to Mechanisms: Quantitative
Mechanistic Explanations.” In Explanation in Biology: An Enquiry into the Diversity of Explanatory Patterns
in the Life Sciences, ed. P.-A. Braillard and C. Malaterre, 345–63. Dordrecht: Springer.
——. 2016. “From Interventions to Mechanistic Explanations.” Synthese 193: 3311–27
Ball, P. 2011. “Physics of Life: The Dawn of Quantum Biology.” Nature 474 (7351):272–74.
Barabási, A.-L., and Z. N. Oltvai. 2004. “Network Biology: Understanding the Cell’s Functional
Organization.” Nature Reviews Genetics 5 (2):101–13.
Barwich, A.-S. 2015. “Bending Molecules or Bending the Rules: The Application of Theoretical Models
in Fragrance Chemistry.” Perspectives on Science 23 (4):1–23.
Bechtel, W. 2006. Discovering Cell Mechanisms: The Creation of Modern Cell Biology. Cambridge: Cambridge
University Press.
Bechtel, W. 2012. “Understanding Endogenously Active Mechanisms: A Scientific and Philosophical
Challenge.” European Journal for Philosophy of Science 2:233–48.
Bechtel, W., and A. Abrahamsen. 2005. “Explanation: A Mechanist Alternative.” Studies in History and
Philosophy of Biological and Biomedical Sciences 36:421–41.
——. 2010. “Dynamic Mechanistic Explanation: Computational Modeling of Circadian Rhythms as an
Exemplar for Cognitive Science.” Studies in History and Philosophy of Science Part A 41:321–33.
——. 2011. “Complex Biological Mechanisms: Cyclic, Oscillatory, and Autonomous.” In Philosophy of
Complex Systems, ed. C. A. Hooker, 257–85. New York: Elsevier.
Bechtel, W., and R. Richardson. 2010. Discovering Complexity: Decomposition and Localization as Strategies in
Scientific Research. Cambridge, MA: MIT Press.
Bogen, J. 2004. “Analyzing Causality: The Opposite of Counterfactual Is Factual.” International Studies in
the Philosophy of Science 18:3–26.
——. 2005. “Regularities and Causality: Generalizations and Causal Explanations.” Studies in History and
Bolker, J. 1995. “Model Systems in Developmental Biology.” BioEssays 17 (5):451–5.
Braillard, P.-A. 2010. “Systems Biology and the Mechanistic Framework.” History and Philosophy of Life
Sciences 32:43–62.
Brigandt, I. 2013. “Systems Biology and the Integration of Mechanistic Explanation and Mathematical
Explanation.” Studies in History and Philosophy of Biological and Biomedical Sciences 44 (4):477–92.
Callebaut, W., and D. Rasskin-Gutman. 2005. Modularity: Understanding the Development and Evolution of
Natural Complex Systems. Cambridge, MA: MIT Press.
Carlson, E. O. 1967. The Gene: A Critical History. Philadelphia: W.B. Saunders.
Cartwright, N. 2002. “Against Modularity, the Causal Markov Condition and any Link between the Two:
Comments on Hausman and Woodward.” British Journal for the Philosophy of Science 53 (3):411–53.
Craver, C. 2006. “When Mechanistic Models Explain.” Synthese 153:355–76.
——. 2007. Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience. Oxford: Clarendon Press.
Craver, C., and L. Darden. 2013. In Search of Biological Mechanisms: Discoveries across the Life Sciences. Chicago,
IL: University of Chicago Press.
Cremer, T., and C. Cremer. 2001. “Chromosome Territories, Nuclear Architecture and Gene Regulation
in Mammalian Cells.” Nature Reviews Genetics 2:292–301.
Darden, L. 1991. Theory Change in Science: Strategies from Mendelian Genetics. New York: Oxford University
Press.
——. 2006a. “Flow of Information in Molecular Biological Mechanisms.” Biological Theory 1 (3):280–7.
316
——. 2006b. Reasoning in Biological Discoveries: Essays on Mechanisms, Interfield Relations, and Anomaly
Resolution. Cambridge: Cambridge University Press.
——. 2008. “Thinking Again about Biological Mechanisms.” Philosophy of Science 75:958–69.
Davidson, E., and M. Levine. 2005. “Gene Regulatory Networks.” Proceedings of the National Academy of
Science 102 (14):4935.
DesAutels, L. 2011. “Against Regular and Irregular Characterizations of Mechanisms.” Philosophy of Science
78 (5):914–25.
Dupré, J. 2010. “It Is Not Possible to Reduce Explanations in Biology to Explanations in Chemistry and/
or Physics.” In Contemporary Debates in Philosophy of Biology, ed. F. Ayala and R. Arp, ch. 2. Chichester,
UK: Wiley-Blackwell.
Ellis, J. 2001. “Macromolecular Crowding: Obvious but Underappreciated.” Trends in Biochemical Sciences
26 (10):597–604.
Elowitz, M., A. J. Levine, E. D. Siggia, and P. S. Swain. 2002. “Stochastic Gene Expression in a Single
Cell.” Science 297 (5584):1183–6.
Fox Keller, E. 2000. The Century of the Gene. Cambridge, MA: Harvard University Press.
Glennan, S. 1996. “Mechanisms and the Nature of Causation.” Erkenntnis 44:49–71.
——. 2002. “Rethinking Mechanistic Explanation.” Philosophy of Science 69:S342–S53.
——. 2010. “Ephemeral Mechanisms and Historical Explanation.” Erkenntnis 72:251–66.
——. 2011. “Singular and General Causal Relations: A Mechanist Perspective.” In Causality in the Sciences,
ed. P. McKay, J. Williamson and F. Russo, 789–817. Oxford: Oxford University Press.
Griffiths, P., and K. Stotz. 2006. “Genes in the Postgenomic Era.” Theoretical Medicine and Bioethics 27
(6):499–521.
——. 2013. Genetics and Philosophy: An Introduction. Cambridge: Cambridge University Press.
Gross, F. 2015. “The Relevance of Irrelevance: Explanation in Systems Biology.” In Explanation in
Biology: An Enquiry into the Diversity of Explanatory Patterns in the Life Sciences, ed. P.-A. Braillard and
C. Malaterre. Dordrecht: Springer.
Hall, N. 2004. “Two Concepts of Causation.” In Causation and Counterfactuals, ed. J. Collins, N. Hall and
L. Paul, 225–76. Cambridge, MA: MIT Press.
Hartwell, L. H., J. J. Hopfield, S. Leibler, and A. W. Murray. 1999. “From Molecular to Modular Cell
Biology.” Nature 6761supp (C47–52).
Heams, T. 2011. “Expression stochastique des gènes et différenciation cellulaire.” In Le hasard au coeur
de la cellule: Probabilités, déterminisme, génétique, ed. J.-J. Kupiec, O. Gandrillon, M. Morange and
M. Silberstein, 31–59. Paris: Éditions Matériologiques.
Hochachka, P. W. 1999. “The Metabolic Implications of Intracellular Circulation.” Proceedings of the
National Academy of Sciences 96 (22):12233–9.
Hoffmann, P. M. 2012. Life’s Ratchet: How Molecular Machines Extract Order from Chaos. New York: Basic
Books.
Illari, P. 2013. “Mechanistic Explanation: Integrating the Ontic and Epistemic.” Erkenntnis 78 (2):237–55.
Illari, P., and J. Williamson. 2012. “What is a Mechanism? Thinking about Mechanisms across the Sciences.”
European Journal for Philosophy of Science 2 (1):119–35.
Jablonka, E. 2001. “The Systems of Inheritance.” In Cycles of Contingency: Developmental Systems and
Evolution, ed. S. Oyama, P. Griffiths and R. Gray. Cambridge, MA: MIT Press.
Jacob, F. 1976. The Logic of Life. New York: Pantheon.
Kay, L. E. 1993. The Molecular Vision of Life: Caltech, the Rockefeller Foundation, and the Rise of the New
Biology. New York: Oxford University Press.
Kitcher, P. 1984. “1953 and All That: A Tale of Two Sciences.” Philosophical Review 93:335–73.
Kupiec, J.-J. 2009. The Origin of Individuals. Singapore: World Scientific Publishing.
Laubichler, M., and G. Wagner. 2001. “How Molecular is Molecular Developmental Biology? A Reply
to Alex Rosenberg’s Reductionism Redux: Computing the Embryo.” Biology and Philosophy 16:53–68.
Levy, A. 2014. “What was Hodgkin and Huxley’s Achievement?” British Journal for the Philosophy of Science
65 (3):469–92.
Machamer, P. 2004. “Activities and Causation: The Metaphysics and Epistemology of Mechanisms.”
International Studies in the Philosophy of Science 18 (1):27–39.
Machamer, P., L. Darden, and C. Craver. 2000. “Thinking about Mechanisms.” Philosophy of Science
67:1–25.
Mathews, C. K. 1993. “The Cell-Bag of Enzymes or Network of Channels?” Journal of Bacteriology 175
(20):6377–81.
317
Tudor M. Baetu
Merlin, F. 2011. “Pour une interprétation objective des probabilités dans les modèles stochastiques
de l’expression génétique.” In Le hasard au coeur de la cellule: Probabilités, déterminisme, génétique, ed.
J.-J. Kupiec, O. Gandrillon, M. Morange and M. Silberstein, 31–59. Paris: Éditions Matériologiques.
Mitchell, S. 2009. Unsimple Truths: Science, Complexity, and Policy. Chicago: University of Chicago Press.
Monod, J. 1972. Chance and Necessity: An Essay on the Natural Philosophy of Modern Biology. London:
Vintage Books.
Morange, M. 1998. A History of Molecular Biology. Cambridge, MA: Harvard University Press.
——. 2002. “The Gene: Between Holism and Generalism.” In Promises and Limits of Reductionism in the
Biomedical Sciences, ed. M. Van Regenmortel and D. Hull. Chichester: John Wiley & Sons. DOI:
10.1002/0470854189.ch9.
——. 2009. “Synthetic Biology: A Bridge Between Functional and Evolutionary Biology.” Biological
Theory 4 (4):368–77.
Morgan, T. H. 1935. “The Relation of Genetics to Physiology and Medicine.” Les prix Nobel en 1933.
Imprimerie Royale: 1–16.
Moss, L. 2003. What Genes Can’t Do. Cambridge, MA: MIT Press.
Nicholson, D. 2013. “Organisms ≠ Machines.” Studies in History and Philosophy of Biological and Biomedical
Sciences 44 (4):669–78.
Olby, R. 1994. The Path to the Double Helix: The Discovery of DNA. Mineola, NY: Dover.
Oyama, S., P. Griffiths, and R. Gray. 2001. “Introduction: What Is Developmental Systems Theory?”, In
Cycles of Contingency: Developmental Systems and Evolution, ed. S. Oyama, P. Griffiths and R. Gray, 1–11.
Cambridge, MA: MIT Press.
Philips, R., and S. R. Quake. 2006. “The Biological Frontier of Physics.” Physics Today 59:38–43.
Psillos, S. 2004. “A Glimpse of the Secret Connexion: Harmonizing Mechanisms with Counterfactuals.”
Perspectives on Science 12:288–391.
Rao, C. V., D. M. Wolf, and A. P. Arkin. 2002. “Control, Exploitation and Tolerance of Intracellular
Noise.” Nature 420 (6912):231–7.
Rosenberg, A. 2006. Darwinian Reductionism, or, How to Stop Worrying and Love Molecular Biology. Chicago:
Schaffner, K. F. 1993. Discovery and Explanation in Biology and Medicine. Chicago: University of Chicago
Press.
——. 2001. “Extrapolation from Animal Models: Social Life, Sex, and Super Models.” In Theory and
Method in the Neurosciences, ed. P. Machamer, R. Grush and P. McLaughlin, 231–49. Pittsburgh:
University of Pittsburgh Press.
Steel, D. 2007. Across the Boundaries: Extrapolation in Biology and Social Science. Oxford: Oxford University
Press.
Waters, C. K. 1994. “Genes Made Molecular.” Philosophy of Science 61:163–85.
——. 2007. “The Nature and Context of Exploratory Experimentation: An Introduction to Three Case
Studies of Exploratory Research.” History and Philosophy of Life Sciences 29 (3):275–84.
——. 2008. “Beyond Theoretical Reduction and Layer-Cake Antireduction: How DNA Retooled
Genetics and Transformed Biological Practice.” In The Oxford Handbook of Philosophy of Biology, ed.
M. Ruse, 238–62. Oxford: Oxford University Press.
Weber, M. 2005. Philosophy of Experimental Biology. Cambridge: Cambridge University Press.
Wimsatt, W. C. 1976. “Reductive Explanation: A Functional Account.” In Boston Studies in the Philosophy
of Science, ed. A. C. Michalos, 671–710. Dordrecht: Reidel.
Woodward, J. 2002. “What Is a Mechanism? A Counterfactual Account.” Philosophy of Science
69:S366–S77.
——. 2010. “Causation in Biology: Stability, Specificity, and the Choice of Levels of Explanation.” Biology
and Philosophy 25:287–318.
318
24
MECHANISMS AND
BIOMEDICINE1
1. Introduction
Our aim in this chapter is to give a “functionalist account” of the ways that mechanisms are
sought, formulated, and used in medicine. Rather than giving a single analytic account of
mechanism, or a review of ways that existing accounts of mechanism fail to describe one or
other aspects of medical practice, we instead work from a starting position that one of us has pre-
viously called the “mosaic view” of causality (Illari and Russo 2014). According to this mosaic
view, the objective of research by practically and historically engaged philosophers of science is
not to find The-One definition of causality, but instead to understand what role related notions
play in our epistemologies and methodologies. In a similar vein, we here focus on exploring the
use of mechanisms in the methodologies and epistemologies of medicine. In this we are moti-
vated by (amongst others) Andrea Woody’s functionalist account of explanation. Her approach
is to “think about where and when explanations are sought and formulated, and subsequently
to consider what role(s) they might play in practice” (Woody 2015: 81).
Our aim is to provide a related inquiry, concentrating on the ways that mechanisms are
sought, formulated, and used in medicine. We take the main aims of medicine to be to under-
stand and intervene on the health of individuals and of populations, where those interventions
seek to cure, mitigate, or prevent disruptions to human health.
Mechanisms, we submit, are found to the point of ubiquity in medicine. Because of this, we will
use six “episodes” to draw out some of the role(s) that mechanisms play in making sense of medicine.
In section 2, we introduce these episodes in a fairly descriptive way. We then, in section 3, analyze
these episodes to draw lessons about mechanisms in medicine. Finally, in sections 4 and 5, we reach
some tentative conclusions about mechanisms in medicine. Here, a major job is to answer the fol-
lowing question: if the mechanisms project often or usually looks to the sciences for inspiration, why
is there such a mis-match between the high prominence of mechanisms in medical practice, and the
much lower level of attention that medical mechanisms have received from philosophers?
2. Examples of mechanisms in medicine

We begin with what seems a simple fact of the matter: talk of mechanisms is nearly ubiquitous
in medical practice. This is not to claim that mechanisms are either necessary or sufficient to
319
establish the causes and effects of health and disease, but just to notice that mechanisms enter,
very, very frequently, into several inferential practices in the medical sciences. The interesting
question then becomes what these different uses of mechanisms have in common or in which
respects they differ. This is in line with Woody’s functionalist approach.
To build our argument for this diversity, we present some “episodes” of medical mechanisms
at work. We borrow from Chang (2011: 110ff ) the term “episode,” rather than “case study,” to
emphasize these are selected as exemplary cases of the numerous uses of mechanisms in medi-
cine, rather than unique instances chosen ad hoc for the purpose of the present discussion. To
develop this point slightly, we intend to tell a diverse group of stories about the practices of the
medical sciences, present and historical. Yet we hope that the term “episode” will suggest that,
while each tells a different story about medicine, these stories have something in common from
the perspective of researchers interested in mechanisms.
Recall that we take the main aims of medicine to concern understanding and intervening
on the health of individuals and of populations. This means we use the term “medicine” to
include all clinical, scientific, and even political forms of engagement with health and disease
(for a detailed discussion, see Clarke and Russo 2016). We therefore use “medicine” as an
umbrella term with the aim of moving toward a broad and inclusive understanding of medicine
which—we hope—will foster a thorough discussion of the role of mechanisms in this broad-
church medicine. It is in this sense that, in this chapter, we explore the applied epistemology
of mechanisms in medicine. We submit that mechanisms are powerful tools for understanding,
establishing, and intervening on causal relations in medicine. However, we worry that insuf-
ficient attention has been paid to the different ways that an understanding of mechanisms can
contribute to this applied epistemology. Hence our six episodes.
From the discussion of the episodes it will become clear (i) that mechanisms contribute to,
rather than constitute, several inferential practices in medicine and (ii) that the effects of these
contributions are extremely heterogeneous. We will go on to think more analytically about
these various contributions in section 3.
As a first case, consider aspirin, which has been widely used as analgesic, antipyretic, and
anti-inflammatory since its synthesis by Hoffmann in 1897 (Schrör 1997: 349). Its effects are
so well known that the example of aspirin is often used by philosophers of medicine to argue
that mechanisms are not needed to establish causal relations (e.g. Howick 2011: 930). At first
sight, the argument seems compelling: the efficacy of aspirin as a painkiller was known for
many decades before its mechanism of action was understood. Many of us, too, will have taken
and trusted aspirin to (say) relieve our headache despite (we venture) few of us possessing the
knowledge of the relevant mechanisms that explained why it was effective. Yet the analgesic,
antipyretic, and anti-inflammatory effects are not the only effects of aspirin of interest to medical
practice. Aspirin is widely used because of its effect on platelet function. Taking regular low-
dose aspirin improves cardiovascular outcomes (Antithrombotic Trialists’ Collaboration 2002).
Yet, we argue, this role of aspirin is not as perspicuous as its painkilling one. Instead, knowledge
of this effect was intimately linked to understanding the mechanism by which the drug worked.
We summarize here the main steps of how the effect of aspirin on platelet function was
established, based on Schrör’s presentation (1997: 349–50). Quick reported that aspirin
increased bleeding time (which is a measure of the overall rate at which blood clots) in 1967.
This was shortly followed by several reports that low-dose aspirin appeared to inhibit platelet
aggregation—in itself, an important part of blood clotting (Weiss, Aledort, and Kochawa 1968;
O’Brien 1968). This inhibition began within two hours of taking the aspirin, and lasted for sev-
eral days. The investigation of this impressive degree of platelet aggregation in turn led to the
finding that aspirin inhibited prostaglandin synthesis (Vane 1971). Incidentally, this inhibition
320
Mechanisms and biomedicine
of prostaglandin biosynthesis was later to provide the roots of the mechanistic explanations of
the other actions of aspirin. Smith and Willis (1971) then identified that the inhibition of pros-
taglandin biosynthesis was “the mechanism of the antiplatelet action of aspirin” (Schrör 1997:
350), which was traced specifically to inhibition of the cyclooxygenase (or COX) enzyme in
1975, and to a specific amino acid residue in the COX-1 enzyme in 1991. Further mechanistic
research on COX would reveal routes for future interventions that could capitalize on aspirin’s
beneficial effects, while hopefully avoiding its adverse effects.
Our second example is high-density lipoprotein (HDL)-raising drugs. Observational evi-
dence suggests high levels of HDL in the blood are inversely correlated with heart disease.
But is this correlation due to an underlying causal (preventive) relationship between heart
disease and HDL?
How about investigating this question by using HDL-raising as an intervention designed to
prevent heart disease? A recently developed class of drugs called cholesteryl ester transfer pro-
tein (CETP) inhibitors seemed to promise just such an investigation. Clinical trials of the drug
torcetrapib seemed effective at raising HDL levels (Brousseau et al. 2004). However, the drug
did not improve clinical outcomes. In fact, it severely worsened them, leading to the aban-
doning of a large phase III clinical trial (known as ILLUMINATE) in 2006, owing to excess
deaths in patients receiving torcetrapib. This result was unexpected—so much so that one com-
mentator suggested that this result should lead us to conclude that “We know so much about
the cholesterol pathway, but we never seem to know what matters” (Lehrer 2011). The trial
authors (Barter et al.) were unsurprisingly more measured in their analysis of the trial result, and
responded with some educated speculation as to its cause(s):
Clinical trials such as ours are not designed to elucidate mechanisms of either benefit
or harm associated with the use of a drug. However, they may provide clues that have
the potential to inform future research. . . . There are at least two possible explanations
for the observation of increased mortality and morbidity associated with the use of
torcetrapib in our study: an off-target effect of torcetrapib, unrelated to CETP inhibi-
tion, and an adverse effect of CETP inhibition per se, with the possible generation of
dysfunctional or even proatherogenic HDL cholesterol.
(Barter et al. 2007)
Further research has suggested that the “off-target” mechanism suggested by Barter et al. may
be correct. It appears that torcetrapib increases blood pressure (Tall, Yvan-Charvet, and Wang
2007)—itself a well-known cause of cardiac death. If that is the case, then the increased mor-
tality and morbidity found in the ILLUMINATE trial appear to be adverse consequences of
torcetrapib specifically, and therefore unlikely to be replicated in clinical research on alternative
CETP inhibitors (such as evacetrapib and anacetrapib).
The third example is that asbestos is known to be responsible for fatal diseases such as asbes-
tosis and mesothelioma. The biochemical mechanisms that connect exposure to asbestos fibers
with the development of cancer have been studied and examined carefully (see e.g. IARC
Working Group 2012). However, work remains to be done studying the (largely social) mecha-
nisms by which exposure to asbestos occurs. For example, occupational medicine researchers
have been studying the disease from the perspective of the workplace. This led to the study
of populations living close to asbestos factories. Examples abound across different geographical
locations. We might mention, for instance, Barking in the United Kingdom (Greenberg 2003)
and Eternit in Italy, for which a memorable sentence was issued in 2009 after a long and difficult
trial (Mossano 2011; Allen and Kazan-Allen 2012): the owners of the asbestos multinational
321
were deemed guilty of fraudulent environmental disaster and omission. In this context, and
because of their bearing on legal and policy questions, some aspects of the mechanisms underly-
ing asbestos exposure remain disputed, for instance latency (see, e.g., Terracini et al. 2014; La
Vecchia and Boffetta 2014). The point here is that mechanisms investigated via occupational
and environmental epidemiology largely do not concern the way that asbestos causes meso-
thelioma. This, as we shall further discuss later, poses a question for the “overbiologization” of
diseases, namely the reduction of disease causation to biochemical reactions in the body.
In the philosophy of medicine our fourth example, the discovery of Helicobacter pylori as a
cause of gastric ulcer, has been used to illustrate how hypotheses are generated in medicine
(Gillies 2005; Hutton 2012; Thagard 1998a, 1998b). The same episode has also been used to
illustrate the mutual need for evidence of difference-making and of mechanisms in establishing
causal claims in medicine (Russo and Williamson 2007). This body of literature showcased how
the use of mechanisms in given inferential practices also depends on the available theoretical
framework. In this case, available background knowledge had it that bacteria could not live in
acid environments such as the stomach. This tended to preclude the hypothesis that H. pylori
could be a cause of gastric ulcer, and until the mechanism by which this bacteria could survive
in low-pH environments had been investigated, the causal claim could not be properly assessed.
Another stock example in the philosophy of medicine, and our fifth, is the story of Ignaz
Semmelweis, a doctor active in nineteenth-century Vienna. His notoriety is due to his hypoth-
esis that puerperal fever was due to some infection. The mechanisms had not been clarified at
that time, but his suggested preventive intervention was remarkably simple: doctors should wash
their hands after performing autopsies and before assisting women in labor. Part of the debate in
philosophy of medicine concerns the question of whether the scientific community was right or
wrong in rejecting Semmelweis’ precautionary measure on the basis of available evidence and
of the theoretical framework supporting the intervention.
The sixth and final episode we want to present is the comparison between approaches to health
and disease. This becomes highly relevant in times where the public is showing skepticism and
mistrust for so-called “Western medicine” and increasingly seeking advice from “alternative”
approaches. “Alternative” or “complementary” medicine is, however, a basket where too often
anything that is “non-Western” is placed. Instead, care is needed in making comparisons, and
non-Western medical traditions ought not to be conflated.
To illustrate the relevance of our point, we rely on the contribution of Hugh Shapiro in the
volume Medicine across Cultures (Shapiro 2003). Shapiro provides an invaluable contextualization
of what is usually called “Chinese medicine,” including a brief history of its relation to Western
approaches. Shapiro then explains that from a practice of forensic medicine and of dissection,
which was common to both Chinese and Western approaches, very different conceptualizations
of the body (anatomy and physiology) and of pathology derived. Moreover, this led to very dif-
ferent developments: the study of the living body in Chinese medicine and the study of the dead
body in Western medicine. Shapiro illustrates these claims with the case of “nerves.”
At the beginning of the last century the medical community gathered to translate and
standardize terms coming from the West. It became clear that there wasn’t a one-to-one
correspondence. For instance, Chinese lacked both the concept and the word for “nerve.”
Since the Renaissance, attempts to translate Western medical concepts into Chinese ones
had to make recourse to periphrases and to neologisms that would make sense in their con-
ceptual framework. Thus, for instance, nerve was used often as synonymous with sinew, and
its function was to transmit the “vital power.” Chinese doctors have known for a long time
how to intervene on this vital power, using the technique of acupuncture. The lack of a cor-
responding term for “nerve” in Chinese testifies, argues Shapiro, to the different conceptions
322
and understanding of the body and of the phenomena of health and disease. This, it should
be noted, holds for Western culture too. In fact, Shapiro continues, in the West the under-
standing of what nerves are and of how they function has been related to the character trait of
“volition,” an idea that traces back to Greek medicine. Moreover, this is intimately connected
with action and especially volitional actions, which are defining features of identity.
3. Mechanisms contribute to inferences in medicine

After this overview of episodes of mechanisms in medicine, it would be tempting to try and
pin down The-One definition that fits them all. In accordance with the “mosaic view” and the
“functionalist approach” we espoused (see section 1), we will not pursue this objective here.
For our purposes, we can safely rely on the working characterization by Jon Williamson and
Phyllis Illari: “A mechanism for a phenomenon consists of entities and activities organized in
such a way that they are responsible for the phenomenon” (Illari and Williamson 2012: 120).
This kind of view is now routinely referred to as “minimal mechanism” (see also Chapter 1).
This characterization avoids equating mechanisms with Cartesian deterministic machines (see
Chapter 3). Also, this characterization is flexible enough to accommodate mechanisms that
travel across a generic or population-level case and a single, individual case. It allows us to deal
with biochemical, pharmacological, or social mechanisms of health and disease. These and other
aspects of mechanisms are discussed next.
Returning to our first episode, aspirin, examples like this are instructive in that they illus-
trate, in a concrete way, the difficulties that can be experienced in finding effects in medicine.
This is hard—effects are complex, and we think that researchers need to know about mecha-
nisms to reliably find and measure many effects. While there are exceptions—you need no
detailed knowledge to know that an anesthetic is effective in sending someone to sleep—it
can be extremely hard to discover and understand all the clinically relevant effects of an
anesthetic (or, in the aspirin case, analgesic) drug. This is also the case when it comes to find-
ing side effects (such as gastric and duodenal ulcers, Reye’s syndrome, and exacerbations of
asthma in the case of aspirin).
The second example, HDL-raising drugs, shows that medical effects of interest may result
from complex systems of mechanisms that are difficult to separate (like the hypertensive effect
of torcetrapid). The complexity comes from the interaction of drug(s) with a pathology, or
rather with several pathologies. In fact, cases of co-morbidity (i.e., a patient that has more than
one pathological condition at one time) are the norm rather than the exception. Knowledge
of mechanisms is necessary to explain, predict, and treat, especially in these cases. Differently
put, the situation of one-drug-one-pathology rarely occurs. Knowledge of mechanisms doesn’t
come from large clinical trials alone (as Barter et al. noted in their quote given above), but is
instead to be complemented with other types of studies, for instance lab experiments.
Cases like the third, asbestos, point to the question of whether, in the light of the stunning
advancement of biomedicine, disease ought to be reduced to biochemical reactions in the body.
On the one hand, in the medical sciences, attention is given to social mechanisms of health and
disease, but these are often considered “distant,” as “adjunct” descriptions of how biological
mechanisms of health and disease correlate with socio-economic differences or inequalities across
individuals. On the other hand, the sociology of health and the behavioral sciences describe the
role that social, psychological, economic, or behavioral factors play in the etiology of disease.
To be sure, this is an area where evidence of difference-making (that social factors make a differ-
ence to the occurrence of disease) exceeds evidence of mechanisms (how social factors intervene
in the mechanisms of disease causation). The philosophy of mechanisms has a chance there in
323
developing an account of mixed mechanisms, where both biological and social factors play an
active role (Kelly et al. 2014). This is not just to elucidate the etiology of disease, but also to
better plan public health interventions. In fact, while most (non-communicable) diseases have a
proper explanation in terms of biochemical mechanisms, successful public health interventions
in the past targeted social determinates, lifestyles, and basic hygiene or sanitary measures; to be
sure, this is still the case nowadays, as evidenced by major public health interventions in Brazil
(see e.g. Barreto et al. 2010; Barreto and Aquino 2009). In a nutshell, bio-social mechanisms of
health and disease are needed to plan, research, and implement interventions. This is because
successful interventions need a complex view of what counts as relevantly beneficial or harmful.
The fourth case, H. pylori and gastric ulcer, is interesting in that it shows that mechanisms
do not function as an “experimentum crucis” for a theory, nor are they “deductively” inferred
from available theories. The search for explanatory and therapeutic mechanisms requires con-
stant interplay with available background knowledge. Hunting for specific mechanisms, like
that linking H. pylori infection with stomach ulcers, might confirm or disconfirm existing back-
ground knowledge. It is undeniable that experiments depend on mechanisms; more precisely,
on knowledge of mechanisms. However, without standards to assess the quality of knowledge of
mechanisms, poor-quality assertions and speculation (“the stomach is sterile”) appear to fill the
gaps that should instead be populated with empirically based, carefully researched knowledge
about mechanisms. This allows us to introduce a normative dimension of the mechanism pro-
ject in medicine, next to the descriptive one carried out in the previous section. Mechanisms
support the infrastructure of medical inference. To be even more precise, they support a vari-
ety of medical inferences at different stages of the scientific process. They aren’t just a way of
expressing existing knowledge, but instead actively participate in producing new knowledge.
Mechanisms have a vital role in the setup of trials or of lab experiments. And the results of these
scientific practices may lend further support to these mechanisms, or they may lead us to signifi-
cantly revise our knowledge base (as Craver and Darden 2013 suggest).
The fifth case, puerperal fever, is an instructive case for several reasons. One is that we should
avoid “presentism”: it is wrong to assess the past with the theories available today. The recent
debate on evidential pluralism (see Broadbent 2011) focuses on whether the scientific commu-
nity, at the time of Semmelweis, was right or wrong in rejecting his preventive measure. The
point here is that this measure was not supported by a solid theoretical background, as this was
happening well before the germ theory of disease had been developed. Given what we know
today about bacteria and so on, it is too easy a judgment to say that the scientific community
at that time was wrong in rejecting Semmelweis. The point here is of course broader than just
this specific episode. It is a point of philosophical methodology and of how to do philosophy of
science, properly informed by the history of science. Yet much is historically disputed—indeed,
“misleading” (Tulodziecki 2013: 1074) in the Semmelweis case. Another reason why this epi-
sode is relevant to us is that, even if the historical quarrel cannot be settled, this can provide
important lessons about the present. In particular, examples like these should encourage reflec-
tion on what to do when knowledge of mechanisms is incomplete or highly uncertain. What
actions are nonetheless justified? This is where proper attention to mechanisms in philosophy
of medicine pushes back the frontier of epistemology and of methodology. In fact, exquisitely
epistemological questions about the (un)certainty of current medical knowledge quickly turn
into ethico-political questions about what justifies the implementation (or refusal) of interven-
tions or preventive measures. A case in point here is the current controversy over vaccination
programs. Apart from the disputed question of the correlation of vaccines and autism, another
question concerns the tension between the individual freedom not to vaccinate versus the pub-
lic health duty to protect the population at large by promoting vaccination to all individuals.
324
Health, disease, and worldviews, the sixth episode, is an area where much work is needed,
from a philosophical, historical, and also socio-anthropological perspective. These sorts of
comparison are instructive in that they remind us that mechanisms, their entities and activi-
ties, are not given. They are instead products of our investigations into the phenomena of
health and disease. From an epistemological point of view, this means that mechanisms are
part of our inferential systems—a point that already arose earlier in the reflection on the case
of Helicobacter pylori. But mechanisms are also a part of our inferential system in another sense:
different understandings of the phenomena of health and disease may lead to incommensurable
mechanisms, or to syndromes that are culture-bound (Guarnaccia and Rogler 1999). Thus all
these phenomena, as well as the investigations and explanations thereof, are deeply embed-
ded in cultural factors, in the East as well as in the West. In our sixth episode, this was shown
in the way the nervous system was conceptualized differently in Western and in Chinese
medicine, and in the different terms used to refer to parts of the body and to its functioning.
This should warn us about the perils of adopting a naïvely scientistic attitude and imperialistic
imposition of Western standards in an uncritical way.
4. Evidence and contrastive focus

The sextet presented in section 2, and then analyzed in section 3, reveals many issues of interest
to those thinking about mechanisms. While we would like to follow the episodes with a detailed
and exhaustive theoretical discussions of the issues, we will restrict ourselves to a discussion of
just two of them. The first—evidence of mechanism—is a hotly disputed issue in the current
debate. We will develop our account of evidence of mechanism in medicine from the six epi-
sodes, as a useful contribution. The second—contrastive focus—has not (we think) received
sufficient attention from philosophers interested in mechanisms. We therefore highlight it here
by way of raising interest in it as a research problem more widely.
We begin by considering evidence, reasoning, and narratives in medicine. As the aspirin
example suggests, mechanisms do not typically appear as ready-made pieces of knowledge. A
great deal of work is needed to acquire and establish knowledge of mechanism. Here, the notion
of evidence plays a crucial role. We borrow from Illari the definition of evidence of mechanism.
She takes evidence of mechanisms as being “evidence of the actual existence of the postulated
mechanism linking cause and effect” (Illari 2011: 120), and we share her position.
While schematic, this definition is sufficiently clear to warrant some initial clarification regard-
ing our thoughts about the use of mechanisms in medicine as evidence. What is at stake, in fact, is
a distinction between evidence of mechanism and mechanistic reasoning, around which substantial
disagreement exists. For instance, Howick et al. define mechanistic reasoning as follows: “the
inference from mechanisms to claims that an intervention produced a patient-relevant outcome.
Such reasoning will involve an inferential chain linking the intervention (such as antiarrhythmic
drugs) with the outcome (such as mortality)” (Howick et al. 2010: 434).2
We take this to mean that mechanistic reasoning would involve making a clinical deci-
sion about the likely efficacy or harms of a medical intervention just by thinking about
mechanisms. We do not endorse this way of using mechanisms to make medical decisions as
a normative aspiration. Nor do we think it is a good description of the generality of medical
practice (it seems hard to connect to our episodes). We instead think that reasoning about
mechanisms alone is highly unlikely to make for reliable clinical inferences. As we have
argued in other places (Clarke et al. 2013, 2014), a kind of pragmatic evidential pluralism
featuring some contribution from evidence of relevant mechanisms may instead help make
better medical inferences. According to this view, evidence of mechanism participates, along
325
with other kinds of relevant evidence such as that produced by randomized clinical trials
(RCTs), in various inferential practices. For the purposes of this chapter, though, it suffices
to note that our evidence of mechanism is just one kind of evidence (amongst many), rather
than the ambitious and free-standing mechanistic reasoning discussed by Howick et al. Along
with Solomon (2015: 123–4), we therefore note that evidence of mechanism and mechanistic
reasoning cannot be used interchangeably.3
A further distinction that we wish to draw is between evidence of mechanism and narratives.
While the role of narratives in medicine (see Kleinman 1989; Greenhalgh 1999; Montgomery
2006) and in science more generally (Roth 1988, 1989) is a topic that deserves more philo-
sophical attention than it gets, it is important to note that we think there is a difference between
evidence of mechanism and of the narrative(s) that may be used to describe, present, discuss, or
criticize said evidence of mechanism.
It’s possible to use (empirically grounded) narratives to describe mechanisms, but that is not
the same thing as saying that mechanisms just are narratives (see also Chapter 31). We think
that any scientific result that is expressed in ordinary language can be understood as a narrative,
without the converse assertion that all scientific results are just narratives. More positively, we
think that there are standardized rules that govern the relation between a result and its narratives,
which supply rigid structure on the presentation of those narratives. Consider how a scientific
article must present a research question, the data, the analyses, and the results.
We turn now to our second main point. One aspect that many mechanisms in the episodes
above share is contrastive focus. Mechanisms, as also discussed elsewhere in this volume (see
Chapter 16), help answer the question “How does C cause E?” Medicine is no exception to this
use of mechanisms. Typically, emphasis is given to those characteristics of mechanisms (most
notably, organization) that participate in explaining how C causes (in the sense of producing) E.
In this section, though, we want instead to draw attention to another way that we might
interpret the question “How does C cause E?” Rather than give an answer that highlights the
productive continuity that obtains between C and E, we might instead give an answer that
explains why our C causes E rather than E1, or that C, rather than C1, causes E. This is the ques-
tion of contrastive focus, which has been fairly extensively discussed in relation to causality (Dretske
1972; Schaffer 2005; Northcott 2008). Here, we aim to show that attention to contrastive focus
is key to several inferential practices in medicine, and that dealing with questions of contrastive
focus often involves thinking about mechanisms, in different ways, also illustrated by the sextet
of episodes. We highlight here two specific aspects of contrastive focus as applied to medicine:
(i) normal vs pathological and (ii) biological vs social. Simply put, the idea is that the aspects of
contrastive focus we discuss here concern (i) different mechanisms describing normal or patho-
logical behavior and (ii) the factors—biological or social—that are at work in such mechanisms.
To begin with, mechanisms that deal with the etiology of pathology typically do so in a
contrastive way with normal physiology (see also Chapter 8). For example, the mechanism that
explains how infection with H. pylori leads to ulcers does so in a contrastive way with normal
physiology (see the extremely thorough review by Kusters, van Vliet, and Kuipers 2006 for
further details). This means that we need a generic mechanism about normal physiology and a
suitable contrastive model of a mechanism that explains pathogenic behavior. Many experi-
mental studies in biomedicine are based on this idea—they effectively compare subjects that do
and do not receive some intervention. The very same idea is at the basis of extrapolation from
animal studies to humans (see e.g. Steel 2007).
This means that contrastive focus is at the basis of RCTs too. In fact, one way to use RCTs
is precisely to differentiate between the effect E1 rather than E2, or between the role of C1 rather
than C2. This adds to the recent debates on the role of mechanisms in RCTs. Specifically,
326
according to evidence-based medicine (EBM)-theorizers, mechanisms play little or no role at all

in establishing causal relations—a view challenged by supporters of “evidential pluralism,” who
hold the view that to establish causal relations one typically needs multifarious evidence (for a
discussion, see Clarke et al. 2014).
The contrastive focus can also participate in establishing causal relations at the individual,
single-case level, precisely using considerations about what is (thought to be) normal and what
is not. This, however, is far from being an easy task, as testified by the several challenges of
diagnosis. We shall focus here on just one challenge. To make a diagnosis in the single case, we
need to contrast it with a generic mechanism. But how generic should a generic mechanism be?
There is no principled answer to this question. One problem has to do with finding
mechanisms that are stable in sufficiently homogeneous reference classes. A reference class is
homogenous when all its members have the same characteristics. This is usually not the case in
medicine—we expect those in the reference class of “patients with heart disease” to be suffer-
ing from a range of different conditions. Homogeneous reference classes—where all individuals
have exactly the same conditions—are either very scarce or very hard to find in medicine.
A notable example in this respect is the study of pathologies that frequently co-occur with
other pathologies. As the HDL-raising drug example above shows, it can be extremely hard to
unpick the contributions of different conditions with similar effects. In the HDL-raising case,
we can presumably assume that this isn’t just one mechanism at work, but several nested and
interconnected mechanisms. This is because people using HDL-raising drugs have very differ-
ent conditions, and most of the time multiple pathologies, not just heart disease. Consequently,
the contrastive focus of mechanisms should not simply be in terms of differentiating the effect
of a drug against a placebo. Instead, drugs should be tested taking these other mechanisms into
account, including interactions with other drugs.
The second aspect of the issue of contrastive focus is whether only biological factors should
contribute to answering the question “How does C cause E?” We encountered this issue in
the asbestos example, where a relevant question is whether workplace or other social factors are
appropriate to be included in the etiology of the disease. So, when we check for appropriate
contrasts between reference classes, we shouldn’t just consider biochemical factors (i.e. those
biochemical factors marking normal vs pathological behavior), but also psycho-demo-socio-
economic ones. Asbestos is of course not the only case. Another example is epigenetics, which
attempts to reconstruct the effect of the environment, including in utero events, at the genetic
level even one generation later. The contrastive focus here might be with events that happened
much earlier on in our lives, if not generations before. So the mechanisms of health and disease
may be acting over long time spans, and the causes may reside in events that cannot be assessed
at the time of diagnoses. Studies on the “Dutch famine” of 1944–5 are a good case in point.
According to these studies, the children of women that were pregnant during the famine are
more susceptible to develop diseases such as obesity, diabetes, cardiovascular conditions, etc.
Epigenetics attempts to shed light on the long-term mechanisms that, by affecting women dur-
ing pregnancy, also affected their offspring even decades after birth (see Roseboom, de Rooij,
and Painter 2006 for an introduction).
The health sciences (including sociology of health) have recognized for a long time that
health and disease are not independent of socio-economic factors (for an overview, see Kelly
et al. 2014). But this recognition is typically explicated in two ways. On the one hand, socio-
economic factors are merely correlated with health and disease, and they are used to map which
parts of the population are healthier and which are more ill. But here socio-economic factors do
not actively participate in the production of health and disease. On the other hand, the action of
psycho-socio-demo-economic factors is totally explained away by reducing them to the action
327
of biochemical factors. Here, a proper understanding of how these factors actively contribute to
health and disease is missing. Differently put, the mechanisms of health and disease are not just
biochemical or just socio-eco-psychological. They are most often mixed, a blend of those cat-
egories, and in need of further investigation within medicine and public health, and philosophy
too. Most non-communicable diseases, from obesity to type-2 diabetes, and alcohol-related
diseases arguably require mixed mechanisms both for understanding and for intervening on their
causes and effect (see also Kelly et al. 2014). Admittedly, these (and many others) are cases where
what we don’t know exceeds by far what we know. And, precisely for this reason, thinking in
terms of how mechanisms contribute to several inferential practices (rather than just working
out The-One definition of mechanism) may be of great help.
5. Conclusions
As shown by the above sextet of examples, we have a very broad and inclusive understanding
of what medicine is and does. We hope that this will help redress part of the existing debates in
philosophy of medicine, particularly for what we call the “narrow view of medicine,” held by
a number of authors, both from medicine and from philosophy.
For many philosophers in recent years, medicine has been largely synonymous with EBM.
This is perhaps an artifact, due to how philosophers of science came to pay attention to medi-
cine. The establishment of EBM as a dominant paradigm prompted a peculiar reaction in the
philosophical community, which was captured by the iconic paper by John Worrall (2002).
He asked: “What on earth was medicine based on before?” This might explain why so much
attention was, since then, devoted to this sub-field of medicine. There is much more to medi-
cine than evaluating the efficacy and harms of drugs. At the same time, particularly in the
way that philosophical work on EBM intersects with other interests of analytic philosophers
of science (particularly the philosophy of statistics), we see just why this narrowing might be
both useful and comforting. However, when it comes to thinking about mechanisms, we
worry about this narrowness, largely because (we think) it has led to the pre-eminence of a
narrow set of related questions that are ostensibly about medicine, but are really about statisti-
cal inference in the context of the clinical trial: evidence, bias, and so on. In turn, this has led
to a caricature of medical practice such that mechanisms are either (a) described as playing a
negligibly minor evidential role or (b) treated as a simple (but faulty!) alternative method of
inference to the clinical trial. Both of these caricatures, we submit, are inaccurate. It is high
time to broaden the scope of the debate, by looking at the very many practices in medicine,
including, but not restricted to, EBM.
The approach taken in this chapter points to a pluralistic account of medicine, and of mecha-
nisms in medicine. This stems from the different ways in which mechanisms contribute to
medical practices. It is in this sense that we have explored the applied epistemology of mecha-
nisms in medicine. We submit that mechanisms can be used as powerful conceptual tools for
establishing and intervening on causal relations in medicine. However, we worry that insuf-
ficient attention has been paid to the different ways that an understanding of mechanisms can
contribute to this applied epistemology. This is all the more important, because the philosophy
of mechanisms itself became somewhat specialized, discussing mechanisms in specific, isolated
domains: in biology, in psychology, or in sociology. But mechanisms, in medicine or elsewhere,
are seldom intrinsically biological, psychological, or sociological. We hope that our approach
to mechanisms in medicine will be an encouragement to start crossing the borders of these sub-
fields in the philosophy of mechanisms.
328
Notes
1 The authors gratefully acknowledge financial support from the Arts and Humanities Research Council
(AHRC) (UK) for the research done under grant number AH/M005917/1 (“Evaluating Evidence in
Medicine”).
2 Howick also gave a similar formulation in a later solo article: “Mechanistic reasoning is an inferential
chain (or web) linking the intervention (such as HRT) with a patient-relevant outcome, via relevant
mechanisms” (Howick 2011: 929).
3 A brief clarification is in order here. We do not agree that our previous work (Clarke et al. 2013, 2014)
endorsed the equivalence of mechanistic reasoning and mechanistic evidence, as Miriam Solomon seems
to suggest: “Although there are differences between Howick (2011b), Andersen (2012), and Clarke et al.
(2013), they all share the assumption that mechanistic reasoning should be regarded as mechanistic
evidence, moreover, evidence that has a place in the hierarchy. This is an assumption that I challenge”
(Solomon 2015: 120).
References
Allen, D., and L. Kazan-Allen. 2012. Eternit and the Great Asbestos Trial. London: The International Ban
Asbestos Secretariat. http://www.ibasecretariat.org/eternit-great-asbestos-trial-toc.htm.
Andersen, H. 2012. Mechanisms: What are they evidence for in evidence-based medicine? Journal of
Evaluation in Clinical Practice. 18(5): 992–9.
Antithrombotic Trialists’ Collaboration. 2002. Collaborative meta-analysis of randomised trials of anti-
platelet therapy for prevention of death, myocardial infarction, and stroke in high risk patients. British
Medical Journal. 324(7329): 71–86.
Barreto, M. L., and Aquino, R. (2009). Recent positive developments in the Brazilian health system.
American Journal of Public Health, 99(1): 8. doi:10.2105/AJPH.2008.153791.
Barreto, M. L., Genser, B., Strina, A., Teixeira, M. G., Assis, A. M. O., Rego, R. F., . . . Cairncross, S.
2010. Impact of a citywide sanitation program in northeast Brazil on intestinal parasites infection in
young children. Environmental Health Perspectives. 118(11): 1637–42. doi:10.1289/ehp.1002058.
Barter, Philip J., Caulfield, M., Erikson, M. . . . Brewer, B. 2007. Effects of torcetrapib in patients at high
risk for coronary events. New England Journal of Medicine. 357:2109–22. doi:10.1056/NEJMoa0706628.
Broadbent, A. 2011. Inferring causation in epidemiology: Mechanisms, black boxes, and contrasts. In Illari,
P. M., Russo, F. and Williamson, J. (eds). Causality in the Sciences. Oxford University Press. 45–69.
Brousseau, M. E., Schaefer, E. J., Wolfe, M. L., Bloedon, L. T., Digenio, A. G., Clark, R. W., Mancuso,
J. P., and Rader, D. J. 2004. Effects of an inhibitor of cholesteryl ester transfer protein on HDL choles-
terol. New England Journal of Medicine. 350(15): 1505–15. doi:10.1056/NEJMoa031766.
Chang, H. 2011. Beyond case-studies: History as philosophy. In Mauskopf, Seymour and Schmaltz, Tad
(eds). Integrating History and Philosophy of Science. Springer. 109–24.
Clarke, B., Gillies, D., Illari, P., Russo, F., and Williamson, J., 2013. The evidence that evidence-based
medicine omits. Preventive Medicine. 57(6): 745–7.
Clarke, B., Gillies, D., Illari, P., Russo, F., and Williamson, J. 2014. Mechanisms and the evidence hierar-
chy. Topoi. 33(2): 339–60. doi:10.1007/s11245-013-9220-9.
Clarke, B., and Russo, F., 2016. Causation in medicine. In J. Marcum (ed.). Companion to Contemporary
Philosophy of Medicine. Bloomsbury. 297–322.
Craver, C. F., and Darden, L. 2013. In Search of Mechanisms: Discoveries across the Life Sciences. University of
Chicago Press.
Dretske, F. I. 1972. Contrastive statements. The Philosophical Review. 81(4): 411–37.
Gillies, D. 2005. Hempelian and Kuhnian approaches in the philosophy of medicine: The Semmelweis
case. Studies in History and Philosophy of Science Part C. 36(1): 159–81. doi:10.1016/j.shpsc.2004.12.003.
Greenberg, M. 2003. Cape asbestos, Barking, health and environment; 1928–1946. American Journal of
Industrial Medicine. 43(2): 109–19.
Greenhalgh, T. 1999. Narrative-based medicine in an evidence-based world. BMJ. 318(7179): 323–5.
Guarnaccia, P. J., and Rogler, L. H. 1999. Research on culture-bound syndromes: New directions.
American Journal of Psychiatry. 156(9): 1322–7. doi:10.1176/ajp.156.9.1322.
Howick, J. 2011. Exposing the vanities – and a qualified defence – of mechanistic evidence in clinical
decision-making. Philosophy of Science. 78(5): 926–40. Proceedings of the Biennial PSA 2010.
329
Howick, J., Glasziou, P., and Aronson, J. K. 2010. Evidence-based mechanistic reasoning. Journal of the
Royal Society of Medicine. 103(11): 433–41. doi:10.1258/jrsm.2010.100146.
Hutton, J. 2012. Composite paradigms in medicine: Analysing Gillies’ claim of reclassification of disease
without paradigm shift in the case of Helicobacter pylori. Studies in History and Philosophy of Science
Part C. 43(3): 643–54. doi:10.1016/j.shpsc.2012.05.003.
IARC Working Group. 2012. Arsenic, metals, fibres and dusts. International Agency for Research on Cancer.
http://monographs.iarc.fr/ENG/Monographs/vol100C/index.php.
Illari, P. M. 2011. Mechanistic evidence: Disambiguating the Russo-Williamson Thesis. International
Studies in the Philosophy of Science. 25(2): 139–57.
Illari, P. and Russo, F., 2014. Causality: Philosophical Theory Meets Scientific Practice. Oxford: Oxford
University Press.
Illari, P. M., and Williamson, J. 2012. What is a mechanism?: Thinking about mechanisms across the
sciences. European Journal of the Philosophy of Science. 2: 119–35.
Kelly, M. P., Kelly, R. S., and Russo, F. 2014. The integration of social, behavioural, and biological
mechanisms in models of pathogenesis. Perspectives in Biology and Medicine. 57(3): 308–28.
Kleinman, Arthur. 1989. Illness Narratives: Suffering, Healing and the Human Condition. Basic Books.
Kusters, J. G., van Vliet, A. H. M., and Kuipers, E. J. 2006. Pathogenesis of Helicobacter pylori infection.
Clinical Microbiology Reviews. 19(3): 449–90 doi:10.1128/CMR.00054-05.
La Vecchia, C., and Boffetta, P. 2014. A critique of a review on the relationship between asbestos exposure
and the risk of mesothelioma: Reply.” European Journal of Cancer Prevention. 23(5): 494–6. doi:10.1097/
CEJ.0000000000000051.
Lehrer, Jonah. 2011. Trials and errors: Why science is failing us. Wired. https://www.wired.com/2011/12/
ff_causation/all/1.
Montgomery, K. 2006. How Doctors Think: Clinical Judgment and the Practice of Medicine. Oxford University
Press.
Mossano, Silvana. 2011. The Eternit trial: The verdict is close. Epidemiologia E Prevensione. 35(3–4): 175–7.
Northcott, R. 2008. Causation and contrast classes. Philosophical Studies. 139(1): 111–23.
O’Brien, J. R. 1968. Effects of salicylates on human platelets. Lancet 1: 779–83.
Quick, A. J. 1967. Salicylates and bleeding: The aspirin tolerance test. The American Journal of the Medical
Sciences. 252(3): 265–9.
Roseboom, T., de Rooij, S., and Painter, R. 2006. The Dutch famine and its long-term consequences for
adult health. Early Human Development. 82(8): 485–91.
Roth, P. A. 1988. Narrative explanations: The case of history. History and Theory. 27(1): 1–13.
Roth, P. A. 1989. How narratives explain. Social Research. 51(2): 449–78.
Russo, F., and Williamson, J. 2007. Interpreting causality in the health sciences. International Studies in the
Philosophy of Science. 21(2): 157–70.
Schaffer, J. 2005. Contrastive causation. The Philosophical Review. 114(3): 297–328.
Schrör, Karsten. 1997. Aspirin and platelets: The antiplatelet action of aspirin and its role in thrombosis treat-
ment and prophylaxis. Seminars in Thrombosis and Hemostasis. 23(4): 349–56. doi:10.1055/s-2007-996108.
Shapiro, H. 2003. How different are Western and Chinese medicine? The case of nerves. In Selin,
Helaine (ed.). Medicine across Cultures: History and Practice of Medicine in Non-Western Cultures. Springer.
351–72.
Smith, J. B., and Willis, A. L. 1971. Aspirin selectively inhibits prostaglandin production in human
platelets. Nature New Biol 231: 235–7.
Solomon, M. 2015. Making Medical Knowledge. Oxford University Press.
Steel, D. 2007. Across the Boundaries: Extrapolation in Biology and Social Science. Oxford University Press.
Tall, Alan R., Yvan-Charvet, Laurent, and Wang, Nan. 2007. The failure of torcetrapib. Arteriosclerosis,
Thrombosis, and Vascular Biology. 27(2): 257–60. doi:10.1161/01.ATV.0000256728.60226.77.
Terracini, B., Mirabelli, D., Magnani, C., Ferrante, D., Barone-Adesi, F., and Bertolotti, M. 2014. A cri-
tique to a review on the relationship between asbestos exposure and the risk of mesothelioma. European
Journal of Cancer Prevention. 23(5): 492–4. doi:10.1097/CEJ.0000000000000057.
Thagard, P. 1998a. Ulcers and bacteria I: Discovery and acceptance. Studies in History and Philosophy of
Science Part C. 29(1): 107–36. doi:10.1016/S1369-8486(98)00006-5.
Thagard, P. 1998b. Ulcers and bacteria II: Instruments, experiments, and social interactions. Studies in
History and Philosophy of Science Part C. 29(2): 317–42. doi:10.1016/S1369-8486(98)00024-7.
330
Tulodziecki, D. 2013. Shattering the myth of Semmelweis. Philosophy of Science. 80(5): 1065–75.
doi:10.1086/673935.
Vane, J. R. 1971. Inhibition of prostaglandin biosynthesis as a mechanism of action of aspirin-like drugs.
Nature New Biology. 231: 232–5.
Weiss, H. J., Aledort, L. M., and Kochawa, S. 1968. The effect of salicylates on the hemostatic properties
of platelets in man. Journal of Clinical Investigation. 47: 2169–80.
Woody, Andrea I. 2015. Re-orienting discussions of scientific explanation: A functional perspective.
Studies in History and Philosophy of Science Part A. 52: 79–87. doi:10.1016/j.shpsa.2015.03.005.
Worrall, J. 2002. What evidence in evidence-based medicine? Philosophy of Science. 69(S3): S316–S330.
331
25
DEVELOPMENTAL
MECHANISMS1
Alan C. Love
1. Mechanisms in developmental biology

A developmental mechanism is a mechanism that operates during development or ontogeny.
Typically, the scope of development is narrowed to one of four broad problem domains that
constitute the majority of contemporary developmental biology: pattern formation, growth,
morphogenesis, and cellular differentiation (Love 2014, 2015). We can observe this in the fol-
lowing introductory discussion from Mechanisms of Morphogenesis:
The purpose of this book is to bring together in one place some of the most significant
advances that have been made in identifying mechanisms of morphogenesis from a
variety of species, systems and scales. . . . The morphogenetic mechanisms described
in later chapters tend to invoke, and ‘take for granted’ the fine-scale mechanisms
described in earlier chapters.
(Davies 2013, 5)
This passage assumes that there are different types of mechanisms operating in development:
morphogenetic mechanisms and fine-scale mechanisms. The latter is shorthand for molecular
genetic mechanisms and Davies thinks developmental biologists have fixated on them.
[Older descriptive embryology] books, which provided the foundations on which

programs of molecular biology could stand, have been succeeded in recent years
by large and solid works full of well-established mechanisms of signalling, pattern
formation and gene control. . . . The spotlight of developmental books and review
volumes is usually focused on the molecular biology of gene control and pattern for-
mation . . . current texts [focus on] molecular mechanisms of morphogenesis, taking
for granted . . . morphogenetic mechanisms.
(Davies 2013, 4)
Given that molecular mechanisms of morphogenesis (signaling or gene regulatory networks)

are distinguished from morphogenetic mechanisms (cell migration or epithelial invagination),
there is value in marking the distinction explicitly: molecular genetic mechanisms (MGMs) and
332
Developmental mechanisms
cellular-physical mechanisms (CPMs). This distinction plays a significant role in reasoning about
mechanisms that operate during ontogeny and explaining how animals and plants take shape. It
also serves exegetically to accent two conceptual issues: (1) the conservation or stable identity
of MGMs across phylogenetically disparate taxa compared with the non-conservation of CPMs,
and (2) the explanatory generality of MGMs compared with the explanatory completeness of
integrating these with CPMs to comprehend more fully how complex interactions yield mor-
phological outcomes in ontogeny.
The conservation of MGMs is central to the reasoning practices of contemporary develop-
mental biology and a major source of its recent success.
The past two decades have brought major breakthroughs in our understanding of
the molecular and genetic circuits that control a myriad of developmental events in
vertebrates and invertebrates. These detailed studies have revealed surprisingly deep
similarities in the mechanisms underlying developmental processes across a wide range
of bilaterally symmetric metazoans . . . [these] comparisons have defined a common
core of genetic pathways guiding development.
(Bier and McGinnis 2003, 25)
Despite dramatic differences in the phenomena of development (“myriad of developmental

events”), the “deep similarities” of genetic mechanisms underlying these phenomena are striking.
These “conserved mechanisms” serve as a foundation for reasoning from a small set of model
organisms to explanatory generalizations about how all or most animals develop (Love and
Travisano 2013; Ankeny and Leonelli 2011; Halina and Bechtel 2013). However, conserved
MGMs are not identical and therefore a question arises about how deep the similarities must
be to license these inferences. Additionally, mechanisms are individuated by the outcomes they
produce. Since the claim of conservation is a judgment of homology, which is based typically
on structure rather than function or behavior (Love 2007), what constitutes the individuation
conditions for a conserved MGM? And why is it the case that many CPMs, such as those associ-
ated with morphogenesis, appear not to be conserved?
The explanatory generality of MGMs is constitutive of much contemporary developmen-
tal biology. Many take this generality as a fundamental point of departure: “Developmental
complexity is the direct output of the spatially specific expression of particular gene sets and
it is at this level that we can address causality in development” (Davidson and Peter 2015, 2).
However, others disagree: “One might be tempted to think of development simply in terms of
mechanisms for controlling gene expression. But this would be highly misleading. . . . To think
only in terms of genes is to ignore crucial aspects of cell biology” (Wolpert et al. 1998, 15). The
concern expressed pertains to explanatory completeness; something is missing if we only focus
on MGMs. CPMs add crucial detail to mechanistic explanations. What are the prospects for
generating integrated explanations of development using both types of mechanisms? Could gen-
erality and completeness exhibit a tradeoff as explanatory values in the context of developmental
biology (Weisberg 2013; see also Chapter 17, this volume)?
To approach these two conceptual issues, I first provide a framework of interpretive catego-
ries and characterize examples of both types of mechanisms. Then I probe what is involved in
claims about the phylogenetic conservation of MGMs, as well as the difference between the
explanatory generality of MGMs compared with the potential explanatory completeness derived
from integrating these with CPMs. In closing, I identify one further conceptual issue—the
dynamic constitution and organization of MGMs—and provisionally conclude with two broad
lessons derived from my analysis of developmental mechanisms.
333
Alan C. Love
2. Mechanisms: philosophical background and

developmental examples
Philosophical explorations of mechanisms and mechanistic explanation have grown dramati-
cally over the past two decades and encompass a wide variety of sciences (Craver and Darden
2013; Craver and Tabery 2016; Illari and Williamson 2012). Although a number of different
accounts of mechanisms have been offered, four shared elements are discernable (Craver and
Tabery 2016; Illari and Russo 2014, ch. 12; see Chapter 1, this volume): (1) what a mecha-
nism is for, (2) its constituents, (3) its organization, and (4) the spatiotemporal context of its
operation. From these we can generate an ecumenical characterization to serve as a template
for guidance in our analysis.
A mechanism is constituted by a number of parts and activities or component operations

that are organized into patterns of interacting relationships within a particular spatiotem-
poral context so as to produce a specific behavior or phenomenon (or set thereof ).2
Mechanistic explanation works by decomposing systems into their constituent parts, localizing
their characteristic activities, and articulating how they are organized to produce a particular
effect at or within a specific time or place. These explanations illustrate and display the genera-
tion of specific phenomena at particular times and places by describing the organization of a
system’s constituent components and activities.
Consider an established MGM: the initial formation of segments in Drosophila due to the
segment polarity network of gene expression (Wolpert et al. 2010, 70–81; Damen 2007). By
Stage 8 of development (~3 hours post-fertilization), Drosophila embryos have 14 parasegment
units that were defined by pair-rule gene expression in earlier stages. The transcription factor
Engrailed accumulates in the anterior portion of each parasegment. This initiates a cascade
of gene activity that defines the boundaries of each compartment of cells that will eventually
become a segment. One element of this activity is the expression of hedgehog, a secreted signal-
ing protein, in cells anterior to the band of cells expressing engrailed, which marks the posterior
boundary of each nascent segment. This, in turn, activates the expression of wingless, another
secreted signaling protein, which maintains the expression of both engrailed and hedgehog in a
feedback loop so that segment boundaries persist.
This mechanism description can be expanded to illustrate the productive continuity of inter-
actions involved in the feedback loop. On one side, the Hedgehog signaling pathway is activated
when Hedgehog binds to the membrane protein Patched (Lum and Beachy 2004). In the absence
of Hedgehog, Patched inhibits another membrane protein (Smoothened). Once Hedgehog
binds to Patched, Smoothened is able to block the production and operation of repressors of
the transcription factor Cubitus interruptus, which then turns on the expression of wingless. On
the other side, Wingless (a member of the Wnt protein family) jointly binds to a member of the
Frizzled family and lipoprotein receptor-related membrane protein, which disrupts a complex of
proteins that continually degrade β-catenin in the cytoplasm. The phosphoprotein Dishevelled
is also activated, further blocking this complex from operating by anchoring it to the plasma
membrane. β-catenin then accumulates and is able to reach the nucleus in sufficient concentra-
tions to initiate transcription and expression of engrailed and hedgehog.
The segment polarity network is a MGM that exhibits the features of our ecumenical char-
acterization. It is constituted by a number of parts (e.g., Engrailed, Wingless, Hedgehog) and
activities or component operations (e.g., signaling proteins bind receptors, transcription factors
bind to DNA and initiate gene expression), which are organized into patterns of interacting
334
relationships (the positive feedback loop, the Wnt and Hedgehog signaling pathways) within a
spatiotemporal context (in parasegments of the Drosophila embryo, ~3 hours post-fertilization)
so as to produce a specific behavior or phenomenon (a set of distinct segments with well-
defined boundaries). The Wnt and Hedgehog signaling pathways are also distinct mechanisms
(Craver and Tabery 2016; Lum and Beachy 2004; van Amerongen and Nusse 2009). The posi-
tive feedback loop of the segment polarity network of gene expression has these as component
mechanisms. However, whereas these component mechanisms are stable and present through-
out the organism’s life history, the segment polarity network is transient and operative for only a
specific period of ontogeny. The expression of hedgehog and wingless becomes decoupled later in
embryogenesis so that the feedback loop is no longer present. There is a change in the organiza-
tion of the mechanism as the spatiotemporal context changes subsequent to the production of
a particular outcome.
Having observed a MGM, consider the CPM of branching morphogenesis, which refers
to combinations of cellular proliferation and movement that yield branch-like structures in
kidneys, lungs, glands, or blood vessels. There are many types of branching morphogenesis,
but one primary mechanism is epithelial folding, which involves cells invaginating at different
locations on a structure to yield branches (Davies 2013, ch. 20). Different CPMs can produce
invaginations that lead to branching structures (Varner and Nelson 2014): the constriction
of one end of a subset of columnar cells in an epithelium (“apical constriction”); increased
cell proliferation of one epithelial sheet in relation to another (“differential growth”); and
compression of an epithelium leading to periodic invaginations (“mechanical buckling”).
That different mechanisms can lead to the same morphological outcome means it can be
difficult to discern which mechanism is operating in an embryonic context. Sometimes diver-
gent requirements of different CPMs can be isolated, such as the stiffness of the extracellular
matrix, and divergent predictions can be identified for cellular properties, such as the dynam-
ics of cell shape changes, which permit different CPMs to be teased apart and confirmed to
apply to a particular system (Davidson et al. 1995).
CPMs like apical constriction bear complex relations to MGMs. The transcription factor
Trachealess initiates a signaling cascade during Drosophila embryogenesis that leads to apical
constriction and thereby epithelial invagination to form the reticulate system of branching mor-
phology that is characteristic of the trachea (Affolter and Caussinus 2008). However, CPMs
are a distinct type of mechanism and exhibit the features of our ecumenical characterization
differently. The parts constituting the CPM are no longer transcription factors or signaling
molecules but cells and tissues. The activities or component operations (e.g., apical constriction,
differential growth, mechanical buckling) are organized into patterns of interacting relation-
ships (apical constriction leading to epithelial invagination) within a spatiotemporal context
(in tracheal precursors within the Drosophila embryo around Stage 7 and 8). This organization
produces a specific behavior or phenomenon (a set of branching structures—the trachea).
It is tempting to interpret the complex relations between MGMs and CPMs in terms of
nested mechanisms. The MGMs generating apical constriction could be considered component
mechanisms of the CPM of branching morphogenesis. However, caution is required because
MGMs are stable over evolutionary time whereas associated CPMs vary significantly.
The molecular signaling pathways [MGMs] that control branching morphogenesis appear
to be conserved across organs and species. However, despite this molecular homology,
recent advances in cell lineage analysis and real-time imaging have uncovered surprising
differences in the [cellular-physical] mechanisms that build these diverse tissues.
(Varner and Nelson 2014, 2750)
335
Alan C. Love
This evolutionary decoupling of the different types of mechanisms brings us to our first issue:
what does it mean to label MGMs as “highly conserved”?
3. Conserved genetic mechanisms of development

The Wnt signaling pathway is found across animal phyla (van Amerongen and Nusse 2009;
Komiya and Habas 2008). Whether biologists look in fruit flies, soil worms, or humans, this
pathway is present in some form; it is a conserved MGM. The claim of conservation is a judg-
ment of homology. The core idea behind homology is identifying the same trait in different
taxa under every variety of form and function, where the sameness derives from a common
evolutionary heritage due to descent with modification. Similarity is neither necessary nor suffi-
cient for determining whether two traits are homologous (Ghiselin 2005). Many morphological
features are similar because of natural selection (i.e., analogous), but not as a result of common
descent (i.e., homologous).
The application of homology reasoning in the context of conserved mechanisms has not
received sustained attention (but see Halina and Bechtel 2013; Love 2007). Halina and Bechtel
(2013) offer the following definition: “A biological mechanism is conserved when it can be
identified as the product of evolutionary descent” (1201). To some degree, this is too liberal. All
biological mechanisms are the product of evolutionary descent. The suppressed premise con-
cerns the comparative nature of the judgment. A biological mechanism cannot be “conserved”
except with respect to another mechanism in another taxon. Thus, the Wnt signaling pathway
is designated as conserved because it is found in fruit flies and soil worms and humans. Yet why
do we think the “same” Wnt signaling pathway is present in these taxa? Is it because of similari-
ties in the pathways? These could be due to natural selection operating at the molecular level
in these taxa, leading to similar (but not conserved) mechanisms (Halina and Bechtel 2013).
This highlights a critical link between our understanding of what a conserved mechanism is and
the criteria for determining whether a mechanism is genuinely conserved. Standard criteria for
identifying homologs include relative position in the body with respect to other traits, similarity
of structural detail, special quality, and embryological origin. Each of these is defeasible since
they can manifest because of other factors (e.g., similarity due to natural selection) or fail despite
homology obtaining (e.g., many homologous traits do not share the same embryological origin
because of developmental system drift; Haag 2014).
What might count as criteria for a conserved mechanism? One thing to note is that phy-
logenetic reconstruction is necessary but not sufficient. Thus, the claim that “conservation is
ultimately established through phylogenetic analyses” (Halina and Bechtel 2013, 1201) is not
strictly true because reconstructing a phylogeny does not necessarily involve applications of
particular criteria of homology for the trait (or mechanism) in question. (Methodologically, one
should explicitly avoid including the trait in question when constructing a phylogeny to test for
conservation.) More constructively, we can use our ecumenical characterization and explicitly
map its four features onto the standard criteria for homology.
The first mapping is between the constituents of a mechanism and the criterion of special
quality. If the same constituents are identified as standard components (e.g., Wnt proteins), then
this is a particular feature (e.g., an initiating signaling molecule) that is pertinent to determin-
ing whether something is conserved. This accounts for why many conserved mechanisms are
named in terms of constituent molecules that have these special qualities (e.g., Wnt, Hedgehog,
Hox). The second mapping is between the organization of a mechanism and similarity of struc-
tural detail. If a mechanism is organized in similar ways in terms of which families of molecules
interact with one another (e.g., in the segment polarity network), then this is another potential
336
indicator of conservation. This supplements the first criterion because many of the molecules
whose name labels a conserved mechanism are involved in other molecular processes that do
not exhibit the same organization among participating constituents. The third mapping obtains
between the criterion of relative position and the spatiotemporal context of a mechanism. The
conserved mechanism of establishing anterior-posterior axes in the embryo using combinations
of Hox gene expression occurs at distinct temporal junctures and spatial regions during develop-
ment (e.g., early on in the entire embryo, but later only in appendages).
The fourth mapping introduces a twist into our analysis. A stereotypical behavior or phe-
nomenon produced by a mechanism does not correspond to any of the criteria, and for a good
reason. If a homolog is the same under every variety of form and function, then a stereotypi-
cal behavior or function that does not vary clashes with typical homology judgments based on
structure. Unlike most traits evaluated for homology, the individuation of mechanisms is func-
tional, not structural:
The boundaries of a mechanism—what is in the mechanism and what is not—are fixed

by reference to the phenomenon that the mechanism explains. The components in a
mechanism are components in virtue of being relevant to the phenomenon.
(Craver and Tabery 2016)
In essence, change the phenomenon or behavior and you change what counts as the mecha-
nism. Yet similarity of function is a problematic criterion of homology (Abouheif et al. 1997);
what a trait does typically should not enter into an evaluation of homolog correspondence
because similarity of function is often the result of adaptation via natural selection to com-
mon environmental demands, not common ancestry. As a consequence, the idea of “functional
homology” has long been thought suspect, though there are ways to recover a conceptually
coherent notion of homology of function (Love 2007). However, this strategy assumes that a
functional trait can be defined in terms of its activity (what it is) rather than use (what it is for).
Is the notion of a “conserved genetic mechanism” inherently problematic because of its reliance
on functional individuation?
Given that conserved MGMs play important roles in the reasoning of developmental biolo-
gists (Love and Travisano 2013; Ankeny and Leonelli 2011; Halina and Bechtel 2013), we
should be hesitant to jettison it outright and instead attempt to account for these roles. One
plausible strategy for understanding the situation is to recognize that claims about conserved
mechanisms are not simply claims about homology. A judgment of correspondence for mecha-
nisms across diverse taxa does not hold under every variety of form and function. Instead, we
can use our analogies between the traditional criteria for homology and those for conserved
genetic mechanisms to recognize a richer characterization:
Conserved genetic mechanisms are shared, derived traits composed of particular con-
stituents, organized in a specific way, and found in delimitable spatiotemporal contexts
where they manifest a stereotypical behavior or phenomenon.
This is more circumscribed than the typical understanding of homology because sameness is not
maintained under every variety of form and function; substantive claims about the sameness of
constituents, organization, context, and function are involved. Importantly, this characterization
underwrites the role of conserved MGMs in securing explanatory generality from investigating
model organisms: “researchers assume deep commonalities between the mechanisms such that
when a part or operation is found within a model mechanism, they expect to find it and search
337
Alan C. Love
for it within the target mechanism” (Halina and Bechtel 2013, 1202). A typical judgment of
homology would not necessarily license this kind of inference.3
This characterization retains several advantages of homology judgments, such as the abil-
ity to talk about sameness in the midst of evolutionary modifications across taxa. One can still
accommodate variation in a mechanism’s composition (form) or what it is for (function). First,
although particular types of constituents are necessary for a conserved MGM, there can be vari-
ation in the number of components and intensity of activities for each type (Halina and Bechtel
2013). For example, the number of Wnt proteins involved in segment formation is variable
across protostomes even though the interactions between Wnt and Hedgehog pathways are
maintained (Janssen et al. 2010). Second, the requirement of specific organization need not
imply identical organization across taxa. One well-known example is the diverse rearrangement
of gene interactions in conserved mechanisms related to the initial patterning of insect embryos
(Chipman 2015). Third, spatiotemporal contexts and functional individuation can be treated at
different levels of abstraction to establish correspondence. One reproductive signaling pathway
is considered conserved even though it produces outcomes of larval formation in soil worms,
metamorphosis in insects, and puberty in mammals (Antebi 2013). At a higher level of abstrac-
tion, larval formation, metamorphosis, and puberty are all outcomes related to reproduction in
a life cycle. Similar reasoning permits identifying conserved mechanisms across spatiotemporal
contexts, such as the coordinated expression of Hox genes in establishing major body axes early
in development and individual appendage axes later in development.
With this picture of MGM conservation in view, we need to ask why many CPMs (e.g.,
branching morphogenesis) appear not to be conserved. The short answer is that they exhibit a
form of multiple realization (Bickle 2016). Constraints on what counts as a conserved MGM
prevent it from being defined solely in terms of an outcome. However, the same is not true for
many CPMs. There are numerous different ways to achieve epithelial invagination to gener-
ate branching morphologies. Another illustration comes from the developmental stage labeled
“gastrulation,” which refers to when a single-layered ball of animal cells is transformed into a
three-layered embryo of endoderm, mesoderm, and ectoderm. The MGMs involved are highly
conserved (Tam and Loebel 2007), but a plethora of CPMs manifest during gastrulation in
different taxa, including invagination, ingression, involution, delamination, epiboly, intercala-
tion, and convergent extension (Stern 2004). Similar studies on diverse anatomical features in
ontogeny corroborate this theme: “a conserved molecular mechanism can control convergent
morphogenesis through different cell behaviours” (Steventon et al. 2016, 1732).
This evolutionary decoupling of MGMs from CPMs has challenged the thinking of devel-
opmental biologists.
Despite the fact that many of the same signaling molecules are used during branching
morphogenesis of the Drosophila trachea and vertebrate lung, the physical mechanisms
that generate branches are distinct . . . how do homologous signaling pathways yield
such a diversity of physical mechanisms? . . . Is the molecular homology simply mis-
leading our search for blueprints that govern the development of branching epithelia?
These questions are not limited to branching systems. . . . Many of the transcription
factors involved in early cardiogenesis . . . are conserved across species, but the tissue
deformations that drive heart development can be very different.
(Varner and Nelson 2014, 2756)
In some ways this recapitulates what happened two decades earlier when developmental genetic
expression was overextended to make problematic claims of homology for many embryological
338
and anatomical features across metazoans (Abouheif et al. 1997). In that situation, the molecu-
lar homologies were misleading. Another question that naturally emerges about the relation
between MGMs and CPMs pertains to the explanatory expectations of developmental biol-
ogists. Do diverse CPMs bear commensurate explanatory weight with conserved MGMs?
Do they individually or in combination provide answers in the “search for blueprints that
govern development”?
4. Integrating genetic and physical developmental mechanisms?

Our discussion of how MGMs are evolutionarily decoupled from CPMs vindicates the cau-
tion in interpreting the former simply as component mechanisms of the latter. However, the
significance of this decoupling can be probed further. Building explanatory models of complex
phenomena (like developing embryos) is a central challenge in diverse sciences. The problem
intensifies when the aim is to offer an integrated account of how different types of causes or
mechanisms make contributions to complex biological phenomena (Mitchell 2003). How are
different types of causes “combined” to explain how an effect results from their joint opera-
tion? Is the integration a relationship of realization, where some mechanisms are parts of other
mechanisms, or a relationship of causation, where productive continuity is established between
diverse causal factors in a single mechanistic description?
This framing arises directly out of scientific practice. It is manifested in questions about how
to combine causes associated with CPMs and causes associated with MGMs that operate during
embryogenesis (Miller and Davidson 2013). Genetic mechanism explanations appeal to changes
in the expression of genes and interactions among their RNA and protein products to causally
explain how processes of differentiation, pattern formation, growth, and morphogenesis pro-
duce anatomical structures. Physical mechanism explanations appeal to mechanical forces (e.g.,
mechanical buckling) resulting from the geometrical arrangements of soft condensed materi-
als within the embryo to causally explain the same effects. The issue is not whether they both
operate—“both the physics and biochemical signaling pathways of the embryo contribute to
the form of the organism” (Von Dassow et al. 2010, 1)—but how to combine different types of
causes to understand their joint contribution to the effect of organismal form.
A degree of pessimism about achieving this type of integration among causal contributions
can be found in prior philosophical discussion. One analysis claims that apportioning causal
responsibility (i.e., how much a cause contributes to an effect) is impossible for developmen-
tal phenomena because there is no common currency for measuring the contributions (Sober
1988). Sober distinguishes questions of relative contribution from questions of whether a cause
makes a difference. Although there are sciences where questions of relative contribution can
be addressed (e.g., Newtonian mechanics), only difference-making questions can be asked in
developmental biology. Careful experimental methodology can answer whether one type of
difference-maker accounts for more of the variation in an effect variable for a particular popula-
tion (e.g., vary the genetic constitution while growing a plant in identical environments or vice
versa), but the measure of interaction between variables says only that there is a joint effect, not
how they jointly bring about their effect. Since the latter is a primary aim of mechanistic explana-
tion, this paints a bleak picture for building integrated explanatory models of the morphological
outcomes of embryogenesis that include both genetics (the presence, absence, or change in fre-
quency of RNA molecules or proteins) and physics (stretching, compression, fluid shear stress).
Representations of time are an underappreciated dimension of explanatory models.
Difference-making accounts of causal reasoning typically have no representation of time apart
from an ordinal relation of causes preceding their effects: “dependence concerns a relation
339
Alan C. Love
between cause and effect, without concern about what happens in between cause and effect”
(Illari and Russo 2014, 252). This is a common idealization in these accounts and facilitates dis-
secting causal relationships within complex phenomena. (Idealizations are reasoning strategies
that purposefully depart from features known to be present in nature to achieve prediction,
explanation, or control.) Production accounts of causal reasoning, including those associated
with mechanisms, do not share this idealization and explicitly attend to the representation of
time, especially to illustrate the productive continuity of causal factors in generating an out-
come: “production concerns the linking between cause and effect” (Illari and Russo 2014,
252; compare Craver and Darden 2013). One can distinguish the contributions of some
causes from others because they do not operate instantaneously. Combinations of causes can
be conceptualized not only as “this and that,” but also as “this then that.” However, even
though descriptions of mechanisms can incorporate temporal duration between difference-
makers or stages of a mechanism, there is a question about whether the forms of integration
common to mechanistic explanation are suited to combining genetic and physical causes (see
Chapters 10 and 11, this volume).
Consider an account of the origin of aortic arch asymmetry (Yashiro et al. 2007). Although
the morphological outcome results from both gene expression and physical dynamics, the
organization of these causal factors within a representation of developmental time permits
a dissection of how contributions are made. Gene expression at an earlier time makes a
difference in the structure of the outflow tract, which leads to a differential distribution of
blood flow on the left and right. The decreased blood supply on the right down-regulates
genes normally expressed on the left, which causes the branchial arch artery to disappear
from the right and leave only a left-sided aortic arch. While this type of explanatory model
does not permit a quantitative evaluation of the relative contribution of each type of cause,
it does illuminate how types of causes are compounded to bring about an effect. It also
aligns with the reasoning of developmental biologists: “an increasing number of examples
point to the existence of a reciprocal interplay between expression of some developmental
genes and the mechanical forces that are associated with morphogenetic movements or with
hydrodynamic flows during development” (Brouzés and Farge 2004, 372). This reciprocal
interplay reinforces the need to conceptualize MGMs and CPMs as causally interacting,
possibly within a single mechanism.
Mechanistic explanations involve decomposing systems into their constituent parts, local-
izing their characteristic activities, and articulating how they are organized to produce a
particular effect at a specific place and time (Darden 2006; see section 2 of this chapter). The
element of organization appears crucial to building integrated explanatory models of how an
effect is produced by genetic and physical mechanisms during development. A natural strategy
is to integrate these two types of mechanism according to tactics that have applied in other
contexts. Craver and Darden (2013, ch. 10) describe three ways that integration can occur in
mechanistic explanation. The first is “simple mechanistic integration” where different fields
are working at the same level or spatial scale of a mechanism, such as the role of RNAs in
protein synthesis. Molecular biologists, focusing on an earlier time in protein synthesis, elu-
cidated how messenger RNAs act as a template to guide the incorporation of amino acids
via transfer RNAs into a polypeptide that corresponds with the triplet codon derived from
a DNA sequence. Biochemists, focusing on a later time, elucidated how individual amino
acids attached to transfer RNAs bonded to one another to form a stable polypeptide chain.
Although different methods and experimental systems were required to ascertain how these
aspects of the mechanism operate, they eventually were integrated into a single mechanistic
description of continuous organization.
340
The second form of mechanistic integration is “interlevel integration”—exploring mecha-

nisms at different spatial scales or hierarchical levels understood as part–whole relations. For
example, to comprehend mechanisms of learning and memory, component mechanisms at the
level of the hippocampus need to be integrated with component mechanisms for how long-
term potentiation occurs in individual neurons. These, in turn, need to be integrated with
component mechanisms for how specific protein receptors are activated by signaling molecules.
Integrating these nested, component mechanisms yields a more robust explanatory model of the
mechanisms of learning and memory.
A third form of mechanistic integration is “sequential intertemporal integration,” such as
different research communities investigating distinct steps of a complex mechanism where
the steps are temporally disjointed, sometimes across generations. The mechanism of hered-
ity involves a complex series of events (e.g., gene replication and expression), each of which
contain different relevant component mechanisms (e.g., meiosis or transcription factor bind-
ing). Although interlevel integration helps to dissect various steps in this series, one also needs
intralevel integration to bridge temporal gaps and purchase productive continuity of organiza-
tion for the mechanistic description. Sequential intertemporal integration requires combining
the other two modes of integration across spans of time; otherwise, there are black boxes
between various steps in the series.
Although integrated explanatory models exhibiting these three types can be found in bio-
logical practice, it is unclear whether they are suitable for building integrated explanatory
models of genetic and physical causes in development. Simple mechanistic integration is inap-
propriate because the causal dynamics occurring through the temporal duration are interlevel
(e.g., gene expression and fluid flow). Interlevel integration is inadequate because the mode
of explanatory reasoning does not involve conceptualizing MGMs as hierarchically organ-
ized parts of CPMs. This is the justified caution identified earlier of not simply interpreting
MGMs as nested components of CPMs. By implication, sequential intertemporal integration,
which combines features of the other two types of integration across substantial temporal gaps,
is not applicable either. None of these appear suitable for combining genetics and physics in
explanations of development, even though the causal reasoning associated with mechanisms
foregrounds the temporal dimension that is crucial in overcoming the idealization about time
inherent to difference-making approaches.
Why are these approaches inadequate to the task of building integrated explanatory models
of genetic and physical factors? First, the reciprocal interaction between genetic and physi-
cal causes does not conform to the expectation that mechanism descriptions “bottom-out” in
lowest-level activities of molecular entities (Darden 2006). The interlevel nature of the causal
dynamics between genetic and physical factors runs counter to this expectation and is not ame-
nable to an interpretation in terms of nested mechanisms realizing another mechanism. Second,
the reciprocal interaction between genetic and physical causes does not require stable, com-
positional organization, which is a key criterion for mechanisms (Craver and Darden 2013).
The productive continuity of a sequence of genetic and physical difference-makers can be
maintained despite changes in the number and types of elements in a mechanism. Although
compositional differences can alter relationships of physical causation (fluid flow or tension),
these relationships do not require the specificity of genetic interaction predominant in most
mechanistic explanations from molecular biology. (The multiple realizability of CPM outcomes
is central to this conclusion.) Standard mechanistic strategies of representation and explanation
appear inadequate to capture these mechanisms.
An alternative form of integration that could illuminate how different genetic and physical
causes jointly operate involves departing from a standard assumption about time and mechanisms.
341
Alan C. Love
Temporal duration for a mechanisms approach is typically time “in” the mechanism. In contrast,
most developmental explanations use standardized representations of times—periodizations—
that are measured and calibrated apart from specific mechanisms. The most ubiquitous of these in
developmental biology are normal stages (Hopwood 2005; DiTeresi 2010). Stages facilitate the
study of diverse mechanisms, with different characteristic rates and durations, within a common
framework for a model organism or its parts. They also permit the study of conserved MGMs in
different species because the mechanism description is not anchored to a particular periodization.
An example of explanatory modeling with genetics and physics using a standardized periodization
illustrates this point.
Experiments have shown that gene expression is initiated by mechanical deformations
of tissue structure in Drosophila embryos (Farge 2003). As germ-band extension occurs dur-
ing early ontogeny,4 the associated physical motions of morphogenesis induce expression
of the gene twist in the anterior foregut and stomodeal primordium (precursor cells of the
mouth). Although the protein Dorsal regulates twist, three experiments showed that twist
is also controlled by mechanical movements: (a) a transient lateral deformation induces the
ectopic expression of twist in the dorsal region of the embryo independent of dorsal expression;
(b) in mutant Drosophila where mechanical movements are blocked, normal expression of
twist in foregut and stomodeal primordium fails to occur, and artificial compression can rescue
the mutant phenotype; and (c) the laser ablation of cells in the dorsal region, which reduces
mechanical compression, leads to decreased expression of twist in the stomodeal primordium. In
the absence of the protein Dorsal, these mechanical movements permit an accumulation of the
regulatory protein Armadillo, which subsequently translocates into the nucleus and—in associa-
tion with other proteins—activates twist.
In this account of how aspects of the gut develops, a standard periodization for Drosophila
frames the description of causal interactions between physical and genetic factors (see also
Desprat et al. 2008). The pertinent events (e.g., mesoderm invagination or germ-band exten-
sion) correspond to the formal structure of normal stages: Stage 6 (mesoderm invagination),
Stage 7 (early germ-band extension/endodermal anterior midgut invagination), and Stage 10
(stomodeal foregut invagination). These stages, which are calibrated independently of the
mechanisms of gut formation, permit the linkage of difference-makers into chains of produc-
tive continuity. The difference-makers are not ranked according to how much variation in
the effect variable they account for, but are combined in a periodization that orders different
kinds of causal factors that contribute to an effect. Researchers capture the dependency rela-
tions between genetic and physical factors by mapping them onto a temporal sequence during
which the mediation of causes and effect occurs. The same set of stages can accomplish a similar
integration with respect to different sets of causes pertinent to other developmental processes.
Although no standard composition of causes emerges from this strategy (i.e., quantified relative
contributions), we do achieve an understanding of how multiple causes combine mechanistically
to yield a morphological outcome.
Is this integrated explanatory model best understood as a single mechanism that incorporates
both physical and biochemical interactions? We have the linkage of difference-makers into
chains of productive continuity—a standard criterion for mechanistic descriptions (Craver and
Darden 2013). This interpretation is possible, though biologists don’t adopt it and continue
to distinguish molecular and physical mechanisms. The reason why they might not adopt the
perspective of a single mechanism returns us to a question of explanatory values: should “inte-
grated” mechanisms be valued over separate mechanisms that are more widely generalizable,
especially conserved MGMs? A consequence of using standardized periodizations to combine
genetic and physical causation in mechanistic explanations of embryogenesis is a reduction in
342
the scope of resulting generalizations. To achieve an integrated or more complete explanation

for two modes of causation, we must give up the explanatory generality secured for one mode
of causation treated separately. The periodizations used to assign responsibility to genetic and
physical difference-makers in chains of productive continuity are relatively specific to a model
organism. Additionally, “there are no one-to-one correspondence principles between gene
functions and the mechanical events that they affect” (Miller and Davidson 2013, 741); this is
the diversity of CPMs that can yield the same outcome in different contexts.
There is a tradeoff between the specificity of diverse causal interactions in an integrated
explanatory model and the generality found in non-integrated explanatory models, especially
those constructed from conserved MGMs. It is not a small thing to give up on the explanatory
generality of the latter since it is a large part of the rationale for the use of model organisms in
developmental biology. Since integrated explanatory models of genetic and physical mecha-
nisms remain rare, explanatory generalizations of wide scope based on conserved MGMs appear
more valued by developmental biologists than combined models of genetic and physical causes.
A preference for generality fits with the standard genetic orientation of developmental biol-
ogy as a discipline, but also has a wider application. Building integrated mechanistic models
is frequently the result of interdisciplinary efforts, where differences in explanatory standards
are exposed; generalization within a discipline is likely to take priority over integration across
disciplines. This implies that the tradeoff between explanatory generalization and explanatory
integration (or completeness) identified for genetics and physics in developmental biology is
broadly applicable across the sciences.
5. Further issues and provisional conclusions

The two issues analyzed in this chapter do not exhaust the space of conceptual questions
related to developmental mechanisms. To see this, let us briefly describe one more issue: the
dynamic constitution and organization of MGMs. This issue is observable in an overview of
the Developmental Mechanisms Section of the National Institute of Dental and Craniofacial
Research.5 The institute’s research focuses on “how signaling interactions control regulatory
state transitions during embryogenesis” and it “emphasizes that signals not only activate new
gene regulatory networks, but also eliminate or alter pre-existing ones.” MGMs, such as gene
regulatory networks, come into existence, transform, and go out of existence during the pro-
cesses of embryogenesis.
The contrast between the stability of the Hedgehog and Wnt signaling mechanisms and the
transient nature of the segment polarity network of gene expression illustrates this issue con-
cretely. Various components and activities are still present but their patterns of interaction within
a particular spatiotemporal context have changed. Additionally, MGMs can exhibit changes in
constitution. The Wnt signaling pathway has a “noncanonical” version that is involved in planar
cell polarity during ontogeny, the establishment of directionality in epithelial tissue (e.g., hairs
pointing in the same direction). Again, a Wnt protein jointly binds to a Frizzled transmembrane
protein and another receptor, which activates Dishevelled. However, β-catenin and its degrada-
tion complex are not involved. Instead, Dishevelled forms a complex with different elements to
activate proteins that regulate actin polymerization in the cytoskeleton to facilitate the establish-
ment of polarity (Komiya and Habas 2008).
A natural response to changes in constitution or organization is that these are no longer
the same mechanism. And, since the outcome of noncanonical Wnt signaling is different, this
change in the phenomenon being produced could indicate that these are different mechanisms.
Whether this is true remains an open question among scientific practitioners; some researchers
343
Alan C. Love
argue for a unified view of Wnt signaling as a mechanism that includes both canonical and non-
canonical dimensions (van Amerongen and Nusse 2009). However, the point is that changes
of constituents and organization for these mechanisms, and hence their causal capacities, differ
from what philosophers have analyzed in molecular biological mechanisms (but see McManus
2012 and Parkkinen 2014). Most of the biological mechanisms analyzed to date (e.g., DNA
replication) are stable and recurring throughout an organism’s life cycle. Questions about the
individuation of mechanisms that change through developmental time require further analysis.
We began our discussion with the distinction between MGMs and CPMs. Then our ecu-
menical characterization of mechanisms illuminated concrete examples of MGMs involved in
insect segment formation and CPMs of branching morphogenesis. This set the stage for explor-
ing two central issues about developmental mechanisms: the conservation of MGMs and the
explanatory integration of genetic and physical mechanisms to understand complex interactions
and morphological outcomes in contrast to the explanatory generality of conserved MGMs
alone. For the former, it was necessary to incorporate comparative reasoning from evolutionary
biology because “conservation” is a form of homology for mechanisms. However, it is a special
kind of homology and its characteristics help account for the role that model organisms play
in developmental biological research. At the same time, delineating this circumscribed homol-
ogy judgment for conserved MGMs facilitated a better understanding of why CPMs are not
conserved.
For questions surrounding integration and generality, we isolated the significant role that
representations of time play in providing a template for combining different kinds of causes
into chains of productive continuity. Although strategies for integrating mechanisms exist, none
were adequate for combining genetic and physical causes in accord with reasoning patterns in
developmental biology. Instead, standardized periodizations offered one possibility that could
serve this integrative role, but at a cost of generality. The local nature of these periodizations in
conjunction with the diversity of CPMs that are decoupled from conserved MGMs meant that
an integrated explanatory model sacrificed the generality secured by reasoning with conserved
MGMs alone. There is a tradeoff between the localized explanatory completeness achieved in
an integrated mechanistic model and the explanatory generality achieved in a mechanistic model
based on molecular genetics alone. The ubiquity of the latter and rarity of the former suggest
that generality is preferred over integration by most developmental biologists.
Two provisional conclusions can be drawn from the preceding analysis. First, scrutinizing
conserved MGMs showed the importance of incorporating reasoning strategies from evolution-
ary biology with those commonly used in molecular and developmental biology. To date, most
philosophers interested in mechanistic explanation have neglected these dimensions of reason-
ing and therefore questions surrounding the meaning of “conservation” critical for licensing
inferences from model organisms were insufficiently analyzed. Second, the tradeoff between
integration and generality, displayed when comparing mechanistic models that combine genetic
and physical causes with those that rely on MGMs alone, is a fault line in how biologists concep-
tualize adequate explanations. The preference for explanatory generality over integration, with
an attendant loss of completeness in our understanding of how an outcome is actually produced
mechanistically, reveals new terrain for philosophical analysis.
A better understanding of the relative significance of explanatory values such as generality,
specificity, and completeness is necessary. Although philosophers have highlighted the impor-
tance of abstraction and the role of mathematical modeling (Levy and Bechtel 2013; Brigandt
2013), as well as idealization (Love and Nathan 2015), less work has been done on how dif-
ferent explanatory values obtain across diverse biological disciplines. Although our analysis
of developmental biology brings some of these issues into view, more work is needed that
344
explores mechanisms and mechanistic explanation in particular areas of biological (and other
scientific) research with special attention to these differences in explanatory expectations. Just
as the distinction between MGMs and CPMs was crucial to identifying issues in developmental
biology, so also the differences in reasoning about mechanisms in other areas of science, rather
than commonalities, hold promise for a deeper comprehension of how mechanistic explana-
tion operates across the sciences, especially under conditions of interdisciplinarity.
Notes
1 I received helpful feedback on an earlier version of this chapter from Stuart Glennan, Phyllis Illari, Tom
Stewart, and Yoshinari Yoshida. The research and writing of this article were supported in part by a
grant from the John Templeton Foundation (Integrating Generic and Genetic Explanations of Biological
Phenomena; ID 46919).
2 This is similar to other ecumenical versions: “A mechanism for a phenomenon consists of entities
and activities organized in such a way that they are responsible for the phenomenon” (Illari and
Williamson 2012, 120).
3 A justification for the deep commonalities in constituents and organization of conserved mechanisms
derives from the concept of generative entrenchment, which refers to how much a complex system or
behavior depends on these features and thereby maintains them (see Wimsatt 2015).
4 The germ-band is a coordinated group of cells that develops into the segmented trunk of the embryo.
During embryogenesis it extends along the anterior-posterior axis and narrows along the dorsal-
ventral axis.
5 http://www.nidcr.nih.gov/Research/NIDCRLaboratories/DevelopmentalMechanisms.
References
Abouheif, E., M. Akam, W.J. Dickinson, P.W.H. Holland, A. Meyer, N.H. Patel, R.A. Raff, V.L. Roth,
and G.A. Wray. 1997. Homology and developmental genes. Trends in Genetics 13:432–433.
Affolter, M., and E. Caussinus. 2008. Tracheal branching morphogenesis in Drosophila: new insights into
cell behaviour and organ architecture. Development 135:2055–2064.
Ankeny, R., and S. Leonelli. 2011. What’s so special about model organisms? Studies in History and
Philosophy of Science 42:313–323.
Antebi, A. 2013. Steroid regulation of C. elegans diapause, developmental timing, and longevity. In
Current Topics in Developmental Biology, eds. A.E. Rougvie and M.B. O’Connor, 181–212. San Diego:
Academic Press.
Bickle, J. 2016. Multiple realizability. In The Stanford Encyclopedia of Philosophy, ed. E.N. Zalta, http://
plato.stanford.edu/archives/spr2016/entries/multiple-realizability/.
Bier, E., and W. McGinnis. 2003. Model organisms in the study of development and disease. In Molecular
Basis of Inborn Errors of Development, eds. C.J. Epstein, R.P. Erickson, and A. Wynshaw-Boris, 25–45.
New York: Oxford University Press.
Brigandt, I. 2013. Systems biology and the integration of mechanistic explanation and mathematical expla-
nation. Studies in the History and Philosophy of Biological and Biomedical Sciences 44:477–492.
Brouzés, E., and E. Farge. 2004. Interplay of mechanical deformation and patterned gene expression in
developing embryos. Current Opinion in Genetics & Development 14:367–374.
Chipman, A.D. 2015. Hexapoda: comparative aspects of early development. In Evolutionary
Developmental Biology of Invertebrates 5: Ecdysozoa III: Hexapoda, ed. A. Wanninger, 93–110. Vienna:
Springer.
Craver, C.F., and L. Darden. 2013. In Search of Mechanisms: Discoveries across the Life Sciences. Chicago and
London: University of Chicago Press.
Craver, C.F., and J. Tabery. 2016. Mechanisms in science. In The Stanford Encyclopedia of Philosophy, ed.
E.N. Zalta, http://plato.stanford.edu/archives/spr2016/entries/science-mechanisms/.
Damen, W.G.M. 2007. Evolutionary conservation and divergence of the segmentation process in arthro-
pods. Developmental Dynamics 236:1379–1391.
Darden, L. 2006. Reasoning in Biological Discoveries: Essays on Mechanisms, Interfield Relations, and Anomaly
Resolution. New York: Cambridge University Press.
345
Alan C. Love
Davidson, E.H., and I.S. Peter. 2015. Genomic Control Process: Development and Evolution. San Diego:
Academic Press.
Davidson, L.A., M.A.R. Koehl, R. Keller, and G.F. Oster. 1995. How do sea urchins invaginate? Using
biomechanics to distinguish between mechanisms of primary invagination. Development 121:2005–2018.
Davies, J.A. 2013. Mechanisms of Morphogenesis: The Creation of Biological Form. 2nd ed. Amsterdam:
Academic Press, Elsevier.
Desprat, N., W. Supatto, P.A. Pouille, E. Beaurepaire, and E. Farge. 2008. Tissue deformation modulates
Twist expression to determine anterior midgut differentiation in Drosophila embryos. Developmental Cell
15:470–477.
DiTeresi, C.A. 2010. Taming Variation: Typological Thinking and Scientific Practice in Developmental
Biology. PhD dissertation. University of Chicago, Chicago.
Farge, E. 2003. Mechanical induction of Twist in the Drosophila foregut/stomodeal primordium. Current
Biology 13:1365–1377.
Ghiselin, M.T. 2005. Homology as a relation of correspondence between parts of individuals. Theory in
Biosciences 124:91–103.
Haag, E.S. 2014. The same but different: worms reveal the pervasiveness of developmental system drift.
PLoS Genetics 10:e1004150.
Halina, M., and W. Bechtel. 2013. Mechanism, conserved. In Encyclopedia of Systems Biology, eds.
W. Dubitzky, O. Wolkenhauer, K.-H. Cho, and H. Yokota, 1201–1204. New York: Springer.
Hopwood, N. 2005. Visual standards and disciplinary change: normal plates, tables and stages in embryol-
ogy. History of Science 43:239–303.
Illari, P., and F. Russo. 2014. Causality: Philosophical Theory Meets Scientific Practice. New York: Oxford
University Press.
Illari, P., and J. Williamson. 2012. What is a mechanism? Thinking about mechanisms across the sciences.
European Journal of the Philosophy of Science 2:119–135.
Janssen, R., M. Le Gouar, M. Pechmann, F. Poulin, R. Bolognesi, E.E. Schwager, C. Hopfen,
J.K. Colbourne, G.E. Budd, S.J. Brown, N.-M. Prpic, C. Kosiol, M. Vervoort, W.G.M. Damen,
G. Balavoine, and A.P. McGregor. 2010. Conservation, loss, and redeployment of Wnt ligands in
protostomes: implications for understanding the evolution of segment formation. BMC Evolutionary
Biology 10:1–21.
Komiya, Y., and R. Habas. 2008. Wnt signal transduction pathways. Organogenesis 4:68–75.
Levy, A., and W. Bechtel. 2013. Abstraction and the organization of mechanisms. Philosophy of Science
80:241–261.
Love, A.C. 2007. Functional homology and homology of function: biological concepts and philosophical
consequences. Biology & Philosophy 22:691–708.
Love, A.C. 2014. The erotetic organization of developmental biology. In Towards a Theory of Development,
eds. A. Minelli, and T. Pradeu, 33–55. Oxford: Oxford University Press.
Love, A.C. 2015. Developmental biology. In The Stanford Encyclopedia of Philosophy, ed. E.N. Zalta, http://
plato.stanford.edu/archives/fall2015/entries/biology-developmental/.
Love, A.C., and M. Nathan. 2015. The idealization of causation in mechanistic explanation. Philosophy of
Science 82:761–774.
Love, A.C., and M. Travisano. 2013. Microbes modeling ontogeny. Biology & Philosophy 28:161–188.
Lum, L., and P.A. Beachy. 2004. The Hedgehog response network: sensors, switches, and routers. Science
304: 1755–1759.
McManus, F. 2012. Development and mechanistic explanation. Studies in the History and Philosophy of
Biological and Biomedical Sciences 43:532–541.
Miller, C.J., and L.A. Davidson. 2013. The interplay between cell signalling and mechanics in develop-
mental processes. Nature Reviews Genetics 14:733–744.
Mitchell, S.D. 2003. Biological Complexity and Integrative Pluralism. New York: Cambridge University Press.
Parkkinen, V-P. 2014. Developmental explanations. In New Directions in the Philosophy of Science: The
Philosophy of Science in a European Perspective, Vol. 5, eds. M.C. Galavotti, D. Dieks, W.J. Gonzalez,
S. Hartmann, T. Uebel, and M. Weber, 157–172. Berlin: Springer.
Sober, E. 1988. Apportioning causal responsibility. Journal of Philosophy 85:303–318.
Stern, C. 2004. Gastrulation. Cold Spring Harbor, NY: Cold Spring Harbor Press.
Steventon, B., F. Duarte, R. Lagadec, S. Mazan, J.-F. Nicolas, and E. Hirsinger. 2016. Species-specific
contribution of volumetric growth and tissue convergence to posterior body elongation in vertebrates.
Development 143:1732–1741.
346
Tam, P.P.L., and D.A.F. Loebel. 2007. Gene function in mouse embryogenesis: get set for gastrulation.
Nature Reviews Genetics 8:368–381.
van Amerongen, R., and R. Nusse. 2009. Towards an integrated view of Wnt signaling in development.
Varner, V.D., and C.M. Nelson. 2014. Cellular and physical mechanisms of branching morphogenesis.
Von Dassow, M., J. Strother, and L.A. Davidson. 2010. Surprisingly simple mechanical behavior of a
complex embryonic tissue. PLoS ONE 5:e15359.
Weisberg, M. 2013. Simulation and Similarity: Using Models to Understand the World. New York: Oxford
University Press.
Wimsatt, W.C. 2015. Entrenchment as a theoretical tool in evolutionary developmental biology.
In Conceptual Change in Biology: Scientific and Philosophical Perspectives on Evolution and Development, ed.
A.C. Love, 365–402. Dordrecht: Springer.
Wolpert, L., R. Beddington, J. Brockes, T. Jessell, P.A. Lawrence, and E.M. Meyerowitz. 1998. Principles
of Development. New York: Oxford University Press.
Wolpert, L., C. Tickle, T. Jessell, P.A. Lawrence, E.M. Meyerowitz, E. Robertson, and J. Smith. 2010.
Principles of Development. 4th ed. New York: Oxford University Press.
Yashiro, K., H. Shiratori, and H. Hamada. 2007. Haemodynamics determined by a genetic programme
govern asymmetric development of the aortic arch. Nature 450:285–288.
347
26
MECHANISMS IN ECOLOGY1
Viorel Pâslaru
1. Introduction
Genetics, cell biology, and molecular biology offered to philosophers of science sufficient
examples to challenge the logical empiricist view on scientific inquiry as a search for laws of
nature and scientific explanation as subsumption of phenomena under laws. Those areas of
biology indicated instead that scientific inquiry is a search for mechanisms and explanation is a
matter of describing them (see Chapters 1 and 23). The resulting new mechanistic philosophy
of science offered various accounts of mechanisms whose common assumptions are articulated
by this formulation of a minimal mechanism: “A mechanism for a phenomenon consists of
entities (or parts) whose activities and interactions are organized so as to be responsible for the
phenomenon” (see Chapter 1).
Mechanisms are real and local complex systems or processes that are responsible for the
phenomena scientists study. Mechanisms produce phenomena in virtue of their parts that
are organized and engage in activities or interact. For example, a heart is a mechanism for
pumping blood. It produces this phenomenon because of its parts: right and left ventricles,
right and left aortas, and the valves. They perform activities, such as the right ventricle con-
tracting and pushing the blood into the pulmonary trunk, while the valves open and close.
Parts and activities are organized. Blood is collected in the right atrium first, but not in the
left atrium, and the tricuspid valve opens toward the apex of the heart, but not toward its
top. The mechanistic approach is reductionist in that it explains properties and activities of
the whole in terms of properties and activities of its lower-level parts. That walls of the heart
contract is explained by referring to three layers of cardiac muscles and to the intercalated
disks that enable direct transmission of electrical impulses between cells, which also shows
that mechanisms are hierarchical. The mechanistic approach acknowledges the complexity
of real systems and examines mechanisms in their contexts. Accordingly, operation of the
heart mechanism is considered in the context of the thorax, its interaction with the dia-
phragm, the lungs, and the system of blood vessels.
Ecologists frequently describe mechanisms for purposes of explanation and prediction.
However, philosophers of science have not examined ecological mechanisms, even though
they might be different from the biological mechanisms that have so far received philosophical
348
Mechanisms in ecology
scrutiny. I use the case of the invasive shrub Lonicera maackii (Amur honeysuckle) to examine
ecological mechanisms at the level of individual organisms. I then explore the case of mecha-
nisms involving groups of organisms. But first, I offer a brief overview of the field of ecology.
2. What is ecology?
Ecology is not environmentalism and its focus is not environmental problems and the effects
of human impact on the environment. Ecology is a discipline of biology that studies the world
of plants and animals, including humans, and is widely understood as “the scientific study of
the interactions that define the distribution and abundance of organisms” (Krebs 2008, 5).
Distribution is the place in nature where organisms are found, while abundance is their number
in an area. Distribution and abundance are two facets of the same phenomenon and the factors
that affect the former could also influence the latter. The factors that determine the distribution
and abundance of organisms fall into two categories: biotic and abiotic. Biotic factors comprise
interactions between organisms, such as predation, or competition. Abiotic factors comprise
influences of the environment on organisms, such as temperature, availability of light or water,
its pH, or the nature of the ground.
Distribution and abundance are examined at various levels of increasing complexity and
integration, but of decreasing understanding: populations, communities, ecosystems, landscapes,
and biosphere. To explain phenomena of distribution and abundance at one level, ecologists
typically seek explanatory mechanisms at lower levels. For example, behavioral and physiologi-
cal mechanisms operating at the level of individual organisms are sought to explain changes in
a population. Because its area of investigation spans so many levels, ecology neighbors with
multiple sciences. It draws on physiological and behavioral studies of individual organisms, as
well as on meteorology, geology, and geochemistry for its explanations (Krebs 2008). Ernest
Haeckel introduced the term ecology in 1869, but ecological research and interest predate this
date, making it one of the oldest disciplines of biology.
3. An example of individual-level ecological mechanisms

Research on the competitive success of the invasive shrub Lonicera maackii (Amur honeysuckle)2
against native plants illustrates how ecologists examine and conceive of mechanisms at the indi-
vidual level and their effects at the macro scale.
Specimens of the shrub were introduced to the US in 1898 to the New York Botanical
Garden. From there, humans further dispersed shrub specimens to serve as ornaments or to
stabilize soil. Furthermore, birds, deer, and small mammals consume the seeds of the shrub and
drop them with feces at various locations, which helped the invasive plant establish throughout
vast areas of the Midwestern US.
L. maackii dominates native plants in competition for light and suppresses their growth
because its canopy (its uppermost leaves and branches) is dense, and because it is able to produce
numerous stem shoots and grow rapidly in habitats with different light regimens. When grown
in understory, i.e., beneath the canopy of other plants, L. maackii maximizes height gain and
allocates energy to continuous production of long shoots growing from the base; it consequently
has a lower canopy width, fewer shoot ends, and a smaller diameter of basal stems. Should more
light become available, which happens when a patch of forest is cleared, say, by a fallen tree, the
shrub stops producing basal long shoots and starts producing long shoots at higher levels in the
349
Viorel Pâslaru
canopy, develops wider canopies, and produces more leaves. Compared with indigenous shrub
species, L. maackii is able to equal or exceed the maximum stem growth, higher stomatal density,
and thickness of leaves, higher leaf number, area, and mass in low-light conditions, but exceeds
such a performance in high-light conditions that occur following disturbances in forests. Under
LARGE SCALE
NEW YORK
BOTANICAL GARDEN
Humans Birds, Deer, Humans

NORTHEAST
CHINA
MIDWEST USA
LOCAL SCALE
Causation H R1 (light)
by absence
of herbivores
and of better
L.m. N.p.
STAGE competitors
1
C2
ESTABLISHMENT R2 (water, nutrients)
AND
COMPETITION
Tolerance and resistance to herbivory
Above-ground competition via shoot and leaf production
Allelopathy via leaf and root exudates
Below-ground competition via L.m.’s shallow root system
H R1 (light)
STAGE
2 L.m.
COMPETITIVE R2 (water, nutrients)

EXCLUSION C2
OF NATIVE PLANTS
Causation by absence of herbivores and of

better competitors
A B Direct negative effect of A on B

L.m. Lonicera maackii C2 Better competitor
A Indirect negative effect of A on B
LEGEND: N.p. Native plant R1 Resource 1 light
A B Positive effect of A on B
H Herbivore R2 Resource 2 water & nutrients No direct or indirect effect A on B,
A B
since A is absent
Figure 26.1 Integrated representation of the mechanisms of competitive success of L. maackii. Drawing
by the author
350
both light conditions, L. maackii allocates more energy to branch and leaf mass growth and this
allows it to overgrow the slow-growing native shrubs. The shrub is fully able to regenerate after
the first clipping both in open-growth and in forested areas. It leafs two to three weeks before
the native plants and drops the leaves four to six weeks later than the native plants, and can be
observed as late as December. Its leaves are also freeze-resistant and are not damaged by spring
bouts of cold, in contrast to the native species. Because of all these features, L. maackii is able,
under various conditions and despite various environmental stresses, to get more light than the
native species and deprive them of this precious resource, suppressing their growth. Seedlings
establish throughout a variety of light conditions, and even those that were suppressed because
of clumped seed input can survive and gradually replace adult L. maackii and native shrubs.
In addition to outcompeting native plants for light, L. maackii dominates below-ground com-
petition for nutrients and water. Its roots cover a wider area close to the surface, which allows
it to deprive neighboring plants of water and nutrients, suppressing their growth. Additionally,
it suppresses the growth of native plants by means of chemicals in a process called allelopathy.
Leaves produce thirteen mildly acidic toxic compounds that inhibit or delay germination of
seeds, or alter size, survival, and architecture (number of branches and bolts) of native plants and
deter generalist herbivores from consuming its leaves. Roots produce similar chemicals, but they
have milder inhibiting effects on neighboring plants. Chemicals released into the ground from
decomposing leaves and by roots modify soil and leaf microbial communities, altering ecosystem
function and processes.
Ecologists treat above-ground competition, below-ground competition, and allelopathy
as separate mechanisms, yet none of them is solely responsible for the competitive success
of L. maackii. Since they act jointly, I represent the separate descriptions in an integrated
account (Figure 26.1), which will also facilitate the forthcoming philosophical examination
of these mechanisms.
4. Individual-level mechanisms and the minimal account

Individual-level ecological mechanisms satisfy most, but not all features of the minimal account of
mechanisms and their particularities contribute toward the general examination of mechanisms.
Several decades after the introduction of L. maackii to the US, it became clear that it was not
an innocent shrub that beautifies landscapes, but a shrub that efficiently suppresses and elimi-
nates native plants. This phenomenon of suppression and elimination of native flora required
explanation that ecologists formulated from the beginning, not in terms of laws, ecological or
otherwise, but in terms of mechanisms. And so they articulated the mechanism of competition
for light, of below-ground competition for water and nutrients, and the mechanism of allel-
opathy, which were meant to account entirely for the elimination of native flora or offer partial
explanations of that phenomenon. Accordingly, ecological mechanisms, just like in other areas
of biology, are mechanisms for a phenomenon and are formulated to explain it.
Ecologists often characterize individual-level mechanisms as consisting of individual organ-
isms with certain properties doing certain activities. In Gause’s (1934) account of the mechanism
of competition, the components are the individual cells of yeast differentiated by the property
tolerance to alcohol. Alcohol produced by S. kephir is more toxic to S. cerevisiae than vice versa,
the latter being eliminated. L. maackii is representative of cases in which the phenomenon can-
not be explained by reference to a property of the individual organism and its activity, but to
several properties and activities that engage different parts of the same organism. Accordingly, it
is parts of organisms that are entities in various mechanisms. The roots of L. maackii are compo-
nents of the mechanism for below-ground competition, while its shoots and leaves are parts of
351
Viorel Pâslaru
the mechanism for above-ground competition, not the individual plant, although it is necessary
for the roots, shoots, and leaves to function. The case of L. maackii helps stress that entities that
appear simple are in fact composites of many parts that function in different mechanisms.
The composite nature of the organism L. maackii makes the mechanism of competition
between this invasive plant and native species hierarchical. Its lower-level components are
the mechanism of competition for light, the below-ground mechanism of competition for
nutrients, the mechanism of allelopathy, and the mechanism of tolerance and resistance to her-
bivores. Each of these mechanisms is made up of parts at lower levels that are also investigated.
The mechanism of competition for light consists of leaves and branches, with specific prop-
erties and performing activities. Shrubs that grow in understory have long shoots of a small
diameter, fewer shoot ends, and a low canopy width. Since leaves are a key component of
this mechanism, their properties are closely studied. Ecologists examine their higher stomatal
density and thickness, their number per stem, area, mass, and their photosynthetic performance
in low-light conditions compared with native species. The mechanism of photosynthesis that
allows L. maackii to thrive in low-light conditions and the mechanism for longer leaf duration in
L. maackii are lower-level components of the mechanism of competition for light.
Hierarchical organization of ecological mechanisms shows some entities can be components
in several mechanisms. Leaves and roots, in addition to being parts of the mechanisms of above-
ground and respectively below-ground competition, are also components in the mechanism
of allelopathy. Leaves produce toxic compounds that, once they reach the ground with falling
leaves, inhibit the germination of seeds of native species, while roots produce metabolites that
inhibit the growth of roots of native species. The outcome is the same in both cases: native spe-
cies are prevented from growing. Descriptions of the mechanism of allelopathy include chemical
analyses of the allelochemicals that leaves and roots produce.
Furthermore, ecological mechanisms do not have clear boundaries, and what identifies them
is the operational unification of components of different sizes. For example, the boundaries of
the mechanism of competition for light are set by the operational unification of the parts of dif-
ferent sizes: shoots and leaves of individuals of species involved and light.
The explanatory role played by absent entities is another feature that, while not unique to
ecological mechanisms, is very important. According to the Enemy Release Hypothesis, one
of the hypotheses formulated to explain the success of L. maackii, the shrub spreads well in new
environments because its traditional predators are absent. Similarly, it is not subject to intense
competition, for there are no competitors that could challenge it. Ecologists consider the absent
predators and competitors in light of the assumption that all species are parts of food webs and
have predators and competitors that control their growth where the absence of either of them
is a cause of a species’ growth. Had either a predator or competitor been present, the dynamics
of the species would have been different. Ecologists treat absent predators and competitors as
possible causes that could influence the phenomenon under scrutiny, and as such they are part
of the mechanistic account of the phenomenon along with the detailed descriptions of actual
entities. This is not a departure from the realism of mechanistic philosophy, since the absent
predators and competitors are not abstractions or fictions, but real organisms that happen not
to be present in the habitat under scrutiny. The list of absent causes could also include other
instances of causation by absence involving abiotic or biotic factors. Absence of sunlight explains
the absence of shade-intolerant plants, and absence of water explains the presence of drought-
resistant plants and the absence of drought-intolerant ones. The absence of an entity is not
sufficient by itself to account for the phenomenon. The other components of the mechanism
are required. Mechanisms in ecology involve absences and preventions, and a mechanistic per-
spective must explain their causal relevance. (See also Chapter 11 and Glennan (forthcoming).)
352
While the mechanisms described above seem generally to fit the characterization of minimal
mechanisms, there are two features of ecological mechanisms that seem to distinguish them
from many of the mechanisms discussed by New Mechanists. The first is that ecological mecha-
nisms are often described in terms of properties rather than activities or interactions, and the
second is that ecological mechanisms are often less dependent on organization than many exem-
plars of biological mechanisms.
In the characterization of mechanisms by Machamer, Darden, and Craver (2000), entities
and activities take center stage, while properties seem subordinated and only briefly acknowl-
edged as bases of activities. In Glennan’s latest work causal powers or capacities perform the
metaphysical role that properties have relative to activities in the characterization of Machamer,
Darden, and Craver (Glennan forthcoming). The minimal account of mechanisms does not
either thoroughly examine the role of properties in mechanisms. However, if the philosophy
of mechanisms is to do justice to ecology, one has to emphasize the properties of entities.
Ecologists are explicit about the importance of properties in characterizing mechanisms:
Perhaps I should clarify what I mean by mechanism. The mechanistic basis for popula-
tion ecology is provided by the properties of entities one hierarchical level lower than
populations, that is, by the behavior and physiology of individual organisms.
(Turchin 1999, 156)
And it is this focus on traits that defines ecology’s subdisciplines of functional ecology and
physiological ecology (Shipley 2010). In ecological context, properties of entities are the
actual traits of organs, individuals, populations, and communities that ecologists measure,
rather than the not-yet-manifested capacities. A focus on properties is necessary to better
understand the activities of organisms. L. maackii chemically suppresses the growth of native
plants. To understand this activity, ecologists performed an analysis at the molecular level of
the chemicals and confirmed experimentally their inhibitory effect (Cipollini et al. 2008). In
other instances, description of properties amounts to a description of causes of an organism’s
impact, or of distribution and abundance. Variety of canopy shapes, i.e., leaf-forms, growth
forms, and heights of plants, accounts for better use of the three-dimensional space and,
consequently, better light interception (Tremmel and Bazzaz 1993). Tolerance to high soil
salinity and to deficiency of oxygen in water-logged soils is a cause of the presence of the
grass Juncus gerardi in those areas of marshes that are flooded regularly, while intolerance of
the shrub Iva frutescens to those conditions explains its presence on the terrestrial borders of
marshes that are not flooded (Bertness and Hacker 1994).
“Organization is the least controversial element in any characterization of mechanisms,” yet
how to understand organization is not a trivial matter and it has not been discussed extensively
(Illari and Williamson 2012, 127). The consensus in the literature is that along with entities
and activities, organization is a necessary condition for a mechanism to be able to produce
the phenomenon for which it is responsible. As a condition different from entities and activi-
ties, organization is generally treated as something that can be changed actually or in principle
by a human or natural agent: should the same entities and activities be organized differently,
they would produce a different phenomenon (Illari and Williamson 2012, 127). This view on
organization is a version of what Glennan (forthcoming; see also Chapter 7) calls induced organ-
ization, which he distinguishes from affinitive organization. This distinction between induced
and affinitive organization helps address the nature of organization in ecological mechanisms.
Induced organization is imposed upon entities and/or activities by an agent, human or oth-
erwise, determining the organization of the mechanism. Affinitive organization depends upon
353
Viorel Pâslaru
affinities, or properties of entities and/or activities. The latter are organized as they are because
of their properties. Two ways to organize a dinner party illustrate the two types of organization.
The host could induce the organization of the party and assign the seats of guests at the tables,
determining thereby who talks to whom. The second way is to let the organization of the party
to be determined by accident (where open seats are available) and by the affinities of people
attending the party—foodies, football fans, children, etc. (Glennan forthcoming).
Induced organization of mechanisms can be found in biogeography. A human or natural
agent can change the distance between an island and its source region or change the arrange-
ment of islands in an archipelago. Nature accomplishes this by means of tectonic changes,
while humans do so by modifying the landscape, as in the case of land-based islands. Changing
the distance between islands and their source region modifies the number of species present
on an island (Pâslaru 2014). One could further analyze induced organization and add that
it can be changed by either adding or removing components. That is, the host could invite
additional guests, or disinvite some if they fail to behave properly. Humans induced the
organization of entities—plant species—at the New York Botanical Garden when they intro-
duced a new component: L. maackii. Likewise, humans induce the organization of ecological
communities when they decide to remove L. maackii because it behaves as an invasive species
and suppresses native plants.
An illustration of affinitive organization is the mechanism of dispersal of seeds of L. maackii.
It involves the interaction between birds, deer, and small mammals with the shrub’s seeds and
it was not induced, but resulted from affinities of the entities and their activities. Birds consume
seeds because of their nutritious value and taste, and spread them to other places, where they
can germinate even after they have passed through the digestive tract of birds. No agent induced
the interaction between birds, seeds, and germination of the latter. Their properties deter-
mined them to interact and organization emerged as a result. The mechanism of competition
for light appears organized affinitively: L. maackii leafs out before native plants and suppresses
their growth by depriving them of light. Leafing at a certain time in the spring is a property
of L. maackii that it will show regardless of where and in what plant community it is planted.
Similarly, succession, another central ecological process, shows affinitive organization. An area
is colonized in a certain order by species belonging to certain functional types. Properties of
species determine the order of colonization. Soil-fixing plants establish first, and then trees,
herbivores, and carnivores. Even if an herbivore were introduced first, it could not get estab-
lished in the absence of plants suitable for consumption. These examples indicate that affinitive
organization is restrictive. It restricts an entity’s interactions. They also indicate that properties
of components limit induced organization and, more generally, “[a]ll mechanistic organization
depends to some degree upon the existence of affinities—as parts must have the capacities to
interact with other parts” (Glennan forthcoming). One cannot just organize differently the
same entities and activities to get a different phenomenon. Trading the location of L. maackii’s
canopy with that of its roots will produce no meaningful ecological phenomenon. The shrub
will simply die. Whether induced organizations persist also depends on the properties of entities.
L. maackii was able to establish in New York, as well as throughout the Midwestern US,
because of its properties and activities. Had it not been a good competitor for light, or a
producer of allelochemicals, it would have been limited to the grounds of botanical gardens.
As in the case of non-hierarchical organization, properties determine hierarchical organiza-
tion. Toxic compounds produced by leaves and roots of L. maackii cannot be a component of
the mechanism of competition for light and occupy the same level with leaves of native species,
because the compounds do not have the geometrical and material properties necessary to dimin-
ish the amount of light reaching native plants. And roots cannot be part of this mechanism, since
354
their properties are such that they can only subsist in the ground. Therefore, the hierarchical
organization is a product of the properties of entities and activities and one cannot induce a dif-
ferent hierarchical organization of a mechanism by reshuffling its components. Instead, one has
to introduce different entities which, however, could destroy the mechanism.
The stronger metaphysical finding of the foregoing examination is that induced organization
is on a par with properties in determining the functioning of mechanisms and it requires an
inducer, human or natural. Affinitive organization does not require an inducer and is a product
of entities interacting in a certain manner in virtue of their properties, and so it appears to be an
epiphenomenon. In mechanisms embodying affinitive organization, it is not the central element
that determines how entities and activities produce phenomena. Instead, it is the properties of
entities and activities that do this work. While affinitive organization is a form of organizing
ecological mechanisms, it does not have the causal power that properties of entities and activi-
ties have. This implies that to change the organization of a mechanism, one has to change the
entities and their activities. Only changing the spatial or temporal organization, while preserving
the components unaltered, is rarely possible.
5. Experimental methods for mechanism discovery

Experimental techniques in ecology exemplify a variety of experimental designs that have
been identified by New Mechanists as strategies of mechanism discovery. Some experi-
ments test the causal relevance of an entity, property, activity, or organizational characteristic.
Others examine the activity or the entity that mediates between a cause and its effects,
and are called by-what-activity and by-what-entity experiments, respectively (Craver and
Darden 2013; see also Chapter 19). Experiments examining multilevel mechanisms are
bottom-up or top-down experiments, using the analytic or the synthetic strategies, respectively
(Bechtel and Richardson 1993, 20). Alternatively, one intervenes on the start conditions to
produce the explanandum phenomenon and examines the behavior of the putative compo-
nents (Craver and Darden 2013, 128). Ecologists’ quest to identify mechanisms illustrates all of
the foregoing experiments.
Cipollini et al. (2008) did by-what-entity experiments when they identified the thirteen
phenolic compounds contained in the leaves of L. maackii, and then singled out the four
chemicals (including their formulas) that have an allelopathic effect. They did by-what-activity
experiments when they determined how the identified chemicals impact the growth of other
plants by inhibiting seed germination, and deter the foraging behavior of herbivorous insects.
Additionally, one could identify in ecology by-what-property experiments that detect the prop-
erty of an entity or activity that is solely responsible for or is necessary for the production of a
phenomenon. Such is the experiment of Luken and Mattimiro (1991) of clipping off repeat-
edly for four years all stems at the top of the base of forest- and open-grown L. maackii. The
experiment determined the shrub has the property of being resilient under stress, i.e. it is able
to resprout in different habitats.
Ecologists also perform experiments to test for causal relevance. Luken et al. (1997) created
large gaps in the shrub thicket to determine whether increasing light availability by removing
the shrub enhances establishment of understory plants. This is an experiment that tests the causal
relevance of shrub removal to enhancing abundance of understory plants. Similarly, Cipollini
et al. (2008) tested the causal relevance of leaf metabolites in suppressing the growth of other
plants. The experiment of Luken et al. (1997) illustrates bottom-up stimulation experiments.
They exposed L. maackii and L. benzoin to increasing levels of light (0 percent, 25 percent, 100
percent) to simulate disturbance regimes that increase light availability. Typically, stimulation
355
Viorel Pâslaru
experiments examine putative components. However, light was not a putative component, but
its level of causal efficacy, that Luken et al. examined, was. Bottom-up interference experiments
in ecology decrease to various levels an abiotic or biotic component. An entire population of
organisms could be removed, or parts of the organism, or the amount of an abiotic component
could be decreased. For example, L. maackii was removed in some experiments to examine
how its absence increases species richness (Gould and Gorchov 2000); in others only half of a
plants’ leaves were removed, and the strength of nitrogen fertilization was halved (Lieurance
and Cipollini 2013).
Top-down experiments are also common. Based on the hypothesis that leachate of L. maackii
affects other species, Watling et al. (2011) created artificial pools in which tadpoles were exposed
to water containing leachate and to non-contaminated water. This experimental strategy was
also used to investigate the mechanistic relationship between various levels of species diversity
and ecosystem processes. In light of hypotheses about the role of species diversity in ecosystem
processes, Naeem et al. (1994) created various communities in laboratory conditions, while
Tilman et al. (2006) assembled them in field experiments. Instead of examining the behavior
of specific components as the system is activated, ecologists focused on how species diversity
produces ecosystem processes and compared them with ecosystem processes that are produced
by an actual system of similarly varying species diversity.
It is not always possible to manipulate components of a mechanism, and manipulation
might not change the component in the same way as natural factors do. Additionally, manipu-
lation might not be sufficient to assess the magnitude of the indirect effects correlated with
u
* <.1
.1
COROLLA LENGTH APPROACHES .2
.3
.4
.5
* .6
.7
COROLLA WIDTH * .8
.9
NECTAR *
PRODUCTION
u
*
DRY MASS
PROBES/FLOWER u
u
* *
* *
HEIGHT PROPORTION
* FRUIT SET
OPEN FLOWERS * u
*
* * *
* u
TOTAL FLOWERS TOTAL
u FRUITS
* *
Figure 26.2 Final causal model of the causal relationships between reproductive success of flowering
plants given plant traits and pollinator behavior. From Mitchell (1994, 879). Reprinted by
permission of the University of Chicago Press
356
the component from its direct ones. In such cases, some ecologists use path analysis and
structural equation modeling (SEM) to identify causal relationships making up mechanisms
(Pâslaru 2015). Mitchell (1992, 1994) used these methods to identify the “underlying causal
mechanisms” responsible for the reproductive success of flowering plants given plant traits and
pollinator behavior. The first step in the use of SEM is to conjecture causal relationships among
variables based on available knowledge about the phenomenon in question and to formulate a
path model incorporating hypothesized causal relationships. The proposed model is then tested
for fit with available data. Should the model not fit the data, one can change some paths of
the model in light of a new hypothesis about the underlying causal mechanism and re-test the
model (Figure 26.2). If deemed necessary, one could study the lower-level mechanisms that
produce the causal relations between two variables.
6. Mechanisms above the individual level

Ecologists generally accept individual-level mechanisms, but some question the existence of
mechanisms involving populations. A frequent objection states that explaining population
dynamics in terms of abundances or density-dependence is a phenomenological approach,
and, therefore, unsatisfactory (Tilman 1987; Turchin 1999). According to this view, mecha-
nisms are found exclusively at the individual level; at the population level one finds only
summaries of interactions among individuals, which are necessary, since it is not possible to
follow the interactions among individual organisms (Turchin 1999). However, other ecolo-
gists affirm the existence of mechanisms at various levels beyond individuals, as I show below.
Given ecologists’ disagreement on this issue, further research is required. The following
examination serves this goal.
Ecologists describe mechanisms that involve groups of organisms of different species, or
of the same species, which are populations. Because of this, I prefer the term group-level
mechanisms, instead of population-level, to cover both cases. Groups have properties that are
different from those of individuals. The latter have properties that do not apply to groups:
they have skeletons, physiological systems, organs, and engage in specific behaviors, e.g.,
feeding or mating. Groups are characterized by frequency, density, growth rate, generation,
age structure, and diversity, properties that arise only when there is a collection of individu-
als. Group-level mechanisms contain causal relations that are best understood by analogy to
natural selection being a population-level causal process (Millstein 2006) and according to the
difference-making account of causation. For example, “If one systematically manipulates the
density of a population to change the annual fecundity, then variation in population density
is the cause of variation in annual fecundity.” Groups can be viewed as individuals, based on
an argument by Millstein (2009), and so they are the components of such mechanisms. They
are organized spatially and temporally and changes in the properties of one group bring about
changes in the properties of another group.
An example of group-level mechanisms is what Palmer et al. (2000) call the hypothetical
mechanism by which diverse terrestrial and aquatic plant communities may increase biodi-
versity in aquatic sediments (see Figure 26.3). Its components are not individual organisms,
but groups of organisms of various species. For example, component microbial diversity
groups together populations of microbes of different species. Diversity means here the num-
ber of species to which microbes belong and the functional identity of those species. In
addition to groups of organisms as components, this mechanism features a property as a
component—temporal stability of detrital pool. Components of the mechanism are linked
causally in the sense of the manipulability perspective on causation. Changing the number of
357
Viorel Pâslaru
Aboveground
Diverse Riparian Diverse Aquatic

Plant Assemblages Plant Assemblages
Aquatic Sediments
↑ Diversity of ↑ Diversity of
DOC Resources Detrital Resources
↑ Temporal Stability
Microbial Diversity of Detrital Pool
Diversity or Abundance
of Sediment Biota
Figure 26.3 Graph of a group-level mechanism. From Palmer et al. (2000, 1069). Reprinted by
permission of Oxford University Press
species, say, by decreasing it, or eliminating certain species of functional importance, changes
the number and functional identity of species present in the sediment. And the presence of
a component is responsible for the presence of another one, for, e.g., diverse riparian assem-
blages are partly responsible for the presence of dissolved organic carbon resources. While
components appear organized in how they are causally interrelated, they do not perform any
activity. This example and ecologists’ use of the term mechanism suggest viewing group-level
mechanisms as causal networks.
Description of this group-level mechanism is not sufficient to provide a satisfactory expla-
nation of the phenomenon biodiversity in aquatic sediments, because Palmer et al. require a
description of the lower-level mechanisms underlying the group-level mechanism. Even if the
group-level mechanism is explanatorily insufficient, it appears to be necessary to provide the
context for individual-level mechanisms producing the upper-level causal relations.
Another example of purported group-level mechanism is the increased population size
mechanism that was proposed to explain the macro-scale relationship between the amount
of energy that an assemblage of species receives and the number of species it contains. Adding
energy increases population size that in turn reduces its extinction risk, and since populations
are of different species, this increases the number of species of the assemblage (Evans, Warren,
and Gaston 2005). This seems to be a group-level mechanism in which components, energy,
and populations are linked causally such that they produce the phenomenon—the number of
species in an assemblage. However, I propose viewing this as a phenomenological description
of changes at the group level. The mechanism is at the individual level, as follows. Increased
amounts of energy delivered in the form of nutrients allow individual organisms to grow and
reproduce, which amounts to increases in population size. Because organisms are of different
species, increasing the amount of energy increases the population size of all species, which
results in the community containing a greater number of species compared with the situation
when less energy is available. This interpretation of the increased population size mechanism
358
indicates that at least some purported group-level mechanisms are reducible to individual-level
mechanisms. Ecologists’ reason for speaking about mechanisms is methodological: “[i]t is nei-
ther profitable nor necessary . . . to try to explain all macroscopic phenomena in terms of the
mechanisms at lower levels” (Brown 1995, 155). (This case is analogous to the statistical descrip-
tion of evolutionary theory that sees genuine forces of evolution as taking place at the level of
individuals. See Chapter 22.)
Elements of group-level mechanisms irreducible to individual-level mechanism appear in
the mechanisms responsible for population regulation. This phenomenon occurs in populations
fluctuating around a mean. Regulation happens when the per capita growth rate of the popu-
lation is influenced by density-dependence. If the density of a population in an area is high,
population growth is restricted, but it is amplified if the density is low (Rockwood 2015, 70–1).
One of the proposed mechanisms to account for population regulation is crowding. In it, the
frequency or intensity of interactions among individuals or with predators or parasites increases
as population density increases, resulting in a lower fecundity for the population (Rodenhouse
et al. 2003). As population density decreases, the frequency of interactions decreases, resulting
in higher fecundity followed by population increase. And so the population fluctuates around a
mean. However, interactions among individuals, which constitute the individual-level mech-
anism of crowding, are insufficient to explain population regulation, because they occur at
various densities. To account for population regulation, they have to be related to changes in
population density, yet it is a property of a group-level component. This indicates that contrary
to what ecologists claim, crowding is not a strictly individual-level mechanism but one that
involves both individual- and group-level components.
The foregoing examination points toward further work needed to clarify the nature
of mechanisms above the individual level. Tentatively, there are at least three types of
such mechanisms: (i) group-level mechanisms as causal networks that require individual-
level mechanisms for a satisfactory explanation; (ii) group-level mechanistic descriptions of
phenomena produced by individual-level mechanisms; and (iii) group-level mechanisms
containing individual-level components.
7. Conclusion
Describing mechanisms is a central way of delivering explanations in ecology, especially at the
individual level, where they could be characterized as follows:
An individual-level mechanism for a phenomenon in ecology consists of entities,

biotic or abiotic, that perform specific activities and are organized in certain ways by
virtue of their properties such that they are responsible for the phenomenon.
These mechanisms show the importance of properties in determining activities and organiza-
tion. The category of absent entities could be extended to comprise not just absent components,
but also absent properties, activities, and organization. Such an extended category of absent
entities allows for a more general account of background conditions necessary for the operation
of certain mechanisms.
While the nature of group-level mechanisms is less clear, and additional research is needed, it
can be concluded based on ecologists’ use of the term mechanism that some of them are causal
networks of group-level components, properties, and activities. Others offer useful methodo-
logical approaches despite being reducible to individual-level mechanisms, and yet others are
group-level mechanisms.
359
Viorel Pâslaru
Since ecology studies phenomena on various levels of integration above individuals and
groups, further research should address the nature of mechanisms ecologists describe at those
levels and their relationship with mechanisms at other levels, and answer this question: are there
mechanisms specific for each level, or are they just variations of the individual- and group-level
mechanisms examined above? I conjecture the latter to be the case.
Notes
1 I am grateful to Stuart Glennan and Phyllis Illari for helpful suggestions and comments on earlier drafts of
the chapter. I am also thankful to Dragoș Popa-Miu for helping me improve Figure 26.1 and to Marilyn
Marx for refining the English expression of this text. Hanley Sustainability Institute at the University of
Dayton financially supported part of the research for this project.
2 Summary of the research is based on the review by McNeish and McEwan (2016).
References
Bechtel, William, and Robert C. Richardson. 1993. Discovering Complexity: Decomposition and Localization
as Strategies in Scientific Research. Princeton, NJ: Princeton University Press.
Bertness, Mark D., and Sally D. Hacker. 1994. “Physical stress and positive associations among marsh
plants.” The American Naturalist 144 (3):363–372.
Brown, James H. 1995. Macroecology. Chicago: University of Chicago Press.
Cipollini, Don, Randall Stevenson, Stephanie Enright, Alieta Eyles, and Pierluigi Bonello. 2008.
“Phenolic metabolites in leaves of the invasive shrub, Lonicera maackii, and their potential phytotoxic
and anti-herbivore effects.” Journal of Chemical Ecology 34(2):144–152.
Craver, Carl F., and Lindley Darden. 2013. In Search of Mechanisms: Discoveries across the Life Sciences.
Chicago: University of Chicago Press.
Evans, Karl L., Philip H. Warren, and Kevin J. Gaston. 2005. “Species–energy relationships at the
macroecological scale: a review of the mechanisms.” Biological Reviews of the Cambridge Philosophical
Society 80:1–25.
Gause, G. F. 1934. The Struggle for Existence. Baltimore: The Williams & Wilkins Company.
Glennan, Stuart. forthcoming. The New Mechanical Philosophy: Oxford: Oxford University Press.
Gould, Andrew M. A., and David L. Gorchov. 2000. “Effects of the exotic invasive shrub Lonicera maackii
on the survival and fecundity of three species of native annuals.” The American Midland Naturalist
144(1):36–50.
Illari, Phyllis McKay, and Jon Williamson. 2012. “What is a mechanism? Thinking about mechanisms
across the sciences.” European Journal for Philosophy of Science 2(1):119–135.
Krebs, Charles J. 2008. Ecology: The Experimental Analysis of Distribution and Abundance. 6 ed. San Francisco:
Benjamin Cummings.
Lieurance, Deah, and Don Cipollini. 2013. “Environmental influences on growth and defence responses
of the invasive shrub, Lonicera maackii, to simulated and real herbivory in the juvenile stage.” Annals
of Botany 112(4):741–749.
Luken, J. O., L. M. Kuddes, T. C. Tholemeier, and D. M. Haller. 1997. “Comparative responses of
Lonicera maackii (Amur Honeysuckle) and Lindera benzoin (Spicebush) to increased light.” American
Midland Naturalist 138(2):331–343.
Luken, James O., and Daniel T. Mattimiro. 1991. “Habitat-specific resilience of the invasive shrub Amur
honeysuckle (Lonicera maackii) during repeated clipping.” Ecological Applications 1(1):104–109.
Machamer, Peter, Lindley Darden, and Carl F. Craver. 2000. “Thinking about mechanisms.” Philosophy
of Science 67(1):1–25.
McNeish, Rachael E., and Ryan W. McEwan. 2016. “Invasion ecology of Amur honeysuckle (Lonicera
maackii), a case study of impacts at multiple ecological scales.” The Journal of the Torrey Botanical Society
143(4):367–385.
Millstein, Roberta L. 2006. “Natural selection as a population-level causal process.” The British Journal for
the Philosophy of Science 57(4):627–653.
Millstein, Roberta L. 2009. “Populations as individuals.” Biological Theory 4(3):267–273.
360
Mitchell, Randall J. 1992. “Testing evolutionary and ecological hypotheses using path analysis and
structural equation modelling.” Functional Ecology 6(2):123–129.
Mitchell, Randall J. 1994. “Effects of floral traits, pollinator visitation, and plant size on Ipomopsis aggregata
fruit production.” The American Naturalist 143(5):870–889.
Naeem, Shahid, Lindsey J. Thompson, Sharon P. Lawler, John H. Lawton, and Richard M. Woodfin.
1994. “Declining biodiversity can alter the performance of ecosystems.” Nature 368:734–737.
Palmer, Margaret A., Alan P. Covich, Sam Lake, Peter Biro, Jacqui J. Brooks, Jonathan Cole, Cliff Dahm,
Janine Gibert, Willem Goedkoop, and Koen Martens. 2000. “Linkages between aquatic sediment
biota and life above sediments as potential drivers of biodiversity and ecological processes.” BioScience
50(12):1062–1075.
Pâslaru, Viorel. 2014. “The mechanistic approach of The Theory of Island Biogeography and its current rel-
evance.” Studies in History and Philosophy of Science Part C: Studies in History and Philosophy of Biological
and Biomedical Sciences 45:22–33.
Pâslaru, Viorel. 2015. “Causal and mechanistic explanations, and a lesson from ecology.” In Romanian
Studies in Philosophy of Science, edited by Ilie Pȃrvu, Gabriel Sandu, and Iulian D. Toader, 269–289.
Cham, Switzerland: Springer.
Rockwood, Larry L. 2015. Introduction to Population Ecology. 2nd ed. Chichester, West Sussex: John
Wiley & Sons.
Rodenhouse, Nicholas L., T. Scott Sillett, Patrick J. Doran, and Richard T. Holmes. 2003. “Multiple
density–dependence mechanisms regulate a migratory bird population during the breeding season.”
Proceedings of the Royal Society of London B: Biological Sciences 270(1529):2105–2110.
Shipley, Bill. 2010. From Plant Traits to Vegetation Structure: Chance and Selection in the Assembly of Ecological
Communities. Cambridge: Cambridge University Press.
Tilman, David. 1987. “The importance of the mechanisms of interspecific competition.” The American
Naturalist 129(5):769–774.
Tilman, David, Peter B. Reich, and Johannes M. H. Knops. 2006. “Biodiversity and ecosystem stability in
a decade-long grassland experiment.” Nature 441(7093):629–632.
Tremmel, D. C., and F. A. Bazzaz. 1993. “How neighbor canopy architecture affects target plant perfor-
mance.” Ecology 74(7):2114–2124.
Turchin, Peter. 1999. “Population regulation: a synthetic view.” Oikos 84(1):153–159.
Watling, James I., Caleb R. Hickman, and John L. Orrock. 2011. “Predators and invasive plants affect
performance of amphibian larvae.” Oikos 120(5):735–739.
361
27
SYSTEMS BIOLOGY AND
Ingo Brigandt, Sara Green, and Maureen A. O’Malley
Systems biology is a new and highly interdisciplinary field that combines elements from
molecular biology and physiology with quantitative modeling approaches from disciplines such
as engineering, physics, computer science, and mathematics. The term “systems biology” was
used originally in 1968 by Mesarović to urge the use of systems theory to understand bio-
logical systems (Mesarović 1968); some commentators would trace the historical roots even
further back (Green and Wolkenhauer 2013). But when the term is used in the context of
contemporary bioscience it typically refers to a much more recent approach, initiated in the
late 1990s as a response to the new experimental techniques and fast computers that allowed
the rapid sequencing of DNA and automated measurements of molecular interactions (Ideker
et al. 2001; Kitano 2001). These innovations afforded major new initiatives in the life sciences
but also produced unforeseen challenges. Systems biology addresses one of these, namely the
interpretation of extensive quantitative data via mathematical and computational modeling
(Alberghina and Westerhoff 2005; Boogerd et al. 2007).
Research in systems biology is driven by complex problems that require multidiscipli-
nary integration (Carusi 2014; MacLeod and Nersessian 2014; O’Malley and Soyer 2012).
Consequently, it is a diverse field. Some proponents pursue strategies that extend molecular
biology with sequence-based tools (see Chapter 23), while others explore the relevance of
abstract mathematical systems theory to molecular interactions (O’Malley and Dupré 2005).
Common to all branches of systems biology is the willingness to borrow reasoning and rep-
resentation tools from engineering and the physical sciences, including network diagrams and
graph-theory, other types of mathematical modeling (primarily ordinary differential equations),
and computational simulations. We focus on just one of the many possible questions about
systems biology: To what extent can the modeling strategies and explanations in systems biology be
characterized as mechanistic?
1. Dynamic mechanistic explanation and other modeling aims

A hallmark of mechanistic research is to understand a complex whole by decomposing it
into component parts, and by localizing phenomena of interest to certain parts of the system
(Bechtel and Richardson 1993; Craver 2007; see Chapters 9 and 19). Models in systems biol-
ogy are similarly based on empirically measured molecular entities and interactions. Given the
362
Systems biology and mechanistic explanation
abundance of different molecules and pathways in every cell, modeling involves the selection
of components relevant to the system being investigated (Donaghy 2014). But the role of
mathematical models and computational tools—as distinctive aspects of systems biology—was
not addressed in original philosophical accounts of mechanistic explanations (see Chapter 16),
primarily because molecular systems biology is so new. Lately, the relationship between models
in systems biology and mechanistic accounts has become an important philosophical topic of
debate, with some commentators arguing that a traditional mechanistic account is sufficient to
describe research in systems biology (e.g., Richardson and Stephan 2007), and others instead
stressing the need for a more pluralistic perspective of explanatory integration (e.g., Braillard
2010; Fagan 2016; Mekios 2015).
Although it is possible to focus on differences between dynamic models in systems biol-
ogy and mechanistic explanations in general (Issad and Malaterre 2015; Théry 2015), Bechtel
and Abrahamsen (2010) instead highlight the continuity between the two by introducing the
notion of dynamic mechanistic explanation. Dynamic mechanistic explanations are also based
on concrete entities and interactions, but they extend mechanistic explanation by math-
ematically or computationally capturing the dynamical operation of the system and its parts
across time. In fact, in the case of complex systems, mathematical models are required for the
purpose of mechanistic explanation (Baetu 2015; Bechtel 2012; Brigandt 2013; see Chapter 20).
How is this the case?
In addition to decomposition and localization as strategies for discovering mechanism com-
ponents, mechanistic explanations must reassemble those components and specify epistemically
how their organization and operation result in the overall features of the mechanism to be
explained (Bechtel and Abrahamsen 2005). Bechtel illustrates the importance of mathematical
models with circadian rhythms, which are endogenous oscillations of about 24 hours present in
most organisms. A mechanism diagram can depict various components of the underlying mech-
anism, including specific genes and proteins, some of which have oscillating expression levels.
The diagram can also represent the activation or inhibition of interactions among proteins and
other entities, thereby depicting positive as well as negative feedback loops (see Chapter 18).
For some simple mechanisms, mental simulation (using a mechanism diagram) suffices to show
how the phenomenon to be explained is generated (Bechtel and Abrahamsen 2005). However,
in the case of circadian rhythms, only mathematical models are able to reveal that over time
the component interactions—which involve changing protein concentrations, several feedback
loops, and time-delaying gene expression pathways—actually produce sustained periodic oscil-
lations (see also Brigandt 2015).
Computational modeling strategies in systems biology can thus extend mechanistic
accounts in various ways. By providing mathematical tools for modeling the dynamics of
large systems of nonlinear organization, computational models can help researchers recompose
knowledge about subsystems that have been taken apart for functional analysis. It is well
known that there is cross-talk between different mechanisms, and computational tools can
afford a better understanding of how mechanisms relate to one another dynamically (Bechtel
2016; Fagan 2012). Among the many examples are computer simulations of whole cells
(and even organs) that explore dynamic interactions between different functional subsystems
(Bassingthwaighte et al. 2009; Karr et al. 2012). If a mechanistic explanation is characterized
as accounting for how a system behavior is causally generated by the organized interaction of
its particular parts, computational models of circadian rhythms or large-scale simulations can
indeed be interpreted as vehicles for mechanistic explanations.
However, we should be careful about assuming that modeling in systems biology is always
geared toward mechanistic explanations. First, ethnographic studies of research practices in
363
Brigandt, Green, and O’Malley
systems biology show that scientists may not even frame their modeling aims in terms of
explanation, but instead in terms of control or prediction (MacLeod and Nersessian 2015).
Practical concerns (application) and pragmatic constraints (not all parameters can be measured
and modeled mathematically) direct particular modeling aims. MacLeod and Nersessian make
this point with the example of how reducing the amount of cell-wall hardening lignin in
plants is highly desirable for biofuel production. But because the relevant model of the lignin
synthesis pathway depends on estimated parameters, it indicates primarily a robust relation
between particular system components and the lignin output. The model’s main achievement,
therefore, is to reveal an angle of technological control. Although it does give partial insight
into how the system works, this model might not yield a mechanistic explanation in the sense
of accounting for the behavior of the whole in terms of its parts and their properties (MacLeod
and Nersessian 2015). Second, even when systems biologists have explanation as an explicit
target, they may not be offering causal explanations of specific systems. Instead, they may intend
to provide more abstract functional classifications of all the variants of system organization that
could possibly realize a particular function (see section 3).
The lesson we draw from these observations is that while systems biology can fruitfully
extend philosophical accounts of mechanistic explanation to include dynamical and quanti-
tative aspects by means of mathematical models and simulations, philosophers investigating
systems biology also need to pay attention to the diverse practices across this field, and to the
actual, context-dependent aims of the modelers and experimental researchers. We will take
this finding into account as we explore mechanisms in systems biology further via the use of
network analysis.
2. Network models: from motifs to global topologies

Because cellular systems are highly complex webs of molecular interactions, one approach
in systems biology involves the investigation of networks. A network can be represented and
studied computationally as a graph, in which the nodes correspond to molecular entities,
while an edge between two nodes represents an interaction between them. A graph can be
undirected, such as a protein-interaction network that depicts all the interactions in which the
protein types inside a cell engage, or it can be directed. The latter category includes metabolic
reaction networks, signal transduction networks, and gene regulatory networks that depict
genes regulating other genes.
While decomposition and localization have proven to be useful strategies of mechanistic
research, additional methods (e.g., graph-theoretic and computational) are needed in systems
biology to process large data sets and analyze highly integrated systems. One particular approach
is to screen larger networks for the repeated occurrence of the same type of small connectivity
patterns called network motifs (Figure 27.1; Alon 2007). The functionality of any network motif
can then be investigated computationally. Consider, for example, a feedforward loop, in which
X has a direct as well as a mediated input on Z (Figure 27.1A). Using engineering language,
systems biologists might say that Z processes its two potential inputs as an AND-gate, which
is when both inputs are needed for activation. In this case, the motif will function as a persis-
tence detector. In other words, output Z will be activated only upon sustained activation of X,
which can be turned on by some external signal. The reason is that when receiving an input
by means of the time-delayed pathway via Y (involving two activation steps), Z would not
receive a second input (directly from X) unless X has already been active for some time. Such a
persistence-detector design makes biological sense when it is energetically costly for an organ-
ism to synthesize an enzyme that processes a particular substrate. In that case, synthesis is best
364
A B C
X X X1 X2 X3 ... Xn
Z1 Z2 . . . Zn Z1 Z2 Z3 Z4 . . . Zm
Z
Figure 27.1 Examples of network motifs. A: Feedforward loop; B: Single-input motif; C: Dense
overlapping regulons. Adapted with permission from Alon (2007), copyright Chapman &
Hall/CRC Press
initiated only if the substrate is reliably present. Particular design motifs are expected to function
in the same general way, whatever the particular biological and environmental contexts of their
implementation (but see section 4).
Network motifs abstract away from a good deal of molecular detail. They neither specify
what kinds of entities the nodes are, nor do they indicate the actual means by which one
entity would activate another (e.g., that a eukaryotic gene is transcribed to RNA, which when
transported outside the nucleus is translated to a protein, which later binds to a different gene
so as to activate it). Despite this loss of mechanistic detail, Levy and Bechtel (2013) argue that
the analysis of a network motif is still a dynamic mechanistic explanation. This is because once
abstracted, the network account points directly to the organization of the mechanism that is
responsible for the phenomenon to be explained. Generally, this sort of abstraction occurs
widely in systems-biological modeling, including the examples already mentioned in section 1
(see also Chapters 17 and 34).
Even though the analysis of an individual network motif’s functionality might be largely
mechanistic, what makes research on network motifs distinctively systems-biological stems from
the fact that large networks are screened to determine the frequency with which motifs occur
(e.g., the feedforward loop is known to be highly abundant). Doing so reveals both common
and uncommon elements of biological design, and draws attention to the former, which are
likely to be more biologically important. Moreover, different kinds of large networks, from
gene regulatory to neural networks, can be screened for the same design element. This general-
izability also applies to networks from different taxa, whether prokaryotes or eukaryotes. These
strategies indicate that abstract organizational schemes, which systems biologists call design princi-
ples, transcend the organization of a single mechanism, and even a single species (Green 2015b).
We will elaborate on design principles in section 3.
While the scrutiny of an individual motif pertains to a very small network, research at the other
end of the spectrum investigates large networks for their global properties, also via graph-theoretical
means. Earlier work initially addressed regular networks (where each node has the same number
of edges) and random networks (where a certain proportion of nodes is randomly connected by
edges; Figure 27.2A). In the last fifteen years, however, small-world and scale-free networks have
gained prominence because of their interesting properties and widespread occurrence among real
biological systems. A small-world network is defined in terms of the global property of the average
path length between two nodes—averaged across all pairs of the network’s nodes—which grows
logarithmically as the number of nodes increases. This means that for two randomly chosen
nodes, the shortest distance between them (in terms of a path of intermediate nodes connected
by edges) will be small relative to the size of the network. This global property usually entails that
365
a signal propagates quickly from one part of the network to another. For a biological system, this
can have the advantage of enabling rapid reaction times. Many protein-interaction networks are
small-world for this reason (Albert 2005).
A network’s degree distribution P(k) is the network-wide proportion of edges that is con-
nected to k other nodes (i.e., the network’s proportion of nodes connected to only one other
node, the proportion of nodes connected to two other nodes, and so on). The degree distribu-
tion is thus a global characteristic. A scale-free network is defined as a network that has a degree
distribution that follows a power law of the form P (k ) = c ⋅ k −γ. This exponentially declining
function means that across any scale-free network there are many nodes that are connected to
only one or a few other nodes, while only few nodes are so-called hubs, which are connected to
many other nodes (Figure 27.2B). From this global property, predictions can be made about the
network’s functionality. One is that it will exhibit robustness, which is the biologically important
feature that a system will maintain its functionality despite perturbations. While the elimination
of a node that is connected to one or only a few other nodes is unlikely to affect a network’s
functionality, eliminating a node that is a hub may seriously impact how the network functions.
But in a scale-free network there are comparatively few hubs, meaning that such a network
is generally robust. A variety of actual biological networks of interest to systems biology are
approximately scale-free, including metabolic networks and the gene regulatory networks of
prokaryotes and eukaryotes (Albert 2005).
Huneman (2010) coined the term topological explanation for explanations of phenomena
in terms of topological properties (including the structural properties of a graph). Although
his focus is on ecological systems and evolutionary contexts, one important explanandum he
addresses is equally relevant in systems biology: namely, robustness. Huneman argues that an
explanation of a system’s general robustness to random node elimination in terms of its scale-free
network structure is a topological explanation. A topological explanation appeals to a system’s
basic organization, in the same way the network motif explanations mentioned above do (Levy
and Bechtel 2013), but it abstracts away from even the generic interactions or temporal features
Figure 27.2 An illustration of two kinds of large-scale networks. In the scale-free network, highly
connected hub nodes are visualized in lighter gray. Reprinted from Barabási and
Oltvai (2004) with permission from Nature Reviews Genetics, Macmillan Publishers Ltd,
copyright 2004
366
seen in motifs. This is at odds with mechanistic explanation, if the latter is to include significant
physical detail (Craver 2007; Kaplan and Craver 2011), or if a mechanistic account’s explana-
tory status is taken to increase when more detail is added (Kaplan 2011; see Chapter 20). In any
case, the fact that topological explanations neither list specific activities nor trace their operation
from set-up to termination conditions is Huneman’s primary reason for contrasting this type of
explanation against mechanistic explanation.
Another case of topological explanation in the context of systems biology is the explanation
of vulnerability (the opposite of robustness) in terms of bowtie structures (Jones 2014). A bowtie
structure is a molecular network with the shape of a bowtie (Figure 27.3), in which it is obvious
that the bowtie’s core is the weakest link because its deactivation (compared to any other node)
will probably damage the whole network’s functionality. A concrete example is the explanation
of why the human immune system is vulnerable to attacks on CD4+ T-cells (by HIV). The
reason is that the molecular network of intercellular interactions and signaling pathways forms a
bowtie that has the CD4+ T-cell type as its core (Figure 27.3; Kitano and Oda 2006). Generally,
then, it holds for scale-free networks that they are robust to random perturbations but vulnerable
to attacks on the highly connected nodes (hubs or bowtie cores) that participate in a large num-
ber of interactions. Topological explanations such as these may well be instances of what some
philosophers have discussed as distinctively mathematical explanations in natural science, which
have even been claimed to offer non-causal explanations of physical phenomena (Lange 2013).
Our reason for invoking topological explanation, however, is simply to show how it contrasts
with classic mechanistic accounts (see Bechtel 2015; Woodward 2013).
Despite these insights, the value of graph-theoretical analysis for biological research is a
contested issue, and critics have pointed to problems with the generalizations made about such
networks (Arita 2004). For instance, whether gene regulatory networks are scale-free has been
CDS+ T-cell cytotoxic

macro- T-cell
phage
macrophage
type-1
natural T-helper
killer cell B-cell
cell
plasma
fibroblast B-cell
cell
type-2
immature mature CD4+
T-helper
dendritic dendritic naïve
epithelial cell eosiniphil
cell cell T-cell
cell leukocyte
type-1
natural type-3
regulatory
killer T-helper cell
T-cell
T-cell
mast cell
Figure 27.3 Bowtie network of CD4+ T-cells. Reproduced from Jones (2014) with permission from
Erkenntnis, Springer, copyright 2014
367
disputed, because some biological networks also exhibit properties similar to random networks
(Barabási and Oltvai 2004; Keller 2005). Another common challenge is that networks usually
provide a static picture of cellular systems, because the data that network edges are based on
combine the totality of interactions that have been measured. However, all the edges repre-
sented need not be active at the same time or in the same location in vivo. A recent development,
therefore, is to include temporal change when constructing topological mappings. When time-
course data are used, distinctions can be made between a “party hub,” which interacts with
many entities at the same time, and a “date hub,” which interacts with only a few other entities
despite having many overall connections and interaction partners (Han et al. 2004). In yeast
metabolism research, protein-interaction and gene expression data have been used to develop a
time-dependent network that is sensitive to which proteins interact at a particular phase of the
cell cycle (de Lichtenberg et al. 2005). This has provided new insight into the processes underly-
ing the periodization of protein synthesis.
As section 1 discussed, research methods in systems biology have huge potential not only
for extending mechanistic research but also for providing novel insights into how and why
biological systems are organized into generalizable schemes with broad applications. As we
have demonstrated, graph-theoretical analysis affords a quantitative understanding of biologi-
cal function and makes possible a comparison of organizational schemes in different functional
systems. As well as cellular systems, neuronal, ecological, and even non-biological communica-
tion and transport networks are often scale-free or small-world networks, and can be analyzed
accordingly. Now we will show how network and systems analysis can be taken even fur-
ther epistemically, in a way that provides additional philosophical insight into the relationship
between systems biology and mechanistic research.
3. Searching for and using design principles

An important research question in systems biology is the extent to which biological func-
tions rely on general design principles that are largely independent of specific causal details
and particular contexts of implementation (Poyatos 2012). Design principles are abstractions
that describe characteristic organizational features of importance for a system’s functionality,
such as negative feedback control, network motif configurations, or common architectures
of biological and engineered networks. Aside from understanding how these design princi-
ples are causally instantiated in specific biological systems, an important explanatory question
is why the same basic principles can describe the functioning of so many different systems.
Some philosophers have recently argued that certain abstract models in systems biology, when
answering that question, provide non-mechanistic design explanations that focus on generaliz-
able constraints for biological functions (Braillard 2010; Green 2015b). In contrast, discus-
sions of mechanisms have typically interpreted abstract models solely as heuristic tools or as
mechanism schemas that guide the formulation of more realistic models (Matthiessen 2017;
see Chapter 19). This resonates with the perspective of many experimental biologists, but has
long been opposed by proponents of systems theory (Green and Wolkenhauer 2013). Rather
than assuming that a model is useful only insofar as it explains a biological system in concrete
detail, current philosophical investigations of design explanations (and of topological expla-
nations) are motivated by the goal of making sense of why some scientists rely on abstract
models even in situations where more detailed models exist.
One illuminating example is how biologists investigate systems exhibiting robust perfect adap-
tation (RPA) from an engineering perspective. RPA is the capacity of a system with sensors to
return to the exact pre-stimulus activity after a stimulus-response reaction. This is important
368
because it maintains the responsiveness of sensors. Creating designs with RPA is a goal in
engineering human-made systems. Biological systems also exhibit RPA. Examples include the
regulation of calcium homeostasis in mammals and membrane turgor pressure in yeast (Briat
et al. 2017). In bacterial chemotaxis (movement in response to external chemical stimuli), RPA
pertains to the regained responsiveness of transmembrane receptors (i.e., sensors) that detect
changes in the concentration of chemicals in the environment. Adding a repellent to the bacte-
rial environment leads to changes in the bacterial tumbling frequency (and thereby to random
reorientations in space), but the receptor system returns very quickly to its equilibrium value.
This enables the receptors to become sensitive to new changes, even if the repellent concentra-
tion continues increasing, which occurs when the bacterium swims along a chemical gradient.
In the case of the bacterium E. coli, the mechanistic basis of its chemotactic RPA is known. It
consists of transmembrane receptors, a signal transduction pathway inside the bacterium, and
its connection to the flagellar motor. A feedback loop from the intracellular process back to the
transmembrane receptor is important for achieving RPA (Barkai and Leibler 1997).
The explanatory issue we are highlighting with bacterial chemotaxis is the question of what
generic properties (abstract organizational features) make it possible for any system—not just
E. coli—to exhibit RPA. The answer is integral feedback control (Yi et al. 2000). Used in engi-
neering, integral feedback control is known in mathematical control theory as a special case of
the internal model principle. When the environmental input u changes (see Figure 27.4), the
difference between the actual output y1 and the desired output y0—the equilibrium value of the
receptor—is fed back into the system as the integral of the system error. This feedback functions
as a signal for the renormalization of the receptor, so that integral feedback control is sufficient
for RPA. A crucial insight is provided by a theorem of Yi et al. (2000), which shows that (at
least in linear systems) integral feedback control is necessary for achieving RPA. This explains
why any system that exhibits RPA has to have an organization that instantiates integral feedback
control (Iglesias 2013). E. coli should be no exception in this regard, and Yi et al. (2000) point
out that an influential dynamic mechanistic model of RPA in bacteria (Barkai and Leibler 1997)
does indeed embody the basic principle of integral feedback control.
−y0
y1
u + k + y
−x −∫
.
x=y y(t) → 0 as t → ∞
y = y1−y0 iff
= k(u−x)−y0 k>0
Figure 27.4 Diagram showing the abstract principle of integral feedback control. Reproduced from Yi
et al. (2000), with permission from Proceedings of the National Academy of Sciences, copyright
(2000) National Academy of Sciences, USA
369
We can compare this account of chemotaxis with a standard mechanistic explanation. The
latter would show how a particular structural organization causally generates and thus explains
some function (e.g., RPA). In contrast, what Wouters (2007) calls a design explanation proceeds
in the opposite direction, as the function to be performed (RPA in our case) explains the presence
of some structural organization (integral feedback control). Using examples from physiology and
functional anatomy, Wouters argues that such an explanation is non-causal, because it is based
on law-like dependency relations between structures and functions. It maps out the possible
structural realizers of a certain function, without going into a diachronic account of how the
realizer or the need for the function came about causally. But regardless of where one stands on
the status of non-causal explanations, in the case of bacterial chemotaxis, the design explanation
does not just offer a list of the various concrete mechanisms that perform RPA (e.g., a trans-
membrane receptor, six Che proteins, and other details in E. coli). This is a non-mechanistic
explanation in that it points to the abstract organizational feature of integral feedback control as a
generic property that any system exhibiting RPA must instantiate (Braillard 2010). This explan-
atory aim addresses a why-question that is distinct from the aim of explaining how a behavior is
mechanistically produced in some specific system.
The example of design principles underpinning RPA in engineering and biology also pro-
vides more general philosophical lessons about the theoretical relevance of delineating the space
of biological possibility. Mechanistic accounts have typically taken how-possibly models to
have less explanatory power than how-actually models (Craver 2007; Kaplan 2011; Kaplan and
Craver 2011; see Chapter 19). Yet understanding the wider constraints on biological variation
can in some contexts be of higher importance than describing how a specific function is causally
produced in any specific system. Design principles can help researchers understand why the same
structural patterns are found across different contexts: as a result of the constraints on possible
architectures that can realize a given function. Importantly, this is not to be understood as a
question that presupposes natural selection as the answer. Rather, the why-question addressed
here is about the physically determined boundaries of the design space for a given function.
Design principles do, however, have significance for evolutionary research as well as func-
tional biology. Investigations of the constraints on evolutionary and developmental trajectories
have often been associated with rather speculative “structuralist” accounts, but some of these
ideas have gained new relevance in the context of evolutionary systems biology (Green et al. 2015;
Jaeger and Crombach 2012). Evolutionary systems biology is an umbrella term for very diverse
approaches (O’Malley 2012), but one important aim is to investigate why certain general pat-
terns arise in evolution. This is often done via models that represent the in silico evolution of gene
regulatory networks. Structures like the network motifs discussed in section 2 are often assumed
to be common because of regulatory functions favored by natural selection (Alon 2007). Yet
evolutionary simulation studies suggest that common structural patterns of networks, such as
feedforward loops, may also result from constraints on genome evolution. These constraints are
inherent in the mutational dynamics of gene duplication, deletion, and recombination (Cordero
and Hogeweg 2006). Research on evolutionary design principles, when understood as general
patterns occurring from evolutionary trajectories, can thus generate insight into the potential
and limits of biological variation.
Design principles also identify the generic features that unite diverse systems exhibiting simi-
lar functional patterns (Green 2015b). By relating specific systems to general functional types,
such as signal amplifiers, filters, or homeostatic regulators, these abstract principles facilitate
the transfer of theoretical frameworks across disciplinary borders. Aside from this epistemic
role, such structure-function mappings can serve practical aims. Similarly to MacLeod and
Nersessian’s (2015) emphasis on practical purposes such as control and prediction of modeling
370
in systems biology, research on possibility spaces for biological structures can have practical goals
such as templates for synthetic biology designs. Synthetic biology is the biological construction
of material models, usually guided by mathematical modeling. “How-possibly models” can in
this context be more important than “how-actually models” because they elucidate the neces-
sary structures for a certain function, like RPA, or reveal whether there are simpler possible
designs than the ones found in nature (Briat et al. 2017; Ma et al. 2009).
The upshot of this discussion is that abstract models are not always stepping stones toward
more detailed mechanistic models. Aside from the practical purposes of control and technologi-
cal implementation, abstract design principles afford an understanding of why causally different
systems in biology and engineering share certain organizational features, and how they are situ-
ated within larger spaces of physically possible designs. Consequently, an exclusive philosophical
focus on mechanistic explanation (and even on dynamic mechanistic explanation) risks missing
out on these diverse epistemic activities in systems biology.
4. Discussion and outlook

Research in systems biology shows how strategies of abstraction are used in biology not only to
simplify the task of identifying mechanisms but also to elucidate system-level patterns of organi-
zation that may not be visible at the level of the molecular details. Mechanistic accounts have
usually been framed in opposition to explanatory unification, understood as the subsumption of
the particular to general laws or explanatory schemas. But network modeling and the quest for
design principles suggest an alternative way of thinking about the role of unification in biology:
not via reduction of the particular to the general, but through abstraction from causal details for
the purpose of identifying generic organizational patterns.
Mathematical models (including network models and design principles) serve various roles
in systems biology. Generally, mathematical frameworks provide a more rigorous way of
exploring the extent to which biological functions are underpinned by characteristic organiza-
tional structures. Mathematical frameworks can also make engineering analogies more precise.
Section 2 mentioned the identification of functional network motifs based on mathematically
guided screening for overabundant circuit types. This search is inspired by an analogy to design
principles in electronic networks, and the structural decomposition of the network preceded the
functional analysis of the modules of the network (see Chapter 34). In other cases, the biological
function is known and systems biologists set out to explore the extent to which the function is
similarly realized in engineered systems (e.g., robustness). Mathematical abstractions and design
principles can articulate constraints that delimit the search space for an analysis. Delineating
search space may serve the development of mechanistic explanations. At the same time, net-
work models and design principles provide an understanding complementary to mechanistic
explanation. Although an important virtue of mechanistic explanations is to make sense of
biological diversity through attention to specific causal difference-makers and material composi-
tion, abstraction strategies can help scientists see similarities in the way functional systems—from
airplanes to organisms—are organized.
Generally, a focus by philosophers on the issue of mechanistic explanation has left many
aspects of systems biology unexplored. We have pointed to the use of models for the purpose
of prediction, control, or the creation of simple and efficient designs that can be implemented
in synthetic organisms. Another topic of major interest to systems biologists that is philosophi-
cally rich is robustness. In many cases when a system maintains its functioning despite noise
and even major perturbations, this is due to dynamic reorganization, where the organismal sys-
tem responds flexibly by changing interaction patterns and levels, including establishing new
371
interactions (Wagner 2005). This puts pressure on the assumption that systems biology always
investigates mechanisms (on a machine-like conception), or that all explanations about systems
exhibiting dynamic reorganization or robustness are mechanistic (in the sense of referencing the
mechanism’s specific organization; Brigandt 2015; Gross 2015; Woodward 2013).
Some of the questions that deserve more attention by philosophers pertain to issues that are
currently controversial within the systems biology community itself. One is the question of
whether complex living systems can be understood in terms of engineering notions (Braillard
2015; Green 2015a), and particularly whether the heuristic assumption of modularity is war-
ranted. Research on network motifs is often predicated on the idea that an individual motif is
modular, meaning that its functionality is unaffected by the system context in which it occurs
(section 2). The traditional mechanistic strategies of decomposition into distinct components
and the localization of some function to a certain component also resonate with the assump-
tion that biological systems are modular. However, many systems biologists observe highly
integrated functionality across large-scale networks, which suggests that systems need to be
investigated not in terms of modularity but via more connectivist perspectives that can cap-
ture features emerging from system-wide dynamics (Huang 2004; see also Bassingthwaighte
et al. 2009; Bechtel 2015).
Our discussion draws attention to a wide range of explanatory and modeling strategies in systems
biology. We have shown how some explanatory aims and outputs are not mechanistic according
to standard philosophical interpretations of mechanistic explanations, and indeed, that some of the
practices in systems biology lie outside existing philosophical frameworks. But well beyond these
negative insights, we have depicted the wealth of modeling approaches at work in systems biology,
and how further philosophical scrutiny of them will enhance the investigation of biological systems
and philosophical accounts of mechanistic explanation and explanation in general.
References
Alberghina, L. and H. V. Westerhoff (eds) (2005) Systems Biology: Definitions and Perspectives, Berlin:
Springer.
Albert, R. (2005) “Scale-Free Networks in Cell Biology,” Journal of Cell Science 118: 4947–57.
Alon, U. (2007) An Introduction to Systems Biology: Design Principles of Biological Circuits, Boca Raton,
FL: Chapman & Hall / CRC Press.
Arita, M. (2004) “The Metabolic World of Escherichia coli Is Not Small,” Proceedings of the National Academy
of Sciences USA 101: 1543–7.
Baetu, T. (2015) “From Mechanisms to Mathematical Models and Back to Mechanisms: Quantitative
Mechanistic Explanations,” in P.-A. Braillard and C. Malaterre (eds.), Explanation in Biology: An Enquiry
into the Diversity of Explanatory Patterns in the Life Sciences, Dordrecht: Springer, pp. 345–63.
Barabási, A.-L. and Z. N. Oltvai (2004) “Network Biology: Understanding the Cell’s Functional
Organization,” Nature Reviews Genetics 5: 101–13.
Barkai, N. and S. Leibler (1997) “Robustness in Simple Biochemical Networks,” Nature 387: 913–17.
Bassingthwaighte, J., P. Hunter and D. Noble (2009) “The Cardiac Physiome: Perspectives for the
Future,” Experimental Physiology 94: 597–605.
Bechtel, W. (2012) “Understanding Endogenously Active Mechanisms: A Scientific and Philosophical
Challenge,” European Journal for Philosophy of Science 2: 233–48.
—— (2015) “Can Mechanistic Explanation Be Reconciled with Scale-Free Constitution and Dynamics?”
Studies in History and Philosophy of Biological and Biomedical Sciences 53: 84–93.
—— (2016) “Using Computational Models to Discover and Understand Mechanisms,” Studies in History
and Philosophy of Science 56: 113–121.
Bechtel, W. and A. Abrahamsen (2005) “Explanation: A Mechanist Alternative,” Studies in History and
—— (2010) “Dynamic Mechanistic Explanation: Computational Modeling of Circadian Rhythms as an
Exemplar for Cognitive Science,” Studies in History and Philosophy of Science 41: 321–33.
372
Bechtel, W. and R. C. Richardson (1993) Discovering Complexity: Decomposition and Localization as Strategies
Boogerd, F. C., F. J. Bruggeman, J.-H. S. Hofmeyr and H. V. Westerhoff (eds) (2007) Systems Biology:
Philosophical Foundations, Amsterdam: Elsevier.
Braillard, P.-A. (2010) “Systems Biology and the Mechanistic Framework,” History and Philosophy of the
Life Sciences 32: 43–62.
—— (2015) “Prospect and Limits of Explaining Biological Systems in Engineering Terms,” in
P.-A. Braillard and C. Malaterre (eds.), Explanation in Biology: An Enquiry into the Diversity of Explanatory
Patterns in the Life Sciences, Dordrecht: Springer, pp. 319–44.
Briat, C., A. Gupta and M. Khammash (2017) “Antithetic Integral Feedback Ensures Robust Perfect
Adaptation in Noisy Biomolecular Networks,” arXiv.org e-Print Archive arXiv:1410.6064v7 [math.OC].
Brigandt, I. (2013) “Systems Biology and the Integration of Mechanistic Explanation and Mathematical
Explanation,” Studies in History and Philosophy of Biological and Biomedical Sciences 44: 477–92.
—— (2015) “Evolutionary Developmental Biology and the Limits of Philosophical Accounts of Mechanistic
Explanation,” in P.-A. Braillard and C. Malaterre (eds.), Explanation in Biology: An Enquiry into the
Diversity of Explanatory Patterns in the Life Sciences, Dordrecht: Springer, pp. 135–73.
Carusi, A. (2014) “Validation and Variability: Dual Challenges on the Path from Systems Biology to
Systems Medicine,” Studies in History and Philosophy of Biological and Biomedical Sciences 48: 28–37.
Cordero, O. X. and P. Hogeweg (2006) “Feed-Forward Loop Circuits as a Side Effect of Genome
Evolution,” Molecular Biology and Evolution 23: 1931–6.
Craver, C. F. (2007) Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience, Oxford: Oxford
University Press.
Donaghy, J. (2014) “Temporal Decomposition: A Strategy for Building Mathematical Models of Complex
Metabolic Systems,” Studies in History and Philosophy of Biological and Biomedical Sciences 48: 1–11.
Fagan, M. B. (2012) “Waddington Redux: Models and Explanation in Stem Cell and Systems Biology,”
Biology & Philosophy 27: 179–213.
—— (2016) “Stem Cells and Systems Models: Clashing Views of Explanation,” Synthese 193: 873–907.
Green, S. (2015a) “Can Biological Complexity Be Reverse Engineered?” Studies in History and Philosophy
of Biological and Biomedical Sciences 53: 73–83.
—— (2015b) “Revisiting Generality in Biology: Systems Biology and the Quest for Design Principles,”
Biology & Philosophy 30: 629–52.
Green, S., M. Fagan and J. Jaeger (2015) “Explanatory Integration Challenges in Evolutionary Systems
Biology,” Biological Theory 10: 18–35.
Green, S. and O. Wolkenhauer (2013) “Tracing Organizing Principles: Learning from the History of
Systems Biology,” History and Philosophy of the Life Sciences 35: 553–76.
Gross, F. (2015) “The Relevance of Irrelevance: Explanation in Systems Biology,” in P.-A. Braillard and
C. Malaterre (eds.), Explanation in Biology: An Enquiry into the Diversity of Explanatory Patterns in the Life
Sciences, Dordrecht: Springer, pp. 175–98.
Han, J.-D. J., N. Bertin, T. Hao, D. S. Goldberg, G. F. Berriz, L. V. Zhang, D. Dupuy, et al. (2004)
“Evidence for Dynamically Organized Modularity in the Yeast Protein-Protein Interaction Network,”
Nature 430: 88–93.
Huang, S. (2004) “Back to the Biology in Systems Biology: What Can We Learn from Biomolecular
Networks?” Briefings in Functional Genomics and Proteomics 2: 279–97.
Huneman, P. (2010) “Topological Explanations and Robustness in Biological Sciences,” Synthese 177:
213–45.
Ideker, T., T. Galitski and L. Hood (2001) “A New Approach to Decoding Life: Systems Biology,” Annual
Review of Genomics and Human Genetics 2: 343–72.
Iglesias, P. (2013) “Systems Biology: The Role of Engineering in the Reverse Engineering of Biological
Signaling,” Cells 2: 393–413.
Issad, T. and C. Malaterre (2015) “Are Dynamic Mechanistic Explanations Still Mechanistic?” in
P.-A. Braillard and C. Malaterre (eds.), Explanation in Biology: An Enquiry into the Diversity of Explanatory
Patterns in the Life Sciences, Dordrecht: Springer, pp. 265–92.
Jaeger, J. and A. Crombach (2012) “Life’s Attractors: Understanding Developmental Systems through
Reverse Engineering and in Silico Evolution,” in O. S. Soyer (ed.), Evolutionary Systems Biology, New
York: Springer, pp. 93–119.
Jones, N. (2014) “Bowtie Structures, Pathway Diagrams, and Topological Explanation,” Erkenntnis 79:
1135–55.
373
Kaplan, D. M. (2011) “Explanation and Description in Computational Neuroscience,” Synthese 183:

339–73.
Kaplan, D. M. and C. F. Craver (2011) “The Explanatory Force of Dynamical and Mathematical Models
in Neuroscience: A Mechanistic Perspective,” Philosophy of Science 78: 601–27.
Karr, J. R., J. C. Sanghvi, D. N. Macklin, M. V. Gutschow, J. M. Jacobs, B. Bolival, N. Assad-Garcia,
J. I. Glass and M. W. Covert (2012) “A Whole-Cell Computational Model Predicts Phenotype from
Genotype,” Cell 150: 389–401.
Keller, E. F. (2005) “Revisiting ‘Scale-Free’ Networks,” BioEssays 27: 1060–8.
Kitano, H. (ed.) (2001) Foundations of Systems Biology, Cambridge, MA: MIT Press.
Kitano, H. and K. Oda (2006) “Robustness Trade-Offs and Host–Microbial Symbiosis in the Immune
System,” Molecular Systems Biology 2: 22.
Lange, M. (2013) “What Makes a Scientific Explanation Distinctively Mathematical?” British Journal for the
Levy, A. and W. Bechtel (2013) “Abstraction and the Organization of Mechanisms,” Philosophy of Science
80: 241–61.
de Lichtenberg, U., L. J. Jensen, S. Brunak and P. Bork (2005) “Dynamic Complex Formation During the
Yeast Cell Cycle,” Science 307: 724–7.
Ma, W., A. Trusina, H. El-Samad, W. A. Lim and C. Tang (2009) “Defining Network Topologies That
Can Achieve Biochemical Adaptation,” Cell 138: 760–73.
MacLeod, M. and N. J. Nersessian (2014) “Strategies for Coordinating Experimentation and Modeling
in Integrative Systems Biology,” Journal of Experimental Zoology Part B: Molecular and Developmental
Evolution 322: 230–9.
—— (2015) “Modeling Systems-Level Dynamics: Understanding without Mechanistic Explanation in
Integrative Systems Biology,” Studies in History and Philosophy of Biological and Biomedical Sciences 49:
1–11.
Matthiessen, D. (2017) “Mechanistic Explanation in Systems Biology: Cellular Networks,” British Journal
for the Philosophy of Science 68: 1–25.
Mekios, C. (2015) “Explanation in Systems Biology: Is It All about Mechanisms?” in P.-A. Braillard and
C. Malaterre (eds.), Explanation in Biology: An Enquiry into the Diversity of Explanatory Patterns in the Life
Sciences, Dordrecht: Springer, pp. 41–72.
Mesarović, M. D. (1968) “Systems Theory and Biology—View of a Theoretician,” in M. D. Mesarović
(ed.), Systems Theory and Biology: Proceedings of the III Systems Symposium at Case Institute of Technology,
Berlin: Springer, pp. 59–87.
O’Malley, M. A. (2012) “Evolutionary Systems Biology: Historical and Philosophical Perspectives on an
Emerging Synthesis,” in O. S. Soyer (ed.), Evolutionary Systems Biology, New York: Springer, pp. 1–28.
O’Malley, M. A. and J. Dupré (2005) “Fundamental Issues in Systems Biology,” BioEssays 27: 1270–6.
O’Malley, M. A. and O. S. Soyer (2012) “The Roles of Integration in Molecular Systems Biology,” Studies
in History and Philosophy of Biological and Biomedical Sciences 43: 58–68.
Poyatos, J. (2012) “On the Search for Design Principles in Biological Systems,” in O. S. Soyer (ed.),
Evolutionary Systems Biology, New York: Springer, pp. 183–93.
Richardson, R. C. and A. Stephan (2007) “Mechanism and Mechanical Explanation in Systems Biology,”
in F. C. Boogerd, F. J. Bruggeman, J.-H. S. Hofmeyr and H. V. Westerhoff (eds.), Systems Biology:
Philosophical Foundations, Amsterdam: Elsevier, pp. 123–44.
Théry, F. (2015) “Explaining in Contemporary Molecular Biology: Beyond Mechanisms,” in P.-A. Braillard
and C. Malaterre (eds.), Explanation in Biology: An Enquiry into the Diversity of Explanatory Patterns in the
Life Sciences, Dordrecht: Springer, pp. 113–33.
Wagner, A. (2005) Robustness and Evolvability in Living Systems, Princeton, NJ: Princeton University Press.
Woodward, J. (2013) “Mechanistic Explanation: Its Scope and Limits,” Aristotelian Society Supplementary
Volume 87: 39–65.
Wouters, A. (2007) “Design Explanation: Determining the Constraints on What Can Be Alive,” Erkenntnis
67: 65–80.
Yi, T.-M., Y. Huang, M. I. Simon and J. Doyle (2000) “Robust Perfect Adaptation in Bacterial Chemotaxis
through Integral Feedback Control,” Proceedings of the National Academy of Sciences USA 97: 4649–53.
374
28
IN NEUROSCIENCE1
1. Introduction
Perhaps the most striking thing you notice when thumbing through the pages of a neuroscience
textbook like Principles of Neural Science (Kandel et al. 2012) are the elaborate diagrams of
the central nervous system, brain, spinal cord, synapses, neurons, and molecules. It is equally
striking that whatever topic you look at, whether the action potential, synaptic transmission,
cognition, or perception, it is inevitably described and explained using the word “mechanism.”
You can become so accustomed to seeing and hearing about mechanisms in neuroscience that
it never occurs to you to question what mechanisms in fact are, or what that choice of termi-
nology implies. As it turns out, there are tricky philosophical problems lurking beneath the
surface of mechanism talk in neuroscience.
In this chapter we explore some of the ways that mechanisms are invoked in neuro-
science, and look at a selection of the philosophical problems that arise when trying to
understand mechanistic explanations (several chapters in this volume go into more detail
about particular philosophical problems encountered in mechanistic explanation in neuro-
science, and Chapter 6 also describes some of the history we discuss below). We begin in
section 2 by introducing a series of historical case studies that illustrate how neuroscien-
tists have depended on mechanistic metaphors in their efforts to understand the mind and
brain, and how their mechanistic explanations have developed over time. We revisit these
examples throughout the remainder of the chapter. In section 3, we use these case studies
to highlight what contemporary philosophers have identified as the fundamental features of
mechanisms and mechanistic explanation. In section 4, we consider some of the methodo-
logical issues that arise in neuroscience including (1) how to integrate psychological with
neural models, (2) how to generalize findings in model organisms like the sea slug Aplysia to
human learning and memory, and (3) whether to favor top-down or bottom-up methods.
2. A short history of neural mechanisms of learning and memory

The historical examples we focus on are episodes in the search for the neural mechanisms of
learning and memory, which are among the most important cognitive phenomena studied
in neuroscience. These examples illustrate how mechanisms are discovered, reasoned about,
represented, and how they figure in explanations.
375
In the seventeenth century, in Treatise on Man, the French philosopher René Descartes
likened human beings to machines. Descartes drew an analogy between the movements of
human beings and the movements of the automated figures in the fountains of the Royal
Gardens at St. Germain. Descartes described how when visitors to the gardens step on certain
tiles, statues of Roman Gods, Goddesses, and other mythical creatures move, gesture, play
music, spray water, and speak. Pressing on the tiles triggers a flow of water from storage tanks
beneath the fountains through a network of hidden pipes. The flow of water then causes the
figures, which are connected to machinery like springs and cogs, to move. Descartes also
compared these motions to those of a clock or a mill, which can be made to move continu-
ously, not just in response to an external push.
Descartes claimed that a similar set of events takes place in the human nervous system when
simple reflexes are triggered. The brain, according to Descartes, contained ventricles filled
with “animal spirits” or “a very fine air or wind,” which reached the ventricles via the blood
(Descartes 1664/1985, 100). He believed that the ventricles were connected to networks of
nerves, which he thought were mostly hollow save for a set of small fibers running their length.
According to Descartes, the nerves are connected to the brain in such a way that stimula-
tion from the periphery, which tugs on the fibers, is communicated to the brain, triggering a
response. Tugging on the fibers opens pores in the nerve, allowing animal spirits to flow from
the ventricles through the nerve to the musculature, causing motion, he claimed. Descartes
illustrated this with a drawing of a man placing his foot near a flame. He outlined a series
of events that supposedly take place in the man’s nervous system from the moment his skin
contacts the flame to the moment he pulls his foot away. According to Descartes, the “tiny
particles” or molecules that comprise the fire cause the area of skin that they touch to move.
When the skin moves, a nerve fiber attached to it is pulled, causing a pore at the other end of
the nerve to open, in turn allowing animal spirits to flow through the nerve to various muscles,
causing the muscles to change shape, and finally pulling the man’s foot away from the flame.
Figure 28.1 Descartes’ illustration of the man pulling his foot away from the flame. Reproduced from
Descartes (1664/1985), out of copyright
376
Mechanistic explanation in neuroscience
Flow of animal spirits down other nerves also causes the man’s head and eyes to turn to look at
the flame, he says (Descartes 1664/1985, 102) (see Figure 28.1).
From this simple mechanical account of reflexes, Descartes built up a model of the nervous
system to explain more complex phenomena like learning and memory. He suggested that asso-
ciative memory traces—the heat of the flame and how it looks—“are imprinted on the internal
part of the brain”; however, he did not have much to say about how that imprinting happens.
If we move ahead to the mid-nineteenth century, the Russian physiologist Ivan Pavlov
made the next significant advances in discovering the neural mechanisms of learning. In the
process of investigating the alimentary or salivation reflex in dogs, Pavlov discovered that his
canine subjects salivate not only in the presence of food, but also in the presence of stimuli
that regularly precede the presentation of food, such as a tone, or the experimenter entering
the room. He described the first type of reflex (e.g., to food) as “inborn,” involving “regular
causal connections between definite external stimuli acting on the organism and its necessary
reflex actions” (Pavlov 1927/2003, 16). However, he hypothesized that a second type of reflex
(e.g., to a tone) involves different “mechanisms” operative in “higher nervous centres” (Pavlov
1927/2003, 25) and is “built up gradually in the course of an animal’s own individual existence”
(Pavlov 1927/2003, 25). In contrast to Descartes, who thought the mind influences the body
through the pineal gland, Pavlov claimed non-physical or psychic causes are not responsible
for either innate or conditioned reflexes. Rather, reflexes can be explained solely in terms of
neural mechanisms mediating between stimuli and responses. This was in line with the views of
mechanist physiologists like Hermann Helmholtz.
To identify “the precise conditions under which new conditioned reflexes are established”
(Pavlov 1927/2003, 26), Pavlov and his colleagues ran many rigorously controlled experi-
ments. On the basis of their data, Pavlov concluded that a conditioned reflex can be established
if: (1) the presentation of the conditioned stimulus (e.g., a tone) precedes the unconditioned
stimulus (e.g., food), (2) the two stimuli overlap in time, (3) the animal is alert and healthy,
(4) the conditioned stimulus is an environmentally familiar one to which the animal is otherwise
indifferent, and (5) the investigator ensures that the only stimuli operative in the experiment
are the conditioned and unconditioned stimuli.
Having reliably produced conditioned reflexes in many canine subjects, Pavlov hypothesized
the physiological conditions that allow their formation: “the linking up of impulses in different
areas of the brain, by the formation of new nervous connections” is the “nervous mechanism”
by which “new conditioned reflexes” are formed (1927/2003, 37). More specifically, Pavlov
said “it appears that the cells predominantly excited at a given time” by an unconditioned
stimulus (food) “become foci attracting to themselves the nervous impulses aroused by” the
conditioned stimulus (tone), and that these impulses “on repetition tend to follow the same
path and so to establish conditioned reflexes” (1927/2003, 38). Pavlov illustrated what he had
in mind by appeal to “telephonic installation.” He explained that he could telephone his labo-
ratory directly, or he could call the operator to connect him to the laboratory. (In those days,
operators would manually connect lines by plugging cables into jacks on a switchboard.) Both
methods would result in the same outcome. However, “whereas the private line provides a
permanent and readily available cable” much like the neural pathway of innate reflexes, “the
other line necessitates a preliminary central connection to be established” much like how the
neural pathway carrying information about the conditioned stimuli must be connected to
the innate pathway. Pavlov did not know precisely the location of the formation of these
new connections—he thought that it was possible that it could occur “within the cortex” or
“between the cortex and subcortical areas” (Pavlov 1927/2003, 37). He also had no explanation
for how such changes in neural connectivity might occur.
377
An explanation began to emerge at the end of the nineteenth century. Wilhelm His (1886),
working with growing nerve cells, August Forel (1887), working on nerve cell degeneration,
and the great Spanish histologist Santiago Ramón y Cajal (1888), using Camillo Golgi’s (1873)
silver nitrate stain on unmyelinated nerve cells, suggested that nerve cells are independent ana-
tomical and functional units (rather than a physically connected web of fibers as previously
believed). Golgi’s illustrations demonstrated many variations of the nerve cell’s typical struc-
ture of cell body, single axon, and branching dendrites, in different brain regions. Cajal (1890)
showed how growing neurons push their growth code outwards, and gradually form more
dendrites and axon collaterals.
In 1891, Cajal discovered that sensory nerves have their dendrites in the periphery and
axons projecting toward the brain, while motor cells are the opposite way around. His “law of
dynamic polarization” hypothesized that conduction of impulses travel in one direction only,
from dendrite to cell body to axon. In the eighteenth century, Luigi Galvani had established
that it is electric currents, not corpuscles (i.e., animal spirits), that transmit nerve impulses.
However, Cajal and many of his contemporaries believed that neurofibrils contained in nerve
cells “underlie the mechanism of neuronal impulse transmission” (Cajal 1890, 95), harkening
back to Descartes’s account (see Figure 28.2).
These combined discoveries led to speculation in the 1890s that the growth or retraction of
dendritic connections and axon collateral branches might account for learning. Cajal (1894a)
suggested that genius in a subject such as music might involve increased branching of cer-
tain neurons’ dendrites and axons. This was pure conjecture, however. Cajal was an anatomist
Figure 28.2 Cajal’s drawing of Purkinje cells with basket endings. Reproduced from Cajal (1894a), out
of copyright
378
working almost exclusively with histological methods (slicing and staining specimens, then
examining them under a light microscope), which did not lend themselves well to discovering
how learning occurs, nor indeed to discovering much about how nervous impulses are com-
municated between nerve cells. Physiological methods were required to discover the functional
import of Cajal’s anatomical findings.
The English physiologist Charles Scott Sherrington first introduced the concept of the syn-
apse in 1897 in a textbook he helped edit (Foster 1897). Based on his work on spinal reflexes,
Sherrington had deduced that there is a significant delay in the speed of neural impulses where
there are connections made between several nerve cells along the way to or from the spinal cord,
rather than single axons traveling the whole distance. He attributed this delay to an “intercel-
lular barrier” or membrane (Bennett 1999). Cajal had convincingly argued that axon collaterals
do not directly fuse with the cells they come into contact with, but the hypothesized junction
could not be seen under a light microscope. Sherrington’s work suggested that the synapse
acted as a valve, explaining why conduction occurred in only one direction. He also explored
the relationship between inhibitory and excitatory connections. He remarked that the synapse
offered “an opportunity for some change in the nature of the nervous impulse as it passes from
one cell to the other” (Foster 1897). Physiologists and pharmacologists continued, in the early
twentieth century, to uncover the electrical and chemical mechanisms of synaptic transmission.
Advances in neurophysiological theorizing and methodology in the first half of the twentieth
century were instrumental in connecting this developing knowledge about synaptic transmis-
sion to the phenomena of learning and memory. One such advance came in the form of a
simple neurophysiological postulate put forward by Donald Hebb in The Organization of Behavior
(1949). Synthesizing a broad selection of research from psychology (e.g., Pavlov), neuroanat-
omy (e.g., Cajal), and neurophysiology (e.g. Sherrington, Lorente de Nó), Hebb hypothesized
that just as associative learning at the level of behaving organisms required the repetition and
contiguity of stimuli or stimuli and responses, so, too, did the permanent metabolic changes or
growth processes thought to underlie learning require the contiguous and repetitive excitation
of the neurons carrying information about those stimuli and/or responses.
More specifically, Hebb claimed that
when an axon of [a] cell A is near enough to excite a cell B and repeatedly or persis-
tently takes part in firing it, some growth process or metabolic change takes place in
one or both cells such that A’s efficiency, as one of the cells that fires B, is increased.
(Hebb 1949/2002, 62)
Although the main idea at the heart of Hebb’s postulate was not new, as Hebb acknowledged,
the postulate provided insight into the kinds of methods that could be employed by physiolo-
gists to determine if neurons were plastic in the way that Hebb’s predecessors, like Pavlov and
Cajal, had claimed.
By the 1960s, neurophysiologists had a working model of the neuron as consisting of (1) an
input component (the dendrites), (2) an integrative component (the axon hillock), (3) a conductile
component (the axon), and (4) an output component—the synaptic terminal from which the
neurotransmitter is released (Kandel and Spencer 1968, 69–70). However, what was missing was
“a crucial experiment identifying specifically a change occurring in neural tissues as learning takes
place” (Hilgard 1956, 481). Such crucial experiments came much later in the form of Nobel Prize
winner Eric Kandel and colleagues’ development of a simplified preparation for studying the cel-
lular and molecular mechanisms of simple forms of associative and non-associative learning in the
invertebrate sea mollusc, Aplysia californica.
379
Siphon
Facilitating
interneuron Sensory
neuron
Sensory
Tail neuron
Motor
neuron
Gill
Figure 28.3 The simple neuronal circuit involved in sensitization of the gill-siphon withdrawal reflex
in Aplysia. (Kandel, E.R., Schwartz, J.H., Jessell, T.M., Siegelbaum, S.A., Hudspeth, A.J.
Principles of Neural Science, 5th edition, 2012, McGraw-Hill Education. Reproduced with
permission of McGraw-Hill Education)
Aplysia has a defensive reflex known as the gill-siphon withdrawal reflex. When a tactile
stimulus is applied to the animal’s siphon—a small fleshy spout located above the gill that
expels seawater and waste—it retracts or withdraws the siphon and the gill. In one early set
of experiments, Kandel and colleagues experimentally isolated the sensory neurons that carry
stimulus information from the siphon, the motor neurons to which these sensory neurons
project, and a set of excitatory and inhibitory interneurons that receive input from the sensory
neurons and project to the motor neurons. By isolating the neurons that comprise this simple
circuit (and other circuits to which it was connected), Kandel and colleagues were able to
identify specific cellular and molecular changes that accompany a set of simple forms of asso-
ciative and non-associative learning in Aplysia.
More specifically, they studied a form of learning known as sensitization. In a sensitiza-
tion experiment, an investigator begins by applying a tactile stimulus (e.g., a Q-tip) to an
Aplysia’s gill or siphon, so as to measure the extent and duration of the withdrawal reflex.
The experimenter then delivers a set of noxious shocks to the organism’s tail. Following
these shocks, she reapplies the tactile stimulus and again measures the extent and dura-
tion of the withdrawal reflex. An increase in duration of the withdrawal reflex prior to
the tail shocks compared to after the tail shocks is taken as indicative that the animal has
learned that there is a noxious stimulus in its environment. Textbooks like Principles of
Neural Science (5th edition, 2012) reveal in detailed diagrams what we now know about
the changes in the strength of the synaptic connections that underlie this form of learn-
ing and that they are mediated by specific changes in cellular and molecular activity
(see Figure 28.3).
These historical examples should give an idea of some of the ways that scientists discover
and explain the mechanisms of the brain. In the following sections we’ll revisit these cases,
highlighting the fundamental features of mechanistic explanations, and considering some of the
methodological issues that arise in neuroscience.
380
3. Philosophical accounts of mechanisms in neuroscience

One aim of philosophy of science is to understand the structure of science. Another is
to account for scientific progress. Up until the latter half of the twentieth century, the
examples considered in philosophy of science were taken primarily from the history of
physics. This exclusive focus led to an understanding of science that conceived of its his-
tory as involving the discovery of laws (e.g., planetary motion, gravitational attraction) and
the development of grand unifying theories (e.g., relativity theory). By the middle of the
twentieth century, philosophers characterized scientific explanations as arguments, where
statements of laws and initial conditions were taken to logically imply the observations
to be explained or predicted. Different branches of science were thought to be hierarchi-
cally organized with those that studied the most fundamental things (e.g., particles) at
the bottom and those that studied the least fundamental things (e.g., societies) at the top.
Branches of science were regarded as compartmentalized, and progress within a given
branch (e.g., psychology) was not taken to rely on developments within other branches
(e.g., neurophysiology). Progress between branches was taken to involve “intertheoretic
reduction”—the reduction of theories in “higher-level” sciences like biology to theories
in “lower-level” sciences like physics (see Oppenheim and Putnam 1958; Nagel 1961;
Chapter 16 in this volume).
The history of scientific research on learning and memory that we described above defies
these characterizations in a number of ways. Descartes, Pavlov, Cajal, Hebb, and Kandel
were aiming neither to discover large-scale scientific theories nor to reduce those theo-
ries to physical ones. The kinds of explanations for learning and memory phenomena they
sought combined findings and insights from different areas of science including anatomy,
physiology, psychology, and later biochemistry. These diverse branches of science all aimed at
understanding learning and memory from different angles and seemed to be making progress
interactively rather than independently.
An alternative account of scientific explanation has recently been proposed that provides a
more congenial analysis of the discovery strategies and markers of progress described in these
historical cases, as well as in the biological sciences more generally. This account focuses on
the role of mechanisms in scientific explanation. (See, for example, Bechtel and Richardson
1993/2000; Craver 2007; Glennan 1996; Illari and Williamson 2012; Machamer, Darden,
and Craver 2000.)
The first crucial step in providing a mechanistic explanation is to identify the phenomenon
to be explained (see, for example, Glennan 1996; Bechtel 2008; Craver and Darden 2001).
In each case considered above, the phenomenon for which a mechanism is sought is precisely
delineated. Consider how Descartes conceives of a reflex; it begins with a stimulus—a man
putting his foot into a fire—and ends when the man looks at the fire and retracts his foot
from the flame. Similarly, Pavlov’s conditioned reflexes begin with repeated and contiguous
presentation of the unconditioned stimulus and conditioned stimulus and end with elicitation
of the conditioned response (i.e., salivation) to the conditioned stimulus. Sherrington, Hebb,
and Kandel postulate a very specific set of inputs—repetition and contiguity in firing of two
cells that comprise a synapse—and a very specific output—a change in the way that the
two cells communicate. As Peter Machamer, Lindley Darden, and Carl Craver (MDC 2000, 3)
claim, the phenomena of interest in mechanistic explanations have clear starting points or set-
up conditions and clear endpoints or termination conditions.
Another important feature of mechanistic explanations is that they “account for the behavior
of a system in terms of the functions performed by its parts and the interactions between these
381
parts” (Bechtel and Richardson 1993/2000, 17), rather than in terms of general laws or theories.
In Descartes’ example, molecules, nerve fibers, pores, animal spirits, and muscles are all parts or entities
of a human organism. Tugging, opening, flowing, and moving are all activities in which these parts
or entities engage in the production of reflex behaviors. In Cajal’s anatomical work, axon col-
laterals, dendrites, dendritic spines, and growth cones are the entities. Sherrington, Hebb, and Kandel
later contributed knowledge about the activities of those entities.
William Bechtel and Robert Richardson (1993/2000) emphasize that a central heuristic
strategy operative in developing mechanistic explanations is the decomposition of the phenom-
enon into its component parts and their operations. Decomposing a system in this fashion and
explaining its behaviors mechanistically is not something that can be accomplished in a single
area of science. As the “New Mechanists” emphasize, it requires input from many different areas
of science. In the process, rather than one branch or area of science being reduced to another,
input from different areas of science is “integrated into descriptions of multi-level mechanisms”
(e.g., Craver 2007).
Cajal’s work is a prime example of the decomposition strategy at work. It was critical in
convincing anatomists of the “neuron doctrine,” which extended cell theory to neural tissues,
stating that the brain is made up of anatomically discrete cellular units. Cajal showed how neu-
rons of different types, such as the basket and Purkinje cells of the cerebellum, are connected
in organized patterns. He also worked to discover the anatomical properties of the neuron’s
sub-parts like dendrites, axon collaterals, growth cones, and dendritic spines. Cajal’s discover-
ies of the anatomical properties of neurons and their component parts, in combination with
Sherrington and Pavlov’s discoveries about how these parts function, shaped the development
of Hebb’s postulate, which informed Kandel’s work in developing simplified preparations that
decomposed reflex operations in Aplysia to a simple neuronal circuit and its component parts.
Bechtel and Richardson also note that the mechanistic explanatory strategy is often con-
strained by available technology and that scientists “will appeal analogically to the principles
they know to be operative in artificial contrivances as well as in natural systems that are already
understood” (1993/2000, 17) to provide mechanistic explanations. Descartes’ appeal to the
mechanical statues in St. Germain to explain reflex action and Pavlov’s appeal to a telephone
switchboard to explain how conditioned reflexes come about are clear examples of how the
available technology of a time period can shape how investigators conceive of a mechanism.
This raises another important feature of mechanistic explanations detailed by Darden (2002):
they are gradually discovered over time (see also Chapter 19). In terms of their empirical support,
candidate mechanisms can have the status of “how-possibly,” “how-plausibly,” or “how-actually”
explanations (Craver 2007). In terms of their completeness, mechanistic explanations start out
as sketches (MDC 2000) that have gaps in their productive continuity or black boxes left to be
filled in with detail. Sketches are revised, filled in, and fit into their surrounding contexts, until
they eventually gain the status of adequately complete mechanistic explanations, or are rejected
as false starts. The activities and sub-entities that mediate the connections between neurons
were black boxes for Cajal before Sherrington developed the concept of the synapse. While
Hebb put forward a “how-possibly” mechanism for permanent changes in communication
between neurons, this was not yet considered an adequate explanation. Later work by Kandel
and colleagues was directed at understanding “how-actually” such changes come about during
real learning events.
The process of discovery sometimes requires more substantial revisions to how the phenom-
enon was originally individuated and circumscribed (see Bechtel and Richardson 1993/2001;
Bechtel 2008; Craver 2007, 2009). Experimentation may result in discoveries that prompt
a revision to the original taxonomy of kinds of phenomena identified in a given field of
382
research. For example, it may be discovered that what was once considered one phenomenon
(e.g., memory) includes at least two forms (e.g., declarative and procedural). Experimentation
may reveal that we were looking in the wrong place for a mechanism, or that a single mecha-
nism performs what were originally thought to be two separate functions (e.g., see Eichenbaum
and Cohen (2014) on the hippocampus’s dual role in memory and navigation). It is well
known today that rather than being the conduit between the body and immaterial soul, the
pineal gland produces and secretes melatonin, which is involved in the modulation of circadian
rhythms in the vertebrate brain.
These historical vignettes as a whole demonstrate another key feature of mechanistic
explanations: they are multi-level. From Pavlov to Kandel, for example, we move from obser-
vations of mid-scale entities and activities like dogs, bells, and salivation, to micro-scale entities
and activities like ions and neurotransmitter release. Mechanistic explanations involve
entities and activities at multiple scales, some of which are sub-mechanisms that constitute
higher-level components. Even in Descartes’ description of the mechanism of the reflex, the
behavior of the whole organism is explained by appeal to some of its constituent parts and
their sub-parts, like nerves, nerve fibers, pores, and animal spirits.
A related and notable characteristic of mechanistic explanations is that they are not
byproducts of a single area of science. Rather, they rely for their development on information
emanating from multiple different areas of science that study entities and activities at varying
scales (compare Darden and Maull 1977; Craver 2007). Consider Descartes’ explanation of
the reflex—it combined a corpuscular theory of matter with a rudimentary understanding of
the anatomy of the nervous system prevalent in his day, and a theory of animal spirits originat-
ing with Galen. Although Pavlov thought that physiology could advance an understanding of
the mechanisms of conditioned reflexes without appeal to psychology, he recognized that it
could not do so in the absence of advances in anatomy and cell biology. Cajal’s histological
preparations could not reveal the functional nature of the connections between contiguous
neurons without the addition of physiological work, which Sherrington later contributed.
Kandel and colleagues’ research into the mechanisms of simple forms of non-associative
learning in Aplysia combines anatomical, electrophysiological, biochemical, behavioral, and
pharmacological techniques.
4. Discovering mechanisms: open philosophical problems

The last two features of mechanistic explanation we mentioned—their multi-level nature, and
the fact that they integrate results from various branches of science—are very much at odds with
traditional thinking about scientific explanation. As mentioned briefly, in the mid-twentieth
century, scientific phenomena at different scales, and the fields of science that study them, were
thought to be related to one another in terms of reduction. Chemistry, for instance, was sup-
posed to occupy itself with a circumscribed range of chemical phenomena, which the methods
of chemistry alone were appropriate for investigating. Furthermore, all of chemistry, it was
thought, would eventually prove to be reducible to physics in the way that heat is reducible
to the average kinetic energy of physical particles. Higher-level sciences, according to this way
of thinking, may serve pragmatic and heuristic purposes along the way to finding the fundamen-
tal theory, but eventually should turn out to be superfluous.
Mental phenomena have long posed a challenge to this picture; many philosophers
(and others) want to deny that the mind is reducible to more fundamental physical entities
and activities. The multi-level nature of mechanistic explanations is meant to provide an
alternative to reduction. All of the levels in a mechanism, from low to high, contribute
383
to it performing its function. Going down lower does not provide a more fundamental
understanding, even if it might provide finer grained details; in fact, scientists sometimes
purposely focus their investigations at higher levels, because that’s where the functions
they’re interested in are performed.
Many questions remain about how exactly this plays out in practice. Craver (2007) describes
a picture of “integrative unity” in which a psychological capacity, such as spatial memory, is
brought about by anatomically differentiated parts of the brain (area CA1 of the hippocampus),
its physiological component parts (neural networks, neurons, synapses), and activities (firing,
transmitter release), which in turn are composed of smaller-scale parts (receptors, molecules)
and their activities (activation, phosphorylation). If Craver is right, we should be able to fit the
results from our historical vignettes into a hierarchy of mechanisms with Pavlov’s conditioned
reflexes at the top, Hebb’s associative synaptic mechanisms slightly below, Cajal’s anatomical
picture of the neuron and Sherrington’s physiological insights into synapses another step down,
then finally Kandel’s molecular mechanisms of learning at the bottom.
Some of the entities and activities involved do fit together as parts to wholes, such as Kandel’s
molecular mechanisms, which describe parts of Cajal and Sherrington’s neurons and synapses.
However, it is not clear that the levels will always connect in such a tidy way, especially at the
higher levels. Craver’s account seems to presuppose that psychological and neural mechanisms
are part of the same ontological hierarchy, yet psychological mechanisms do not necessarily have
neural mechanisms as parts (Stinson 2016).
Consider, for example, an information-processing mechanism that explains how an organ-
ism learns to respond to stimuli like burning flames or noxious shocks. That mechanism needs
to store the relationship between stimulus and response in some memory medium. Reflexes
mediated by nerve fibers, as Descartes imagined, can’t do the whole job, because we can learn
not only to pull our foot away from a flame, but also to do many other things with our limbs
in response to many other kinds of stimuli. The nerve fiber doesn’t have enough bandwidth
to represent all of these learned relationships. This notion of bandwidth is an abstract concept
that doesn’t appeal specifically to any parts of the stimulus-response system, and yet it provides a
psychological-level, mechanistic explanation of why Descartes’ fibers can’t be the whole story.
Another issue is that what look like the natural boundaries of a phenomenon from the
perspective of one science (including start and finish conditions, and the way components are
picked out) might not match up with what look like the natural boundaries of the same phe-
nomenon from the perspective of another science (see Chapter 9). In Stinson (2016) one of
us argues that the science of memory has this problem. From the perspective of psychology, it
seems clear that memory encoding, storage, and recall are distinct processes, for example. Yet
from the perspective of neuroscience, there do not appear to be clear distinctions between these
memory processes. When you try to integrate the mechanisms of memory studied in these two
sciences, you do not find a neat relationship where neural mechanisms turn out to be the parts of
psychological mechanisms. At the neural level, encoding, storage, and recall are all intertwined,
so neural mechanisms of memory don’t turn out to be related to psychological mechanisms of
memory as parts to wholes.
Thus, rather than different areas of science like psychology and neuroscience being seam-
lessly integrated into unified mechanistic explanations, we often find explanations that cross
levels in the mind-brain sciences to be messy and partial. Craver illustrates his mosaic unity with
images like the one on the left of Figure 28.4. Instead, we suggest that the inter-field relation-
ships we should expect will look more like the more complex image on the right.
Scientists working in different fields conceive of their phenomena of interest in different
ways, experimentally investigate phenomena in different ways, ask different research questions, use
384
Figure 28.4 A comparison of Craver’s (2007) view of inter-level relations [left] between mechanisms,
and Stinson’s [right]. Copyright 2016 Catherine Stinson. Used with permission
different methods aimed at vastly different scales, and investigate these phenomena in different
species. For example, psychologists have historically been characterized as interested in pro-
viding explanations of cognitive capacities by functional analysis (e.g., Cummins 1983; Fodor
1968). Many psychologists believe this requires a clear specification and decomposition of abstract
cognitive processes (like learning) involved in psychological tasks. When tasks are regarded as inap-
propriate for individuating a discrete function, they are often refined. However, neurobiologists
investigating learning in invertebrates (like Kandel) and physiologists who investigate learning in
non-human mammals (rodents, dogs (like Pavlov)) are often not interested in individuating the
cognitive processes engaged during training in learning paradigms. While not worrying about
abstract cognitive processes makes good sense in the case of Aplysia, which has a simple nervous
system and can be studied using reduced preparations, ignoring the component cognitive processes
that may be involved when rodents are trained in learning paradigms will render the connec-
tions investigators would like to make between cellular and molecular mechanisms and cognitive
capacities tenuous at best (see Sullivan 2009, 2010, 2016).
A related problem with this unity picture is the issue of comparing mechanisms across species.
Kandel’s sea slugs are vastly different from humans, yet it is assumed that the results of experi-
ments undertaken in one species are generalizable to others. Pavlov and Kandel are interested in
mechanisms of human learning, but perform their experiments on canines and invertebrates. For
both ethical and practical reasons, the systems scientists have historically and continue to use are
model organisms like dogs, frogs, birds, rodents, sea slugs, and fruit flies. While certain cellular
and molecular mechanisms are conserved across species, there are obvious differences between
sea slugs and humans that prohibit direct inference from one to the other.
Although neuropsychological research on patients with localized brain damage and fMRI
experiments involving human beings have shed some light on the loci of specific types of
learning and memory in the human brain, we continue to lack a mechanistic understand-
ing of human learning; the explanations we currently have are patchy at best. It is supposed
that advances in imaging technologies will eventually enable a visualization of the loci and
mechanisms of human learning. However, before such discoveries are to be feasible, sci-
entists require better methods for individuating learning phenomena in human beings and
non-human mammals. It is not simply that scientists lack the available imaging technologies;
it is that many experiments in the cognitive neurosciences lack the rigor of work with model
385
organisms that have smaller repertoires of behaviors. It is more difficult to design tasks that
tease apart discrete kinds of learning in human beings than in Aplysia. One reason for this
difficulty is that human beings might use multiple strategies to perform a cognitive task, and
it’s not always possible to predict the range of strategies that might be used, or to detect
whether subjects are using the expected one.
Despite these difficulties in drawing connections between experimental findings using pro-
tocols from different fields, and in phylogenetically distant species, it is necessary for mechanistic
explanations in neuroscience to find ways of bridging these gaps. As we mentioned earlier,
Cajal’s histological experiments could only go so far. He was able to get a fairly accurate picture
of the anatomy of the neuron, but the structure alone couldn’t reveal how neurons communi-
cate. Pavlov could only figure out the functional characteristics of reinforcement learning using
his experimental methods. His functional picture could not reveal what sorts of structures might
give rise to reinforcement learning. As a general rule, neither bottom-up (from structure to
function) nor top-down (from function to structure) methods in isolation can get us all the way
to understanding mechanisms. Instead what is needed is a multi-level approach, with research-
ers simultaneously using many strategies to investigate different phenomena, alongside some
efforts at linking the results of these together, i.e. something very much like how neuroscientific
research is in fact pursued.
Scientists approach the problem of understanding the brain at various levels because there are
robust regularities at various levels, both in neuroscience and in the life sciences more gener-
ally. There are some phenomena that we feel compelled to think of as real or natural kinds, like
molecules, cells, organs, organisms, and species, even when we can’t give them tidy definitions
in terms of their component parts. We think that neurons are a genuine kind of thing despite
the fact that (contra Cajal) nerve cells sometimes do fuse together in ways that challenge their
anatomical and physiological independence. Organisms often end up in symbiotic relationships
with other organisms, like our gut microbiota, without which we couldn’t live, challenging the
independence of organisms. The action potential depends on a membrane, ion channels, and
extra and intra-cellular ions, but it only exists within a narrow range of conditions.
A complex biological system like the brain will likely prove impossible to fit into a neat
hierarchy of nested parts, because the borders of mechanisms are fuzzy. This does not mean
that we can’t ever have an integrated science that links together different levels. There are
connections to be made between the results from different experimental paradigms, experi-
ments on different species, models of different phenomena, and different models of the same
phenomenon. Many of these connections will be partial, and the integrated picture will be
patchy (see Schaffner 2006; Stinson 2016).
5. Conclusion
The historical case studies we considered span several centuries, but a common aim in each
case was discovering mechanisms. Constructing multi-level mechanistic explanations involves
intensive collaboration across different branches of science, and involves many challenges,
both pragmatic and methodological. Available technologies, training in experimental meth-
ods, choice of model organisms, levels of investigation, and inter-field collaborators all can
either ensure success or act as barriers to progress. Integrating the discoveries from various
fields where the phenomena are circumscribed in different ways requires piecing together
results in complex ways, and carefully considering when and how results can be generalized
to different contexts.
386
Note
1 The authors would like to thank Stuart Glennan and Phyllis Illari for very helpful comments on an
earlier draft of this chapter. Co-authors had equivalent input and are listed in alphabetical order.
References
Bechtel, W. (2008). Mental Mechanisms: Philosophical Perspectives on Cognitive Neuroscience. New York:
Taylor and Francis.
Bechtel, W. and Richardson, R. (1993/2000). Discovering Complexity: Decomposition and Localization as
Strategies in Scientific Research. Princeton, NJ: Princeton University Press.
Bennett, M. R. (1999). The early history of the synapse: From Plato to Sherrington. Brain Research Bulletin,
50(2), 95–118.
Craver, C. (2007). Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience. Oxford: Oxford
University Press.
Craver, C. (2009). Mechanisms and natural kinds. Philosophical Psychology, 22(5), 575–594.
Craver, C. and Darden, L. (2001). Discovering mechanisms in neurobiology: The case of spatial mem-
ory, in P.K. Machamer, R. Grush, and P. McLaughlin (eds.), Theory and Method in the Neurosciences
(pp. 112–137). Pittsburgh: University of Pittsburgh Press.
Cummins, Robert. (1983). The Nature of Psychological Explanation. Cambridge, MA: MIT Press.
Darden, L. (2002). Strategies for discovering mechanisms: Schema instantiation, modular subassembly,
forward/backward chaining. Philosophy of Science, 69(3), S354–S365.
Darden, L. and Maull, N. (1977). Interfield theories. Philosophy of Science, 43, 44–64.
Descartes, R. (1664/1985). Treatise on Man in The Philosophical Writings of Descartes Volume I, translated
by John Cottingham, Robert Stoothoff, and Dugald Murdoch. Cambridge: Cambridge University
Press.
Eichenbaum, H. and Cohen, N. (2014). Can we reconcile the declarative memory and spatial navigation
views on hippocampal function? Neuron, 83(4):764–770.
Fodor, J. (1968). Psychological Explanation: An Introduction to the Philosophy of Psychology. New York:
Random House.
Forel, A. (1887/1991) Out of my life and work. In G. M. Shepherd (Ed.), Foundations of the Neuron Doctrine
(pp. 115–116). New York: Oxford University Press.
Foster, M. with Sherrington, C. S. (1897). A Textbook of Physiology, Part Three: The Central Nervous System,
7th ed. London: Macmillan and Co. Ltd.
Glennan, S. (1996). “Mechanisms and the nature of causation.” Erkenntnis, 44: 49–71.
Golgi, C. (1873/1991). On the structure of the gray matter of the brain. Foundations of the Neuron Doctrine,
Gordon M. Shepherd (Trans.) (pp. 84–88). Oxford: Oxford University Press.
Hebb, D. (1949/2002). The Organization of Behavior: A Neuropsychological Theory. Mahwah, NJ: Lawrence
Erlbaum.
Hilgard, E. R. (1956). Theories of Learning, 2nd ed. New York: Appleton-Century-Crofts.
His, W. (1886/1991). On the structure of the human spinal cord and nerve roots. In G. M. Shepherd (Ed.),
Foundations of the Neuron Doctrine (pp. 106–110). New York: Oxford University Press.
Illari, P. and Williamson, J. (2012). What is a mechanism? Thinking about mechanisms across the sciences.
European Journal for the Philosophy of Science, 2(1), 119–135.
Kandel, E. R. and W. A. Spencer. (1968). Cellular neurophysiological approaches in the study of learning.
Physiological Review, 48(1), 65–134.
Kandel, E. R., Schwartz, J. H., Jessell, T. M., Siegelbaum, S. A., and Hudspeth, A. J. (Eds.). (2012).
Principles of Neural Science (Vol. 5). New York: McGraw-Hill.
Nagel, E. (1961). The Structure of Science: Problems in the Logic of Scientific Explanation. London: Harcourt,
Brace & World.
Oppenheim, Paul and Putnam, Hilary. (1958). The unity of science as a working hypothesis. In Herbert
Feigl, Grover Maxwell, and Michael Scriven (Eds.) Minnesota Studies in the Philosophy of Science
(pp. 3–36). Minneapolis: Minnesota University Press.
Pavlov, I. (1927/60). Conditioned Reflexes. Mineola, NY: Dover Publications.
Machamer, P., Darden, L., and Craver, C. (2000). Thinking about mechanisms. Philosophy of Science, 67(1),
1–25.
387
Ramón y Cajal, S. (1888/1991). Structure of the nervous system of birds. In G. M. Shepherd (Ed.),
Foundations of the Neuron Doctrine (pp. 141–148). New York: Oxford University Press.
Ramón y Cajal, S. (1890/1988). On the structure of the cerebral cortex of certain mammals. In J. DeFelipe
and E. G. Jones (Eds.), Cajal on the Cerebral Cortex: An Annotated Translation of the Complete Writings
(pp. 23–54). New York: Oxford University Press.
Ramón y Cajal, S. (1894a/1991). The fine structure of the nervous centers (Croonian lecture). In
G. M. Shepherd (Ed.), Foundations of the Neuron Doctrine (pp. 239–253). New York: Oxford University
Press.
Ramón y Cajal, S. (1894b/1990). New Ideas on the Structure of the Nervous System in Man and Vertebrates.
Neely Swanson and Larry W. Swanson (Trans.). Cambridge, MA: The MIT Press.
Schaffner, K. F. (2006). Reduction: The Cheshire cat problem and a return to roots. Synthese, 151(3),
377–402.
Shepherd, G. M. (1991). Foundations of the Neuron Doctrine. New York: Oxford University Press.
Stinson, Catherine. (2016). Mechanisms in neuroscience: Ripping nature at its seams. Synthese, 193, 1585–1614.
Sullivan, Jacqueline. (2009). The multiplicity of experimental protocols: A challenge to reductionist and
nonreductionist models of the unity of science. Synthese, 167, 511–539.
Sullivan, Jacqueline. (2010). Reconsidering spatial memory and the Morris water maze. Synthese, 177(2),
261–283.
Sullivan, Jacqueline. (2016). Construct Stabilization and the unity of the mind-brain sciences. Philosophy
of Science, 83, 662–673.
388
29
MECHANISMS IN
COGNITIVE SCIENCE1
Carlos Zednik
1. Introduction
A principal goal of cognitive scientific research is to explain behavioral and cognitive phenomena
such as perception, action, categorization, memory, learning, language, and attention. The most
influential account of explanation in cognitive science is due to David Marr (1982). On Marr’s
account, cognitive scientists answer questions at three distinct levels of analysis: the computational
level, which concerns questions about a particular system’s computational goals and about the
appropriateness thereof; the algorithmic level, which is driven by questions about the representations
and algorithms that are used by the system to achieve these goals; and the implementational level,
which addresses questions about the way in which these representations and algorithms are physi-
cally realized (Marr 1982: 24ff). According to Marr, questions at all three levels of analysis must
be answered to “completely understand” a cognitive system, and to thereby explain its behavior.2
Although Marr’s account remains influential to this day, there are reasons to be unsatisfied
about its clarity and scope. First, although it is relatively clear which questions should be asked,
it is not quite as clear how these questions might be answered. Among others, it is unclear
what it actually takes to show that a computational goal is appropriate, and what it means for
an algorithm to be physically realized. Second, because Marr’s account is formulated in terms
co-opted from computer science, rather than in terms endemic to philosophical discussions of
scientific explanation, it is difficult to know how explanations in cognitive science compare to
explanations in other disciplines that center on, for example, subsumption under law, the devel-
opment of models, and/or the discovery of mechanisms. Finally, many of the terms that play a
central role in Marr’s account are far less prominent in cognitive science today than they were
at the time of Marr’s writing. In particular, the computationalist research program which pre-
dominated in the 1970s and 1980s (Pylyshyn 1980) now competes for attention and resources
with alternative research programs such as connectionism (Rumelhart et al. 1986), dynamicism (van
Gelder 1998), and the Bayesian approach (Zednik & Jäkel 2016)—some of which may not rely
on notions of “computation,” “algorithm,” and “implementation” at all.
Despite these reasons to be unsatisfied with Marr’s account of explanation in cognitive
science, its lasting influence recommends it as a productive starting point for discussion.
Indeed, the aim of this chapter is to show that many of the ambiguities in Marr’s account
can be resolved, and that its scope can be extended, by considering Marr to have been an early
389
Carlos Zednik
advocate of mechanistic explanation (Chapter 1). Although Marr’s account was originally
designed to capture explanations in computationalist cognitive science, the questions it
identifies at each level of analysis are in fact variations on the types of questions that are
asked in any research program that aims to discover and describe (cognitive) mechanisms.
Specifically, questions at the computational level can be construed as questions about what a
mechanism is doing and why: questions that concern the mechanism’s behavior and contain-
ing environment. Questions at the algorithmic level, in contrast, are questions about how a
mechanism does what it does, and concern its component operations. Finally, questions at
the implementational level of analysis can be understood as questions about where a particular
mechanism’s component operations are carried out—that is, questions about the component
parts in which such operations might be localized. Although Marr showed how each one
of these questions might be answered using the concepts and methods of computationalist
cognitive science, they are also answered in the context of other research programs such as
connectionism, dynamicism, and the Bayesian approach. In other words, there is reason to
believe that all of these research programs seek three-level explanations, and moreover, that
all of them aspire to discover and describe mechanisms.
This chapter will outline and defend a mechanistic interpretation of Marr’s account of expla-
nation in cognitive science, and thereby attempt to resolve the ambiguities that remain in this
account, as well as to extend its scope. Notably, in line with Marr’s claim that all three levels
of analysis are necessary to “completely understand” a cognitive system, on the current inter-
pretation all three levels must be addressed to provide mechanistic explanations of behavioral
or cognitive phenomena.3 In this way, the present interpretation differs from several previous
attempts to relate Marr’s account to the framework of mechanistic explanation (see e.g. Bechtel &
Shagrir 2015; Bickle 2015; Kaplan 2011; Milkowski 2013; Piccinini & Craver 2011). Moreover,
the present interpretation helps to address a number of philosophical debates concerning e.g. the
role of idealization in cognitive modeling, the role of abstraction in mechanistic explanation, and
the nature of the realization-relationship that obtains between functional processes in the mind
and physical structures in the brain.
2. The computational level: “what?” and “why?”

In Marr’s original formulation, the computational level of analysis is defined by questions
about a cognitive system’s computational goals, as well as questions about the appropriateness
thereof (Marr 1982: 24ff). More generally, these can be understood as questions about what a
cognitive system is doing, and questions about why it does what it does (see also McClamrock
1991; Shagrir 2010).
What-questions can be answered by describing the relevant system’s behavior. Within the
computationalist research program, this involves specifying an information-processing function
that maps the system’s (sensory) inputs onto its (behavioral) outputs. That said, recent contributions
by e.g. dynamicist researchers suggest that many behavioral and cognitive capacities—especially
those that depend on continuous feedback from the environment—cannot be easily described as
mappings between input and output (van Gelder 1995). Rather, these capacities are better described
as continuous trajectories through state space, the dimensions of which might correspond to neural
activity, a system’s bodily position or motion, and/or features of the environment (Kelso 1995).
Although it may be terminologically confusing to associate dynamical state-space trajectories with
the “computational” level of analysis, they are analogous to information-processing functions in
that they too can be used to describe a cognitive system’s behavior, and thus to answer questions
about what the system is doing.
390
Mechanisms in cognitive science
Answers to what-questions play an important role not only in Marr’s account of explanation
in cognitive science, but also in the framework of mechanistic explanation (see also Chapter 16).
Descriptions of a cognitive system’s behavior are descriptions of an explanandum phenomenon
(Cummins 2000). In many scientific disciplines, such descriptions are the starting point of mech-
anistic explanation: The explanandum phenomenon is identified with the overall behavior of a
mechanism, and an attempt is made to describe that mechanism’s component parts, operations,
and overall organization (Bechtel & Richardson 1993; Darden & Craver 2012; Chapter 19,
this volume). Notably, many mechanisms are known whose overall behavior has been described
as a form of information-processing (Craver 2013; Kaplan 2011), but also mechanisms whose
behavior has been more effectively described as a trajectory through state space (Bechtel &
Abrahamsen 2010). Therefore, these ways of answering what-questions in cognitive science are
by no means inconsistent with the principles of mechanistic explanation (compare Chemero &
Silberstein 2008). That said, merely describing a mechanism’s overall behavior is insufficient
for the purposes of mechanistic explanation (Kaplan & Craver 2011; Chapter 20, this volume).
Indeed, insofar as descriptions of a mechanism’s overall behavior often resemble law-like regu-
larities (Chapter 12), they may also feature in other kinds of scientific explanation (Chemero &
Silberstein 2008; Zednik 2011). For this reason, determining whether cognitive scientists are in
fact in the business of mechanistic explanation requires taking a closer look at the way in which
they answer questions beyond “what?”
One such question is the question of “why?” In the context of his celebrated cash register
example, Marr deems it important to ask “why the cash register performs addition and not,
for instance, multiplication when combining the prices of the purchased items to arrive at a
final bill” (Marr 1982: 22). Marr claims that such questions are answered by considering the
“appropriateness” of the system’s behavior with respect to the “task at hand” (Marr 1982:
24). Whereas many previous discussions of Marr overlook this aspect of the computational
level, Shagrir (2010) has provided a compelling analysis according to which a cognitive system’s
behavior should be deemed appropriate when a mathematical description of that behavior can
be mapped onto a relevant (potentially abstract or even counterfactual) property of the system’s
environment. One analytic technique that is well suited for answering why-questions in this
way is rational analysis (Anderson 1991; Oaksford & Chater 2007), which lies at the heart of
the recently influential Bayesian approach in cognitive science (Zednik & Jäkel 2016). This
technique involves formally characterizing a cognitive system’s task environment as a form of
probabilistic inference, and deriving an optimal solution in the sense prescribed by probability
theory. A widely reported—and initially surprising—finding of this approach is that many dif-
ferent kinds of cognition and behavior closely approximate optimal solutions (Pouget et al.
2013). Whenever this is the case, why-questions can be answered because the optimal solutions
describe a cognitive system’s behavior while simultaneously reflecting a mathematical property
of the system’s environment—namely, the Bayes-optimal solution to a particular task within
that environment. Intuitively, the method of rational analysis allows researchers to show that
cognitive systems behave as they do because that way of behaving is optimal in the sense pre-
scribed by probability theory (but compare Danks 2008).
Shagrir’s analysis of “appropriateness” implies that answers to why-questions cannot be found
by looking solely at the cognitive system whose behavior is being explained; it also involves
looking at the environment in which that behavior unfolds. It is far from obvious how looking
at properties external to a cognitive system might be conducive to revealing the mechanisms
internal to it. Moreover, it is tempting to think of why-questions as pertaining to a teleologi-
cal approach to scientific explanation which is traditionally contrasted with the mechanistic
approach.4 Nevertheless, philosophical proponents of mechanistic explanation have recently
391
Carlos Zednik
argued that environmental and teleological considerations play a significant role in mechanis-
tic explanation. In particular, Carl Craver argues that understanding the role a mechanism is
supposed to play in a containing environment can greatly facilitate the task of describing that
mechanism’s actual behavior, and that many investigators for this reason “search for a higher-
level mechanism within which it has a role” (Craver 2013: 153). Similarly, William Bechtel
(2009) argues that mechanistic explanation involves “looking up” at the environment in which
a mechanism is naturally embedded, because doing so greatly facilitates the task of characterizing
the mechanism’s actual behavior: It might reveal complexities that the mechanism must accom-
modate, as well as regularities that it can exploit.
These considerations suggest that answers to why-questions do in fact play a role in mecha-
nistic explanation (see also Chapter 8), and indeed, that they may be instrumental for coming up
with answers to what-questions. Notably, the Bayesian approach illustrates this kind of depend-
ence of the what on the why. Recall that many different kinds of behavior and cognition have
been found to approximate optimal solutions within a particular task environment. Although
initially surprising, this finding is not accidental. Investigators regularly find discrepancies
between an optimal solution and the observed behavioral data, but then often go on to tweak
the specification of the task environment until the optimal solution is closely approximated by
the data (Anderson 1991). Although some commentators denounce this kind of tweaking as a
post-hoc model-fitting exercise of limited explanatory value (see e.g. Bowers & Davis 2012),
others take it to be an efficient means of deriving mathematical formalisms that simultaneously
answer what- and why-questions at the computational level of analysis, and that even facilitate
the search for answers at lower levels (Zednik & Jäkel 2016).
In short, Marr’s computational level of analysis concerns two types of questions: questions
about what a cognitive system is doing, and questions about why. It is a mistake to consider
either kind of question to be antithetical to the principles of mechanistic explanation. On
the contrary, both kinds of questions play an important role in the discovery and description
of mechanisms in several different research programs. But although the computational level
of analysis can therefore be seen to play a critical role in mechanistic explanation, it is by no
means sufficient. Just as Marr deems it necessary to answer questions below the computational
level for the purposes of “completely understanding” a cognitive system and its behavior,
mechanistic explanation also involves describing the internal features of a mechanism—its
component parts, operations, and organization.
3. The algorithmic level: “how?”

Marr’s algorithmic level of analysis is defined by questions about “the representation for the
input and output,” and about the “algorithm by which the transformation [between them] may
actually be accomplished” (Marr 1982: 23). Questions of this kind are traditionally answered
through cognitive modeling, which involves functionally analyzing the complex behavioral or cog-
nitive capacity being explained into an organized collection of simpler capacities (Cummins
1983), and subsequently describing this collection in formal mathematical or computational
terms (Busemeyer & Diederich 2010; Luce 1995). In the computationalist research program,
cognitive models often consist of lists of production rules with which to manipulate symbolic
expressions (Pylyshyn 1980). These models are quite naturally viewed as descriptions of algo-
rithms for transforming representations so as to achieve a particular computational goal. That
said, cognitive models are a commonplace even in non-computationalist research programs.
Indeed, some of the most influential cognitive models today consist of mathematical equations
that determine the numerical values of output variables (Rumelhart et al. 1986; Nosofsky 1986),
392
the evolution of state variables over time (Busemeyer & Townsend 1993), or the probability
distribution over a hypothesis space (Pouget et al. 2013). It may not always be useful or even
possible to think of these models as describing representation-transforming algorithms (Ramsey
2007; van Gelder 1995). Nevertheless, insofar as they can be used to reproduce a particular
input-output transformation or a series of state-changes, they can still be said to compute a
particular information-processing function or state-space trajectory. Insofar as the function or
trajectory being computed accurately describes a particular cognitive system’s behavior, the
relevant cognitive model is a possible answer to a question about how that cognitive system does
what it does (see also Cummins 2000; McClamrock 1991).
Notably, there are often many different ways to compute a particular information-processing
function or state-space trajectory. For this reason, many different cognitive models may be
developed to answer a particular how-question, and investigators need a way of distinguishing
good answers from bad ones. Interestingly, Robert Cummins is sometimes interpreted as being
unwilling or unable to make such a distinction. Consider the following oft-quoted passage:
Any way of interpreting the transactions causally mediating the input-output connec-
tion as steps in a program for doing ϕ will, provided it is systematic and not ad hoc,
make the capacity to ϕ intelligible. Alternative interpretations, provided they are possi-
ble, are not competitors; the availability of one in no way undermines the explanatory
force of another.
(Cummins 1983: 43)
Because Cummins attributes the same degree of explanatory force to a (potentially) wide variety
of models, he has been accused of being unable to distinguish between answers to how-
questions that capture the way a cognitive system actually does what it does and answers that
merely capture the way it might possibly do so (Kaplan 2011; Piccinini & Craver 2011). This
accusation seems unwarranted, however; it fails to acknowledge Cummins’ demand that cogni-
tive models accurately reflect the “transactions causally mediating the input-output connection.”
Moreover, Cummins goes on to acknowledge that the elements of some cognitive models are
more likely than others to be instantiated in a particular cognitive system—and that these models
are for this reason to be preferred (Cummins 1983: 44; see also Feest 2003). Thus, Cummins
does in fact outline a criterion for distinguishing good answers to how-questions from bad ones:
Good answers are provided by cognitive models whose elements are in fact instantiated by the
cognitive system being investigated, and that actually reflect the causally relevant factors that
contribute to that system’s behavior.
This criterion can be fleshed out by aligning it with the idea that cognitive models should
accurately describe the internal features of the mechanism responsible for the explanandum
phenomenon: its component parts, operations, and overall organization. More precisely,
cognitive models should satisfy what has come to be known as the model-to-mechanism
mapping constraint (3M):
In successful explanatory models . . . (a) the variables in the model correspond

to components, activities, properties, and organizational features of the target
mechanism that produces, maintains, or underlies the phenomenon, and (b) the
(perhaps mathematical) dependencies posited among these variables in the model
correspond to the (perhaps quantifiable) causal relations among the components
of the target mechanism.
(Kaplan & Craver 2011: 611, see also Chapter 17, this volume)
393
Carlos Zednik
3M allows for the possibility that many different models have equal explanatory force—as long
as the rules, symbols, equations, and variables posited by each one of these models correspond
to the features of the target mechanism. Moreover, although 3M requires that the features of a
mechanism be described correctly, it does not require that they be described completely. In this
spirit, Piccinini and Craver (2011) argue that cognitive models are designed to provide mecha-
nism sketches, elliptical descriptions that correctly, albeit incompletely, describe a mechanism’s
component parts, operations, and/or organization. Although mechanism sketches satisfy 3M by
correctly describing some of a mechanism’s features, additional “filling in” is required to trans-
form these sketches into full-fledged mechanistic explanations (Craver 2007; Machamer et al.
2000, see also section 4 of this chapter).
Several commentators have argued that cognitive models can in fact be viewed as mechanism
sketches. To this end, they have considered examples from a variety of research programs (for
discussion see e.g. Abrahamsen & Bechtel 2006; Milkowski 2013; Zednik 2011). That said,
other commentators have identified counterexamples in the form of models that contain ideali-
zations: constructs that do not correspond to any one of a particular mechanism’s features (see
also Chapter 17). For example, Weiskopf (2011) introduces Hummel and Biederman’s (1992)
model of object-recognition, which contains Fast Enabling Links (FELs) that “possess physi-
cally impossible characteristics such as infinite speed” (Weiskopf 2011: 331). Similarly, Buckner
(2015) considers the role of backpropagation learning in connectionist networks, the execution
of which depends on the biologically implausible “backwards transmission of information across
neural synapses, the need for prior knowledge of correct output, and the distinct, individualized
error signals used to adjust the thresholds and weights of each node and link in the network”
(Buckner 2015: 3925). Because FELs and backpropagation learning are psychologically, biologi-
cally, and/or physically implausible, cognitive models that incorporate these constructs can be
known with relative certainty to not satisfy 3M.
Although there are counterexamples to the claim that cognitive models generally satisfy 3M,
there is nevertheless reason to believe that these models are constrained by 3M. That is, there is
reason to believe that FELs and backpropagation learning are subject to replacement by less ide-
alized constructs as the explanatory demands on the relevant models increase. Notably, Weiskopf
denies this claim, arguing that FELs are “not clearly intended to be eliminated by any better
construct in later iterations” (Weiskopf 2011: 331). The key premise in Weiskopf’s argument is
that FELs allow Hummel and Biederman to understand the unique contribution of synchronous
firing to object-recognition, independent of other factors such as mutual excitation and/or inhi-
bition. However, it is important not to conflate intelligibility with explanation: Although FELs
may allow investigators to understand the unique contribution of one causally relevant factor,
the use of such idealizations often comes at the cost of obscuring or altogether neglecting the
contribution of other factors. Thus, although models that contain idealizations may reproduce
gross behavioral trends, they are unlikely to capture precise quantitative detail such as reaction-
times and learning curves that may be of significant explanatory interest. Indeed, as Buckner
goes on to argue, it is in order to capture just this kind of detail that connectionist researchers
often seek to replace backpropagation learning with “more biologically plausible training rules”
(Buckner 2015: 3925). In general, as the explanatory demands on a particular model increase,
it seems likely that idealizations will eventually be replaced by constructs that more accurately
reflect the causal structure of a mechanism. Even if the cognitive models being used today do
not satisfy 3M, future iterations of these models will presumably strive to do so. For this reason,
the requirement that cognitive models correctly describe mechanisms in the sense of 3M is use-
ful for determining whether these models provide good or bad answers to questions about how
cognitive systems do what they do.
394
4. The implementational level: “where?”

How-questions at the algorithmic level of analysis are often the primary concern of research pro-
grams in cognitive science. Nevertheless, it would be a mistake to think that an explanation has
been provided just as soon as these questions have been answered. As has already been stressed
repeatedly, on Marr’s account all three levels of analysis are needed to “completely understand”
a cognitive system and explain its behavior. Accordingly, investigations at the computational
and algorithmic levels must be supplemented by investigations at the implementational level,
which are driven by questions about the way in which the constructs of a cognitive model are
“realized physically” (Marr 1982: 25; see also Polger 2004). Unfortunately, it remains unclear
what it actually takes to show that the production rules, symbols, equations, and/or variables
specified by a cognitive model are realized by complex and potentially unruly physical systems
such as the brain. The aim of this section is to show that this lack of clarity can be remedied by
considering what it takes to “fill in” a sketchy cognitive model so as to deliver a full-fledged
mechanistic explanation.5
In general, “filling in” is a matter of describing a mechanism in greater detail than before
(Craver 2007; Machamer et al. 2000). Although the present discussion embraces the view that
cognitive models are mechanism sketches which leave out certain details, it has not yet been
discussed exactly what kinds of detail are typically left out. Because most cognitive models
specify mathematical constructs such as lists of production rules or systems of equations, it is
tempting to think of them as specifying details about a mechanism’s abstract mathematical or
computational properties, rather than about its concrete physical properties. Indeed, this view
is quite widespread in the literature, despite often being left implicit (see e.g. Bechtel 2008;
Bechtel & Shagrir 2015; Chemero & Silberstein 2008; Craver 2007; Danks 2008; McClamrock
1991; Piccinini & Craver 2011; Shagrir 2010; Stinson 2016). Notably, on this view the details
that are included may pertain to any one or more of a mechanism’s internal features. For exam-
ple, on this view cognitive models may describe a mechanism’s component operations in terms
of e.g. their informational properties, rather than in terms of neuronal spike trains (Pylyshyn
1980; Weiskopf 2011). They may also specify a mechanism’s component parts as e.g. filters,
without identifying these with any particular neural structures (Stinson 2016). Finally, they
might characterize a mechanism’s overall organization as a graph, abstracting over anatomical
detail (Bechtel & Shagrir 2015; Levy & Bechtel 2013).
Thus understood, the realization-relationship that obtains between the algorithmic and
implementational levels is one of instantiation; “filling in” involves specifying the concrete physi-
cal properties that instantiate a particular set of abstract mathematical properties. Thus, whereas
investigations at both levels describe the same component parts, operations, and/or organization
of a particular mechanism, they differ with respect to the kinds of properties being described:
Abstract mathematical properties on the one hand, and concrete physical properties on the
other. Put differently, although the questions being asked at each level are fundamentally the
same—they are how-questions in each case—the answers being given differ because they are
articulated at different levels of abstraction.
Although this view is relatively widespread in the literature, it faces an important challenge:
Why should it be necessary to answer a how-question twice? Since a mechanism’s abstract
mathematical properties are instantiated by its concrete physical properties, why should the
former need to be cited in an explanation, in addition to the latter? Polger (2004) argues that
mechanisms can compute information-processing functions or state-space trajectories just in
virtue of their physical properties, regardless of whether these properties are also said to instanti-
ate any particular abstract mathematical properties. In the same vein, Bickle (2015) advances a
395
Carlos Zednik
conception of mechanistic explanation in which a characterization of a mechanism’s physical

properties suffices to explain a wide variety of behavioral and cognitive phenomena; whether
or not a cognitive mechanism can also be characterized in abstract mathematical terms at the
algorithmic level is explanatorily irrelevant.
There is reason to be wary of this conclusion, however. For one, it contradicts Marr’s highly
influential claim that all three levels of analysis are needed to “completely understand” a cogni-
tive system’s behavior. For another, it calls into question the explanatory relevance of cognitive
modeling, a practice that is widely considered to be the centerpiece of cognitive scientific
research (see statements to this effect by e.g. Busemeyer & Diederich 2010; Cummins 1983,
2000; Luce 1995; Stinson 2016; Weiskopf 2011). Unfortunately, several previous attempts to
avoid this conclusion fall short. Consider, for example, Bechtel and Shagrir’s recent claim that
cognitive models facilitate the identification of design principles that “produce the same results
across a wide range of different implementations” (Bechtel & Shagrir 2015: 318). Although
Bechtel and Shagrir show that the description of design principles can render a particular mech-
anism’s organization intelligible and generalizable (see also Levy & Bechtel 2013), they do little
to argue that these principles actually produce anything over and above the physical properties
in which they are instantiated. In other words, design principles are subject to Kim’s (1993)
causal exclusion argument, according to which realized properties possess no causal powers over
and above their realizers. If the abstract design principles identified at the algorithmic level
cannot be thought to possess causal powers over and above the concrete properties that instan-
tiate them, it is unclear why cognitive models are needed to answer questions about how a
mechanism does what it does. A similarly unsuccessful response is given by Craver, who argues
that the algorithmic level describes “realized properties [that] figure in unique causal relevance
relations”—relations that, on his manipulationist account of constitutive relevance, “are true of
realized properties and are not true of their realizers” (Craver 2007: 220). Recent challenges to
Craver’s manipulationist account suggest that interventions always affect realized and realizing
properties simultaneously (Baumgartner & Gebharter 2015), thereby calling into question the
claim that there can actually be any such unique causal relevance relations. Thus, the explana-
tory relevance of the algorithmic level remains unclear.
These unsuccessful responses to the challenge of explanatory irrelevance are weighed down
by the view of cognitive models as descriptions of a mechanism’s abstract mathematical proper-
ties. But there is an alternative view that has yet to receive serious consideration in the literature:
Rather than consider cognitive models as descriptions of a subset of a mechanism’s properties,
they might instead be thought to describe a subset of its internal features.6 Specifically, cognitive
models can be thought to describe a mechanism’s component operations as well as their func-
tional organization, rather than its component parts and their structural organization. Indeed,
insofar as cognitive models are developed by functionally analyzing the behavioral or cognitive
capacity being explained, it should come as no surprise that they are concerned primarily or
even exclusively with a mechanism’s component operations. Piccinini and Craver appear to
have this in mind when they observe that cognitive models are mechanism sketches “in which
some structural aspects of a mechanistic explanation are omitted” (Piccinini & Craver 2011: 284,
emphasis added). Nevertheless, they do not go far enough: Cognitive models may actually be
mechanism sketches in which all structural aspects are left out. For sure, many cognitive models
contain constructs that are labeled with nouns such as “filter,” “channel,” or “representation.”
Nevertheless, these constructs are nearly always defined functionally: A filter is something that
filters, a channel is something that channels, and a representation is something that represents.7
Notably, on this view the fact that a cognitive model’s constructs are specified in mathemati-
cal or computational terms is irrelevant to its ability to explain; in line with the 3M constraint
396
discussed previously, the production rules, symbols, equations, or variables that typically feature
in such a model have just as much explanatory relevance as any other description—abstract or
concrete—of the relevant mechanism’s component operations. That is, it is misleading to align
Marr’s levels of analysis with levels of abstraction.
On this alternative view, “filling in” is a matter of localizing the production rules, symbols,
equations, or variables of a cognitive model by identifying them with a particular mechanism’s
component parts (Bechtel & Richardson 1993). Thus, implementational-level questions
about the way in which certain mathematical or computational constructs are “realized phys-
ically” are not questions about how, but are in fact questions about where a mechanism’s
component operations are carried out. A cognitive mechanism’s component parts may be
situated at several different levels of organization—from the level of molecules to the level
of organisms (Bechtel 2008; Craver 2007)—and they may be highly distributed within any
particular level—spanning whole neural populations, the brain as a whole, but also spanning
the physical boundaries between brain, body, and world (Kaplan 2012; Zednik 2011). Like a
mechanism’s component operations, its component parts can be described in abstract math-
ematical or concrete physical terms: Molecules, nerve cells, neural populations, bodily limbs,
and tools in the environment can all be described in full anatomical or physical detail, but
may also be characterized schematically or abstractly, e.g. as lattice-like structures, geometric
shapes, or graphical topologies. In general, no matter whether they emphasize abstract math-
ematical properties or concrete physical ones, descriptions of operations and their functional
organization answer how-questions at the algorithmic level of analysis, while descriptions of
parts and their structural organization answer where-questions at the implementational level.
In closing, it is worth highlighting an important corollary of this alternative view of the
realization-relationship that obtains between the algorithmic and implementational levels: It
is wrong to assume that cognitive models at the algorithmic level are explanatorily autonomous
(compare Feest 2003; Weiskopf 2011). Investigators have recourse to a plethora of analytic and
experimental techniques with which to characterize the component operations of a cognitive
mechanism, as well as to test the accuracy of any particular characterization. For sure, some of
these techniques are driven by purely behavioral methods, and may be independent of the possi-
bility of localizing operations in a mechanism’s component parts (Busemeyer & Diederich 2010;
Cummins 1983). Nevertheless, the fact that cognitive models are subject to norms that do not
involve localization has no bearing on the issue of whether localization—or more generally, the
ability to identify the physical structures that are involved in the production of a cognitive or
behavioral phenomenon—is also important (compare Stinson 2016). Moreover, it is important
not to exaggerate the ability to describe a mechanism’s component operations independently
of its component parts. There are many historical examples in which the successful characteri-
zation of a mechanism’s component operations was greatly facilitated or even enabled by the
prior identification of its component parts (Bechtel 2008; Bechtel & Richardson 1993; Craver
2013). Indeed, as technological advances amplify investigators’ ability to individuate structures
in the brain and to characterize their functional properties, it seems reasonable to expect that
how-questions at the algorithmic level and where-questions at the implementational levels will
become increasingly intertwined (see also Boone & Piccinini 2015).
5. Conclusion
The primary aim of this chapter has been to resolve ambiguities in Marr’s account of explanation
in cognitive science by subsuming it under the framework of mechanistic explanation. To this
end, it was argued that Marr’s three levels can be individuated by the different types of questions
397
Carlos Zednik
that are typically asked about a cognitive system, and that ambiguities concerning the way in
which these questions are posed by Marr can be resolved by aligning each question with a specific
aspect of mechanistic explanation. Computational-level questions about a system’s computational
goals are in fact what-questions that can be answered by describing a mechanism’s behavior, and
questions about these goals’ appropriateness are in fact why-questions that can be answered by
situating the mechanism in a containing environment. At the algorithmic level, questions about
how a certain computational goal is achieved are in fact questions about a mechanism’s compo-
nent operations. Finally, at the implementational level, questions about the way in which algo-
rithms and representations are physically realized are in fact where-questions that can be answered
by identifying the relevant mechanism’s component parts. Construed in this way, no single level
of analysis bears sole responsibility for delivering mechanistic explanations of behavior and cogni-
tion; all three levels are involved in the discovery and description of cognitive mechanisms.
A secondary aim of this chapter has been to show that several different research programs—
not just the computationalist approach in which Marr was himself embedded—seek to deliver
answers to questions at all three levels of analysis. Thus, the scope of Marr’s account is much
wider than traditionally assumed, encompassing many different areas and traditions in con-
temporary cognitive scientific research. Insofar as researchers across cognitive science answer
questions about the what, why, how, and where of behavior and cognition, they are all in the
business of mechanistic explanation.
Notes
1 The author is indebted to Frank Jäkel, Holger Lyre, the volume editors, and conference audiences in
Düsseldorf and Warsaw for helpful feedback on earlier versions of this chapter.
2 Whereas Marr sometimes speaks of explanations at individual levels of analysis, here the term “explana-
tion” will be reserved for a full three-level account, i.e. the kind of account that Marr deems necessary
for “complete understanding” (Marr 1982: 4ff). This is not meant to be a commitment to any particular
account of the relationship between explanation and understanding, however.
3 Levels of analysis must not be confused with levels of organization within a mechanism (see also Bechtel
2008; Craver 2007). Whereas the former are individuated by the kinds of questions an investigator might
ask about a cognitive system, the latter are individuated by constitution-relations within a mechanism.
Notably, insofar as many real-world cognitive mechanisms are hierarchical (Bechtel & Richardson 1993;
Craver 2007), it may often be profitable to apply all three levels of analysis at several different levels of
organization within a single mechanism.
4 Indeed, some proponents of the Bayesian approach take themselves to be delivering teleological expla-
nations rather than mechanistic explanations (see e.g. Oaksford & Chater 2007). The accuracy of this
characterization has been questioned, however, with some claiming that the Bayesian approach does not
provide explanations at all (Bowers & Davis 2012; Danks 2008), and others arguing that it does in fact
explain in a way that centers on the discovery of mechanisms (Zednik & Jäkel 2016).
5 The present discussion concerns only the realization-relationship that is relevant to Marr’s account
of explanation in cognitive science. No commitment will here be made regarding the realization-
relationship in other contexts (see e.g. Kim 1993). Moreover, the present contribution will not explore
the question of whether the realization-relationship that obtains between the implementational and
algorithmic levels also obtains between the algorithmic and computational levels.
6 Intriguingly, Glennan (2010) has called it a “category mistake” to attribute a mechanism’s causal powers
to its properties, instead of to its component parts, operations, and/or overall organization. The view
advanced here of the relationship that obtains between the algorithmic and implementational levels is
consistent with Glennan’s.
7 The view outlined here is consistent with the proposal that what a particular representation repre-
sents may be secondary to the causal-functional role it plays in a cognitive system (see e.g. Fodor
1980). A particular representation’s causal-functional role is akin to a mechanism’s operation, rather
than, for example, to a part.
398
References
Abrahamsen, A. & Bechtel, W. (2006). Phenomena and mechanisms: Putting the symbolic, connectionist,
and dynamical systems debate in broader perspective. In R. Stainton (Ed.), Contemporary debates in cogni-
tive science. Oxford: Basil Blackwell, 159–186.
Anderson, J. R. (1991). Is human cognition adaptive? Behavioral and Brain Sciences 14: 471–517.
Baumgartner, M. & Gebharter, A. (2015). Constitutive relevance, mutual manipulability, and fat-handedness.
The British Journal for the Philosophy of Science. doi: 10.1093/bjps/axv003.
Bechtel, W. (2008). Mental mechanisms: Philosophical perspectives on cognitive neuroscience. London: Routledge.
Bechtel, W. (2009). Looking down, around, and up: Mechanistic explanation in psychology. Philosophical
Psychology 22: 543–564.
Bechtel, W. & Abrahamsen, A. (2010). Dynamic mechanistic explanation: Computational modeling of
circadian rhythms as an exemplar for cognitive science. Studies in History and Philosophy of Science Part
A 1: 321–333.
Bechtel, W. & Richardson, R. C. (1993). Discovering complexity: Decomposition and localization as strategies in
scientific research. Princeton, NJ: Princeton University Press.
Bechtel, W. & Shagrir, O. (2015). The non-redundant contributions of Marr’s three levels of analysis for
explaining information processing mechanisms. Topics in Cognitive Science 7: 312–322.
Bickle, J. (2015). Marr and reductionism. Topics in Cognitive Science 7: 299–311.
Boone, W. & Piccinini, G. (2015). The cognitive neuroscience revolution. Synthese 193: 1509. doi:
10.1007/s11229-015-0783-4.
Bowers, J. S. & Davis, C. J. (2012). Bayesian just-so stories in psychology and neuroscience. Psychological
Bulletin 138: 389–414.
Buckner, C. (2015). Functional kinds: A skeptical look. Synthese 192: 3915. doi: 10.1007/s11229-
014-0606-z.
Busemeyer, J. R. & Diederich, A. (2010). Cognitive modeling. Newbury Park, CA: SAGE.
Busemeyer, J. R. & Townsend, J. (1993). Decision field theory: A dynamic-cognitive approach to decision
making in an uncertain environment. Psychological Review 100: 432–459.
Chemero, A. & Silberstein, M. (2008). After the philosophy of mind: Replacing scholasticism with science.
Craver, C. F. (2007). Explaining the brain. Oxford: Oxford University Press.
Craver, C. F. (2013). Functions and mechanisms: A perspectivalist view. In P. Huneman (ed.), Functions:
Selection and mechanisms. Dordrecht: Springer, 133–158.
Cummins, R. (1983). The nature of psychological explanation. Cambridge, MA: MIT Press.
Cummins, R. (2000). ‘How does it work?’ versus ‘what are the laws?’: Two conceptions of psychologi-
cal explanation. In F. Keil & R. Wilson (eds.), Explanation and cognition. Cambridge, MA: MIT Press,
117–144.
Danks, D. (2008). Rational analyses, instrumentalism, and implementation. In N. Chater & M. Oaksford
(eds.), The probabilistic mind: Prospects for Bayesian cognitive science. Oxford: Oxford University Press,
59–75.
Darden, L. & Craver, C. F. (2012). In search of mechanisms. Chicago, IL: The University of Chicago Press.
Feest, U. (2003). Functional analysis and the autonomy of psychology. Philosophy of Science 70: 937–948.
Fodor, J. A. (1980). Methodological solipsism considered as a research strategy in cognitive psychology.
Behavioral and Brain Sciences 3: 63–73.
Glennan, S. (2010). Mechanisms, causes, and the layered model of the world. Philosophy and Phenomenological
Research LXXXI: 362–381.
Hummel, J. E. & Biederman, I. (1992). Dynamic binding in a neural network for shape recognition.
Psychological Review 99: 480–517.
Kaplan, D. M. (2011). Explanation and description in computational neuroscience. Synthese 183: 339–373.
Kaplan, D. M. (2012). How to demarcate the boundaries of cognition. Biology and Philosophy 27: 545–570.
Kaplan, D. M. & Craver, C. F. (2011). The explanatory force of dynamical and mathematical models in
neuroscience: A mechanistic perspective. Philosophy of Science 78: 601–627.
Kelso, J. A. S. (1995). Dynamic patterns: The self-organization of brain and behavior. Cambridge, MA: MIT
Press.
Kim, J. (1993). Supervenience and mind: Selected philosophical essays. Cambridge: Cambridge University Press.
Levy, A. & Bechtel, W. (2013). Abstraction and the organization of mechanisms. Philosophy of Science 80:
241–261.
399
Carlos Zednik
Luce, R. D. (1995). Four tensions concerning mathematical modeling in psychology. Annual Review of
Psychology 46: 1–26.
Machamer, P., Darden, L. & Craver, C. F. (2000). Thinking about mechanisms. Philosophy of Science 67:
1–25.
Marr, D. (1982). Vision: A computational investigation into the human representation and processing of visual
information. San Francisco: W. H. Freeman.
McClamrock, R. (1991). Marr’s three levels: A re-evaluation. Minds and Machines 1: 185–196.
Milkowski, M. (2013). Explaining the computational mind. Cambridge, MA: MIT Press.
Nosofsky, R. M. (1986). Attention, similarity, and the identification-categorization relationship. Journal of
Experimental Psychology 115: 39–57.
Oaksford, M. & Chater, N. (2007). Bayesian rationality: The probabilistic approach to human reasoning. Oxford:
Piccinini, G. & Craver, C. F. (2011). Integrating psychology and neuroscience: Functional analyses as
mechanism sketches. Synthese 183: 283–311.
Polger, T. (2004). Neural machinery and realization. Philosophy of Science 71: 997–1006.
Pouget, A., Beck, J. M., Ma, W. J., & Latham, P. E. (2013). Probabilistic brains: Knowns and unknowns.
Nature Neuroscience 16: 1170–1178.
Pylyshyn, Z. (1980). Computation and cognition: Issues in the foundations of cognitive science. Behavioral
and Brain Sciences 3: 111–169.
Ramsey, W. M. (2007). Representation reconsidered. Cambridge: Cambridge University Press.
Rumelhart, D. E., McClelland, J. E. & the PDP Research Group (1986). Parallel distributed processing:
Explorations in the microstructure of cognition. Cambridge, MA: MIT Press.
Shagrir, O. (2010). Marr on computational-level theories. Philosophy of Science 77: 477–500.
Stinson, C. (2016). Mechanisms in psychology: Ripping nature at its seams. Synthese 193: 1585. doi:
10.1007/s11229-015-0871-5.
Strevens, M. (2008). Depth. Cambridge, MA: Harvard University Press.
van Gelder, T. J. (1995). What might cognition be, if not computation? Journal of Philosophy 91: 345–381.
van Gelder, T. J. (1998). The dynamical hypothesis in cognitive science. Behavioral and Brain Sciences 21:
1–14.
Weiskopf, D. A. (2011). Models and mechanisms in psychological explanation. Synthese 183: 313. doi:
10.1007/s11229-011-9958-9.
Zednik, C. (2011). The nature of dynamical explanation. Philosophy of Science 78: 238–263.
Zednik, C. & Jäkel, F. (2016). Bayesian reverse-engineering considered as a research strategy for cognitive
science. Synthese. doi: 10.1007/s11229-016-1180-3.
400
30
SOCIAL MECHANISMS1
Petri Ylikoski
Social mechanisms and mechanism-based explanation have attracted considerable attention in

the social sciences and the philosophy of science during the past two decades. The idea of
mechanistic explanation has proved to be a useful tool for criticizing existing research practices
and meta-theoretical views on the nature of the social-scientific enterprise. Many definitions
of social mechanisms have been articulated, and have been used to support a wide variety of
methodological and theoretical claims. It is impossible to cover all of these in one chapter, so I
will merely highlight some of the most prominent and philosophically interesting ideas.
1. Mechanisms in the social sciences

As in other sciences, mechanism as a notion belongs to the everyday casual vocabulary of many
social scientists. In this context the word “mechanism” could refer to a cause, a causal pathway,
or an explanation without explicit theorizing about the nature of mechanisms. This casual
and occasional mechanistic way of talking is probably as old as the social sciences. Equally
old are certain negative connotations of the word “mechanical” implying simple, rigid, and
reductionist. Being part of everyday casual vocabulary explains much of the intuitive appeal
of mechanistic language, although the negative connotations have made some social scientists
suspicious of the mechanistic turn.
Theorizing about mechanisms has multiple origins in the social sciences. Among the most
prominent early sources is Rom Harré’s (1970) philosophy of science. Although his later work
has been highly influential in social psychology, the biggest impact of his philosophy of science
was through the so-called critical realist movement. The key thinker in this movement is Roy
Bhaskar (1978, 1979), in whose philosophy Harré’s ideas about causation and mechanisms have
a central role. Despite Bhaskar’s transcendental argumentation, his layered account of real-
ity, and his ideas about essences and internal relations that have raised philosophical suspicions
and doubts about their relevance to the social sciences, critical realism is still one of the most
influential meta-theoretical movements in the social sciences (Lawson 2003; Elder-Vass 2011).
Philosophers of mechanisms might be unfamiliar with critical realism, but in the view of many
social scientists the idea of causal mechanisms is strongly associated with this movement.
Another influential early advocate of mechanism-based thinking is Jon Elster, his many
books containing excellent examples of mechanistic thinking in action. His early definition
401
Petri Ylikoski
(Elster 1989), according to which a mechanism explains by providing a continuous and

contiguous chain of causal or intentional links between the explanans and the explanandum,
is quite in line with the general mechanistic perspective. Although Elster’s mechanisms tend
to be psychological rather than social, his work has inspired many social scientists to open up
black boxes and show the cogs and wheels of the internal machinery of social processes. Elster
has also strongly associated the mechanistic attitude with the intellectual virtues of clarity and
precision: he sees mechanism-based theorizing as clearheaded causal thinking about social pro-
cesses and for a large group of social scientists this is the core of the mechanistic perspective.
Note that there are few connections between social scientists inspired by Bhaskar’s critical
realism and those supporting Elster’s approach. The idea of mechanism-based thinking has
multiple interpretations in the social sciences.
Mechanistic thinking started to become mainstream in the 1990s. A work of particular sig-
nificance in this respect was Social Mechanisms (1998), edited by Peter Hedström and Richard
Swedberg. Among contributors to this volume are many scholars who have been influential
in the development of mechanism-based thinking in the social sciences. Later this approach
developed into the analytical sociology movement (Hedström & Bearman 2009; Hedström &
Ylikoski 2010) that has most systematically developed the program of mechanism-based social
science. At the same time, philosophers of the social sciences such as Daniel Little (1991) and
Mario Bunge (1997) started to talk about mechanisms. In political science mechanisms became a
focus of attention a little bit later. There causal mechanisms played an important role in debates
concerning research methodology and causal inference. Especially noteworthy has been the
development of process-tracing methodology for case-study research (Bennett & Checkel 2015).
In general, the debates in political science and sociology have been quite similar: advocates of
mechanisms have criticized the simplistic use of statistical methodology and the downplaying of
the importance of causal process assumptions in causal inference. However, there are also some
important differences, as the later discussion will show.
In the following discussion I focus, first, on what is known as Coleman’s diagram, which
helps to identify the core challenges of mechanism-based theorizing in the social sciences. It
also provides a context in which to discuss the appeal of mechanism-based explanation among
social scientists. In an attempt to make sense of this debate I will introduce a distinction between
causal scenarios and causal mechanism schemes. By way of illustrating social-scientific thinking
about mechanisms, I discuss the use of agent-based simulation as a tool for mechanism-based
theorizing, and introduce the idea of metamechanism. In the concluding discussion I show how
mechanistic ideas have been used in social-scientific debates about causal inference.
2. Coleman’s diagram
A useful starting point for discussing social mechanisms is a diagram known as Coleman’s boat
(Coleman 1987, 1990; see Figure 30.1). It is commonly used in discussions of social mecha-
nisms, and one could say that it has become an emblem of mechanism-based thinking in the
social sciences. Coleman himself did not employ mechanistic vocabulary, but in many ways the
diagram exemplifies mechanistic thinking in the social sciences (Ylikoski 2016).
The recommended starting point for unpacking the diagram is node D, which represents the
macro-social explanandum that the sociologist finds interesting. Node A represents some macro-
social variable that is associated with D. For example, let us assume that A is the implementation
of a government job-training program and D is the decrease in the level of youth unemploy-
ment. Did the job-training program cause the decline in unemployment? In other words, can
arrow 4 be interpreted as causal?
402
Social mechanisms
A 4 D
I 3
B 2 C
Figure 30.1 Coleman’s diagram
Coleman’s main point is the following. To justify causal claims like this, it is necessary to
understand how the suggested cause brings about the effect in question. This idea has two com-
ponents. The first of these concerns the justification of causal claims: the statistical data on the
relevant macro variables is usually so sparse as to be insufficient for establishing such a claim.
Coleman’s suggestion is to test the causal claim by finding out whether there is an empirically
supported mechanism by which A brings about D: if there is, the claim is supported, and if not,
then it should be scrapped. The second component concerns explanation. Coleman and other
supporters of mechanism-based thinking stress the importance of understanding how the effect
was brought about. Thus, even if it were possible to show that variable A is a causal difference-
maker for variable D, it would not suffice for a theoretically satisfactory explanation: the expla-
nation has to spell out the mechanism. These two ideas are at the core of social-scientific
debates about causal mechanisms. I will discuss both in detail later, but first I will consider other
elements of the diagram.
The key point in the diagram is that macro variables have to be connected to activities by
agents. In most cases these agents are human persons, but Coleman also allows for various sorts
of corporate agents to take this role. Thus arrow 1 describes the way in which changes in macro
conditions influence the relevant agents. The change in A may bring about changes in the
beliefs, desires, or other mental attributes of these agents, or it might change the opportunities
or incentives they are facing. Hedström and Swedberg (1998) call these influences situational
mechanisms. They cover the ways in which social structures constrain and enable individuals’
opportunities for action, and how the cultural and social contexts influence individuals’ goals,
beliefs, habits, or cognitive frames.
Arrow 2 in the diagram covers the role of the theory of action in sociological explanation.
The purpose of theory of action is to connect changes in agents’ opportunities and mental states
to changes in their behaviors or actions. Coleman used rational choice theory for this purpose,
but it is also possible to use other theories of action (Hedström & Ylikoski 2014). The important
point here is that mechanical explanations in the social sciences bottom out (Craver 2007) at the
level of individual action (Coleman 1990: 4). The behaviors of individual persons are the basic
components of social mechanisms and social scientists do not look for explanations for these
action-formation mechanisms: this is a job for cognitive scientists and psychologists.
Although both situational and action-formation mechanisms certainly pose their own
challenges, Coleman argues that the transformational mechanisms (arrow 3) are the biggest bot-
tleneck in sociological theory. Social scientists know a great deal about how individuals’ desires,
beliefs, and opportunities, for example, are influenced by the social contexts in which they are
embedded (situational mechanisms), and about how these desires, beliefs, and opportunities
influence actions (action-transformation mechanisms), but when it comes to the link between
individual actions and social outcomes, they are often forced to resort to hand-waving. This is
403
Petri Ylikoski
something Coleman aimed to highlight with his diagram: macro-level patterns are often difficult
to predict from individual-level descriptions, and the way in which individual actions produce
social patterns is rarely a simple process of aggregation. This is not a problem that is specific to
individualistic theories: although holistic theories tend to highlight the contextuality and com-
plexity of everything, in practice their micro-macro assumptions tend to be rather simplistic.
It is frequently assumed that macro facts simply reflect the relevant micro facts, and vice versa.
Schelling’s (1978) well-known checkerboard model shows how wrong this assumption is even
in very simple settings: it shows how residents who do not favor segregation may still end up in
a highly segregated neighborhood. It is clear even from this very simple model that one cannot
assume that macro facts will reflect individual (or average individual) preferences. It is neces-
sary to understand how the macro outcome is brought about by the interdependent actions of
individuals; in other words, one has to understand “the rules of the game” (Coleman 1990: 19).
Coleman was never able to fully unpack the metaphor of the rules of the game, but his dis-
cussion and examples were enough to interest sociologists in the theoretical challenge of the
micro-macro link. Consequently, whereas others might talk loosely about emergent macro-scale
properties, sociologists following Coleman set themselves the much harder task of explaining
how that emergence comes about. The key mechanistic idea captured in the diagram is that only
when we have understood the whole chain of situational, action-formation, and transformational
mechanisms have we understood the relation between the macro-scale social facts. Underlying
each A–D association or causal relation is a combination of these mechanisms. All three are
required, otherwise the explanation would not cover all elements of the micro-macro link.
Users of the diagram have been criticized (Jepperson & Meyer 2011; Little 2012) for being
committed to a reductive ideal of methodological individualism. I believe this criticism is mis-
placed, at least to some extent. The diagram does not describe the reduction of macro facts to
facts about individuals and their relations. The mistake is the assimilation of the Coleman dia-
gram into the supervenience/realization diagrams used in the philosophical debate on mental
causation. Coleman’s arrows represent explanatory rather than reductive relations, and if one
considers the details it is evident that various sorts of structural assumptions play a central role
in situational and transformational mechanisms. Coleman’s main point is that the structural facts
are not explanatory in themselves. It is necessary to understand how they bring about their effects
via the activities and cognitions of individuals, and a full understanding of how the social whole
works requires an understanding of how its behavior is generated by the activities of its members.
In Coleman’s view, individual agents are the basic building blocks of social mechanisms, and
therefore have to be included in any mechanistic social explanation. However, this does not
imply that the explanation entails the reduction of macro-social facts to facts about individuals.
3. Mechanism-based explanations
In the social sciences, as in the philosophy of science, the mechanism-based accounts of expla-
nations have been developed as alternatives to the once dominant covering-law account of
explanation (Hempel 1965). While social scientists have been familiar with famous counter-
examples to that theory and philosophical problems raised by them (such as problems of the
explanatory relevance, the symmetry between explanation and prediction, the asymmetry of
explanation, and many problems in analyzing the notion of law), social scientists have mainly
been concerned with the apparent implausibility of the theory as a model for the social sciences.
There are very few laws in the social sciences, and even those are better described as explananda
rather than explanatia. Most social-scientific generalizations that are not truisms are quite
limited in domain, and include exceptions that neat ceteris paribus conditions cannot cover.
404
Social mechanisms
Furthermore, the strategy of formulating social-scientific theories in terms of axioms and laws
has turned out to be very unproductive, resulting in incomplete collections of sterile generali-
zations marred with unclear domains of application and high levels of conceptual indetermi-
nacy. Finally, the basic idea of a covering-law theory is counterintuitive for social scientists.
The best social-scientific explanations seem to do more than simply subsume the phenomenon
under more general empirical regularities: a good explanation shows how the suggested cause
brings about the effect to be explained. The initial appeal of mechanistic ideas lies in this gen-
erative notion of explanation.
The dissatisfaction with covering-law theory has not led to general agreement on the defi-
nition of a mechanism, however. The intuitive idea can be developed in multiple directions,
especially when the people have different applications in mind. As a result, the literature on
social mechanisms notoriously abounds with apparently incompatible definitions of mecha-
nisms. Mahoney (2001), for example, lists 24, and subsequent contributions have added to
the number. Some critics (e.g., Norkus 2005) regard this multiplicity as a serious problem for
mechanism-based explanation. This could be an overstatement in that, although the definitions
are formally incompatible, most of them could be considered attempts to capture the same basic
ideas. The absence of a generally agreed definition for basic concepts such as gene and species
has not stalled the development of the biological sciences. As long as mechanism-oriented social
scientists agree on the central exemplars of mechanism-based explanations and share a similar
understanding of their general characteristics (Hedström & Ylikoski 2010), they should manage
well without a general definition. Thus the main challenge is not to provide a general defini-
tion of mechanism-based explanations; it is rather to arrive at a consensus about prototypical
examples of mechanistic explanation.
A more serious problem is that much of the mechanism discourse in the social sciences is
quite loose. As noted above, as a notion, mechanism belongs to the general causal vocabulary of
social scientists. Some advocates of social mechanisms seem to go along with this loose talk, and
it is not uncommon to name processes that produce certain kinds of outcome as mechanisms
and leave it at that. In such usage, the mechanism is just a label for a black box, a name for an
effect, not an explanation. The invisible hand, cumulative advantage, and democratization are
not mechanisms in themselves; they are processes that produce specific kinds of outcomes. At
most they are names for families of mechanisms. However, if one does not distinguish at least
some family members, all one has is a placeholder for something substantial. The key point here
is that if one looks into how the invisible hand or cumulative advantage work in practice, one
finds that there may be multiple mechanisms underlying these effects, which sometimes work
separately, and sometimes act together in various combinations. The point here is not about
the proper use of the word “mechanism”: it is about dangerous ambiguity. Confusion between
naming an effect and providing a mechanistic explanation for it may give rise to an illusion of
depth of understanding (Ylikoski 2009). Furthermore, the ambiguity makes one blind to the
possibility of there being multiple mechanisms that are responsible for similar kinds of effect.
4. Causal scenarios and causal mechanism schemes

Social scientists are interested in explaining particular causal outcomes and in developing gen-
eral theories about social mechanisms. In both contexts they refer to mechanisms, which some-
times causes unnecessary confusion. To avoid this, it is useful to distinguish between causal
scenarios and causal mechanism schemes (Ylikoski & Aydinonat 2014). Causal scenarios are
(selective) representations of particular causal processes responsible for some concrete event
or phenomenon. Used thus, causal mechanism refers to a causal narrative that describes the
405
Petri Ylikoski
process that is responsible for the explanandum. This narrative may be highly detailed or a
mere sketch, but in any case it does more than cite a cause that had an effect on the event (see
Chapter 31): it describes the crucial elements in the relevant causal chain. In other words, the
causal scenario describes how the explanandum event came about. The distinction between
how-possibly and how-actually explanations applies here. Usually there are many different ways in
which the outcome could have come about, which competing how-possibly scenarios describe.
The challenge for researchers is to find evidence that could discriminate between these alterna-
tives and enable them to make a judgment about which scenario is the true explanation.
When political scientists talk about process tracing, they are concerned with causal scenarios.
On the other hand, when analytical sociologists discuss causal mechanisms, they tend to refer
to causal mechanism schemes (compare Chapter 19). For example, sociologists talking about self-
fulfilling prophecies or vacancy chains are referring to causal mechanism schemes, which are
abstract representations of mechanisms that could bring about effects of a certain kind. The
explanandum of a causal mechanism scheme tends to be quite abstract, or stylized, reflecting
the fact that such schemes are not primarily explanations of particular facts, but schemes for
constructing them. Thus it is useful to think of them as abstract building blocks that can be
adapted and filled in to serve a role in causal scenarios that explain particular facts. A single causal
scenario might be a combination of many different causal mechanism schemes, and might even
contain mechanism schemes that have opposite causal effects.
Causal mechanism schemes are at the core of analytical sociology’s account of growth of
theoretical knowledge, according to which social-scientific knowledge accumulates through the
development of middle-range theories (Hedström & Udéhn 2009) about social mechanisms. In
this view the core theoretical knowledge comprises a collection of causal mechanism schemes
that can be adapted to particular situations and explanatory tasks. According to this toolbox
view (Hedström & Ylikoski 2010; Elster 2015), social-scientific knowledge is not integrated
into highly abstract general theory, but consists of a growing collection of causal mechanism
schemes that are mutually compatible. The understanding of the social world accumulates as
the knowledge of the mechanism schemes becomes more detailed and the number of known
mechanisms increases. Understanding of more complicated phenomena requires combining
different mechanism schemes, hence knowledge also expands through learning how to create
these “molecular” mechanisms.
The similarity between this view of the architecture of theoretical knowledge in the social
sciences and the mechanistic view of the biological sciences is obvious (see Chapter 19), and
offers a fruitful opportunity to compare two domains of knowledge. However, the relevance
of the toolbox vision is not limited to meta-theory: it also gives new tools to counter the frag-
mentation of the social sciences. Causal mechanism schemes can be shared among the different
subfields, which would allow for a novel type of integration: various subfields employ and
develop the same theoretical toolbox and thereby benefit from each other’s work. The toolbox
vision for sociological theory shows that in the social sciences mechanistic ideas are not confined
to discussions about explanation and causation, but they also play an important role in how
social scientists think about the nature of social-scientific knowledge.
5. Agent-based simulation and generative explanation

One of the main points about mechanism-based explanation is that it should describe how the
properties, activities, and relations of components bring about the phenomenon to be explained
(see Chapters 1 and 19). This concern with generative processes makes generative sufficiency a
central concern in the evaluation of explanations: the suggested explanation should, at least in
406
Social mechanisms
principle, be capable of bringing about the outcome in the specified circumstances. However,
demonstrating generative sufficiency is not easy. The social outcomes of interest typically result
from numerous individuals acting and interacting with one another over extended periods of
time. Furthermore, the aggregate behavior of these complex dynamic systems is extremely
difficult to understand and to predict without the aid of analytical tools. Agent-based simulation
modeling (ABM) provides such a tool (Miller & Page 2007; Squazzoni 2012; Ylikoski 2014).
There is a natural affinity between the components of mechanism-based explanations and ABM.
Like any society, social-scientific ABMs comprise agents with goals and beliefs. These agents
possess resources and influence each other. It is easy to see how the macro-patterns they create
come to be regarded as analogical to the processes that social scientists study. An ABM sheds
light on how the phenomenon to be explained could have been generated, and how changes in
agents’ attributes or relational structures change the macro outcomes.
Joshua Epstein (2006) offers the strongest ABM-based formulation of the idea of generative
explanation. According to Epstein, “If you didn’t grow it, you didn’t explain it.” He meant
that producing the macro-level outcome by means of ABM is a necessary condition for its
explanation. However, it could be argued that simply “growing” the phenomenon of interest
is not sufficient to engender a proper understanding. The crucial challenge is to understand how
the specified micro-configuration produces the phenomenon. Thus, in reformulating Epstein’s
slogan Macy and Flache (2009: 263) make a crucial point: “If you don’t know how you grew it,
you didn’t explain it.” In any case, social scientists using ABM are not satisfied with mechanistic
explanation as mere storytelling. They do not think it is enough to provide a qualitative narra-
tive about the process and its components. They also consider it important to demonstrate that
the suggested mechanism can, in fact, bring about the effect to be explained. Having a detailed
how-possibly scenario is also a precondition for the real empirical testing of it. Only if competing
causal scenarios are clearly articulated is it possible to look for crucial empirical evidence that
discriminates between them. Still, ABM has a long way to go to become a mainstream tool in
sociological research: it is basically a tool for working with causal mechanism schemes rather
than concrete causal scenarios (Ylikoski 2014). However, the connection between ABM and
mechanism-based thinking is strong.
6. Metamechanisms
Jeremy Freese and Karen Lutfey’s (2011) metamechanism is an interesting addition to the concep-
tual toolbox of mechanical philosophy. Behind this idea is Link and Phelan’s (1995) suggestion
that socioeconomic status is a fundamental cause of health differences. They refer to an extremely
robust empirical finding: socioeconomic status and health are strongly correlated. The lower-
status people are, the sooner they die, and the worse health they have while alive. This associa-
tion holds for virtually any society and historical period for which there is adequate empirical
data. The puzzling thing is that the causes of death and disease have changed a lot over the past
hundred years. In other words, the proximate mechanisms of death and ill-health are highly
variable. What explains this puzzling pattern?
Freese and Lutfey suggest that the idea of fundamental cause lends itself to a mechanistic
interpretation: socioeconomic status (SES) is associated with a metamechanism, in other words
a general mechanism that explains the generation of multiple proximate mechanisms that repro-
duce a particular relationship in different places and at different times. This helps to make sense of
Link and Phelan’s findings. People with a higher SES have more resources and education, which
makes it easier for them to use new medical services and health-improving inventions. Even
if healthcare is universal and free, the higher-SES people are better placed to make use of it.
407
Petri Ylikoski
Other metamechanisms may have a similar influence. There are spillover effects, for example,
even among individuals who do not especially care about their health. Higher-SES individuals
will have better health because they tend to gain more benefits from the purposive actions of
others in their social networks (that are partly based on SES). Similarly, health-related institutions
might be biased toward higher-SES people: they are given better service and better understand
the instructions they get. The proximate mechanisms vary, but as long as the metamechanisms
remain in place, there will be health inequalities. The usefulness of such mechanisms is not lim-
ited to the sociology of health, and could extend to some biological contexts, for example.
7. Mechanisms and statistical methodology

Apart from explanation, mechanisms are also important in the context of justification of causal
claims. Especially in non-experimental contexts that are common in the social sciences, they
are said to play a crucial role in distinguishing true causal relations from spurious correlations.
Knowing that there is a mechanism through which X could influence Y supports the inference
that X is a cause of Y. In addition, the absence of a plausible mechanism linking X to Y gives good
reason to be suspicious of any straightforward causal interpretation of the association. Knowledge
of mechanisms is also applied in extrapolating causal findings. The assumption of similarity among
causal mechanisms is a crucial element in making inferences from one setting or population to
another (Steel 2008; see also Chapter 32). However, there is much ambiguity in these mecha-
nistic slogans. It is impossible to ascertain that knowledge of mechanisms is necessary for justifying
causal claims without a clear idea of what such knowledge consists of, and how much of it is
needed. All causal inference presupposes some causal background assumptions, but do all such
assumptions concern causal mechanisms? It should also be recognized that mechanisms are not
a magic wand for causal inference in the social sciences. The problem in many cases is not the
absence of a possible mechanism, but insufficient evidence to discriminate between competing
mechanistic hypotheses. Similarly, lazy mechanism-based storytelling is a constant threat: having
a good story is no substitute for real statistical evidence. It is not rare for a good story about a
(possible) mechanism to make people forget how important it is to test whether such a mecha-
nism really is in place and whether it can really account for the intended explanandum.
Thinking in terms of mechanisms is often set against statistical methodology in the social-
scientific debates on causal inference. This opposition takes different forms. Many critical realists
rely on theorizing about mechanisms as an alternative to using statistics and causal-modeling
techniques. Causal modeling is said to embody problematic Humean ideas about causation
that make it suspect and of limited value, which is why critical realists tend to use statistics for
descriptive purposes only and prefer qualitative evidence and theoretical argumentation.
An alternative is not to give up statistical and causal modeling, but to object to their use
without consideration of the relevant causal mechanisms. Here the claim is that statistical tech-
niques have replaced substantial theorizing. Peter Hedström conveys a fairly common sentiment
in the following:
Although most causal modelers refer to sociological theories in their work, they rarely
pay it any serious attention. More often than not, they simply use theories to justify
the inclusion of certain variables taken from a data set that has often been collected for
entirely different purposes than the one to hand. Theoretical statements have become
synonymous with hypotheses about relationships between variables, and variables have
replaced actors as the active subjects with causal powers.
(Hedström 2005: 105)
408
Social mechanisms
What he is calling for is a fuller incorporation of theory (about causal mechanisms) into the
research design and the interpretation of statistical data. According to this view, sociologi-
cal research should not be limited to the measurement of causal effects among conventional
variables, but should focus on how the social world works. Given this background, it is under-
standable that advocates of mechanisms have been resistant to the common assumption that
mechanisms are just intervening variables. Although the existence of a causal process or mecha-
nism implies that there are intervening variables, not all of them are necessarily of the right sort.
They may tell more about other effects of the mechanism than about the mechanism itself, for
example. Thus, mere intervening variables do not guarantee the explanatory depth that is the
main concern among mechanists.
The promise of mechanisms is to provide something more to causal inference, but what is
that additional element? It is now generally recognized that the uses of statistical tools such as
regression analysis presupposes substantial assumptions about the causal relations that are mod-
eled (Kincaid 2012). A substantial proportion of these background assumptions concern possible
causal mechanisms. There have also been attempts to incorporate mechanistic thinking into
causal modeling (e.g., Knight & Winship 2014; see also Imai et al. 2011) that defines mecha-
nisms as modular sets of entities connected by relations of counterfactual dependence. (Philosophers will
easily recognize the influence of Woodward 2002 here.) According to Knight and Winship,
as long as the mechanisms studied satisfy the requirement of modularity, Judea Pearl’s DAG
(Directed Acyclic Graph) calculus is a powerful tool in terms of facilitating the rigorous con-
sideration of mechanisms in causal analysis. Their main argument is that mechanisms and causal
analysis can be combined fruitfully in a way that could help in identifying causal effects even
when traditional techniques fail. It is to be expected that attempts like this to combine mecha-
nistic thinking with causal modeling will become more frequent. It appears that the issue is not
really about statistical methods, but concerns the way they are used.
Causal graphs are useful when it comes to thinking about social mechanisms, and mech-
anists should welcome them. However, it should be borne in mind that there are serious
limitations in terms of what can be represented as DAGs. Coleman’s diagram discussed above
cannot be interpreted as a DAG, for example. One problematic spot is arrow 1 between
A and B. Consider how the demographic change caused by war affects the number of poten-
tial marriage partners available to women, how laws allowing same-sex marriage change the
opportunities of same-sex couples to arrange their legal relationship, and how the improved
educational level of society is related to the education of individuals. It cannot be said in any
of these cases that the relations are strictly causal, whereas it is plausible to say in these cases
the cited A-facts partially consist of the mentioned B-facts. However, some of the relevant
consequences of changes in A might well be causal at the same time. Thus it seems that A–B
explanatory dependency is based on (various) mixtures of causal and constitutive relations
(Ylikoski 2013, 2016). This makes sense on the theoretical level, but also makes it impossible
to interpret arrow 1 as causal or the diagram as a DAG.
In political science debates, especially in international relations, mechanistic thinking has also
been set against statistical methods. However, the context has been that of qualitative case stud-
ies, the point being that “the standard quantitative template” is ill adapted for such research. The
alternative methodology for causal inference is called process tracing, which Bennett and Checkel
define as “the analysis of evidence on processes, sequences, and conjunctures of events within
a case for the purposes of either developing or testing hypotheses about causal mechanisms
that might causally explain the case” (Bennett & Checkel 2015: 7). As this definition implies,
mechanisms—causal scenarios—play a central role in process tracing. Causation is understood
here as a continuous process and the task is to explain a singular event. A central concern in
409
Petri Ylikoski
process tracing is with the sequence of events and mechanisms involved in the unfolding of the
process. The researcher looks for diagnostic evidence that can be used to discriminate between
alternative causal scenarios that could explain the event.
Process tracing is often presented as a method for “within-a-case causal inference,” but its
functions remain somewhat unclear: it has been presented as a tool for theory testing, theory
development, and for explaining singular outcomes (Beach & Pedersen 2013). These roles are
naturally interlocking and not so easy to distinguish. It is rare to begin a case study with the goal
of developing a theory, but the search for explanation might generate novel theoretical ideas.
The competing hypotheses are competing explanations, and hence it does not make much sense
to distinguish hypothesis testing and explanation as separate activities. Furthermore, it is doubt-
ful that process tracing captures something that is unique to qualitative case-study research given
that similar case-based causal process observations play an important role in the evaluation of evi-
dence in experimental, comparative, and statistical studies. Consideration of the causal processes
that produce the data to be analyzed is a major concern in all research. Thus there is much room
for building new bridges between different research methodologies based on mechanistic ideas.
An interesting contribution in the literature on process tracing is the taxonomy of tests
for causal hypotheses. Originally presented by Van Evera (1997), but later adapted by others
(Beach & Pedersen 2013; Bennett & Checkel 2015), this taxonomy describes different kinds of
tests that people struggling with the problem of multiple competing causal scenarios could look
for. Passing a Smoking-Gun test gives strong support to the hypothesis and substantially weakens
its rivals, but failure does not imply that the hypothesis is eliminated: it is only weakened and the
rivals gain some additional support. The very name of the test is illustrative: finding a person with
a smoking gun straight after a shooting makes him or her a strong suspect, but the lack of such
evidence does not eliminate this person from the list of suspects. In the case of the Hoop test the
implications are the opposite: passing the test affirms the relevance of the hypothesis, but does not
confirm it. However, if the hypothesis fails the test, it is eliminated. Here the illustration is the
familiar idea of an alibi: giving a speech to an audience of dozens of people at the time of the crime
provides strong grounds for elimination from the list of suspects. However, the mere lack of an
alibi does not yet provide positive evidence of guilt. Doubly Decisive tests are rare, but they are the
strongest. A hypothesis that passes the test is confirmed and competing hypotheses are eliminated.
The consequences of failure are also drastic, but only for the failed hypothesis: it is eliminated.
Having a clear video recording of the crime with the shooter’s face clearly visible is an example of
a doubly decisive test: the video evidence demonstrates that the particular suspect is responsible,
and also shows that other suspects took no part in the shooting. Finally, there are Straw-in-the-
Wind tests. Passing such a test gives the hypothesis some support, but failure does not mean that
it is eliminated. The evidence is weak or circumstantial and cannot in itself prove or disprove the
suspect’s guilt. However, in favorable circumstances enough accumulated evidence of this type
may convince the jury. In the context of the social sciences most evidence is inherently of the
Straw-in-the-Wind kind, and social scientists rarely encounter evidence that could be considered
Doubly Decisive, or that would constitute a Smoking-Gun test. However, the taxonomy is useful in
highlighting the fact that the value of evidence depends on the set of alternative hypotheses, not
on some intrinsic relationship between a single hypothesis and the empirical material.
8. Conclusion
In the above review I have covered some prominent and interesting themes in the social-scientific
debate on mechanisms. I have left many things out, some of which are discussed in other chapters
of this book. I have attempted to show that mechanism-based thinking is a strong and expanding
410
Social mechanisms
meta-theoretical idea in the social sciences, and that some of the ideas, such as the distinction
between causal scenarios and causal mechanism schemes, and the notion of metamechanisms,
might also be of interest in other disciplines.
Note
1 This research received funding from the European Research Council under the European Union’s
Seventh Framework Programme (FP7/2007-2013)/ERC grant agreement no. 324233, Riksbankens
Jubileumsfond (DNR M12-0301:1), and the Swedish Research Council (DNR 445-2013-7681 and
DNR 340-2013-5460).
References
Beach D & Pedersen RB (2013). Process-Tracing Methods. Foundations and Guidelines. Ann Arbor, MI: The
University of Michigan Press.
Bennett A & Checkel JT (2015). Process Tracing. From Metaphor to Analytic Tool. Cambridge: Cambridge
University Press.
Bhaskar R (1978). A Realist Theory of Science. 2nd ed. Brighton: Harvester.
Bhaskar R (1979). The Possibility of Naturalism. Brighton: Harvester.
Bunge M (1997). Mechanism and Explanation. Philosophy of the Social Sciences 27:410–65.
Coleman JS (1987). Microfoundations and Macrosocial Behavior, in Alexander, Giesen, Münch & Smelser
(eds.): The Micro-Macro Link. Berkeley, CA: University of California Press: 153–73.
Coleman JS (1990). Foundations of Social Theory. Cambridge, MA: The Belknap Press.
Craver CF (2007). Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience. New York: Oxford
University Press.
Elder-Vass D (2011). The Causal Power of Social Structures: Emergence, Structure and Agency. Cambridge:
Elster J (1989). Nuts and Bolts for the Social Sciences. Cambridge: Cambridge University Press.
Elster J (2015). Explaining Social Behavior: More Nuts and Bolts for the Social Sciences. 2nd ed. Cambridge:
Epstein J (2006). Generative Social Science: Studies in Agent-Based Computational Modeling. Princeton, NJ:
Princeton University Press.
Freese J & Lutfey K (2011). Fundamental Causality: Challenges of an Animating Concept for Medical
Sociology, in Pescosolido et al. (eds.): Handbook of the Sociology of Health, Illness, and Healing. Dordrecht:
Springer: 67–81.
Harré R (1970). The Principles of Scientific Thinking. London: Macmillan.
Hedström, Peter (2005). Dissecting the Social: On the Principles of Analytical Sociology. Cambridge: Cambridge
University Press.
Hedström P & Bearman P (eds.) (2009). The Oxford Handbook of Analytical Sociology. Oxford: Oxford
University Press.
Hedström P & Swedberg R (eds.) (1998). Social Mechanisms: An Analytical Approach to Social Theory.
Hedström P & Udéhn L (2009). Analytical Sociology and Theories of the Middle Range, in Hedström &
Bearman (eds.): The Oxford Handbook of Analytical Sociology. Oxford: Oxford University Press: 25–49.
Hedström P & Ylikoski P (2010). Causal Mechanisms in the Social Sciences. Annual Review of Sociology
36, 49–67.
Hedström P & Ylikoski P (2014). Analytical Sociology and Rational Choice Theory, in Gianluca Manzo
(ed.): Analytical Sociology: Norms, Actions and Networks. New York: Wiley: 57–70.
Hempel C (1965). Aspects of Scientific Explanation. New York: Free Press.
Imai K, Keele L, Tingley D, & Yamamoto T (2011). Unpacking the Black Box of Causality: Learning
about Causal Mechanisms from Experimental and Observational Studies. American Political Science
Review 105, 765–89.
Jepperson R & Meyer JW (2011). Multiple Levels of Analysis and the Limitations of Methodological
Individualisms. Sociological Theory 29, 54–73.
Kincaid H (2012). Mechanisms, Causal Modeling, and the Limitations of Traditional Multiple Regression, in
Kincaid (ed.): The Oxford Handbook of Philosophy of Social Science. Oxford: Oxford University Press: 46–64.
411
Petri Ylikoski
Knight CR & Winship C (2014). The Causal Implications of Mechanistic Thinking: Identification Using
Directed Acyclic Graphs (DAGs), in Morgan (ed.): Handbook of Causal Analysis for Social Research,
Dordrecht: Springer: 275–300.
Lawson T (2003). Reorienting Economics. London and New York: Routledge.
Link BG & Phelan J (1995). Social Conditions as Fundamental Causes of Disease. Journal of Health and Social
Behavior 35 (Extra Issue), 80–94.
Little D (1991). Varieties of Social Explanation: An Introduction to the Philosophy of Social Science. Boulder, CO:
Westview.
Little D (2012). Explanatory Autonomy and Coleman’s Boat. Theoria 74, 137–51.
Macy M & Flache A (2009). Social Dynamics from the Bottom Up: Agent-Based Models of Social
Interaction. In Hedström and Bearman (eds.): The Oxford Handbook of Analytical Sociology. Oxford:
Oxford University Press: 245–268.
Mahoney J (2001). Review Essay: Beyond Correlational Analysis: Recent Innovations in Theory and
Method. Sociological Forum 16, 575–93.
Miller JH & Page SE (2007). Complex Adaptive Systems: An Introduction to Computational Models of Social Life.
Princeton, NJ: Princeton University Press.
Norkus Z (2005). Mechanisms as Miracle Makers? The Rise and Inconsistencies of the “Mechanismic
Approach” in Social Science and History. History and Theory 44: 348–72.
Schelling T (1978). Micromotives and Macrobehavior. New York: W.W. Norton.
Squazzoni F (2012). Agent-Based Computational Sociology. Chichester: John Wiley & Sons.
Steel D (2008). Across the Boundaries: Extrapolation in Biology and Social Sciences. Oxford: Oxford University
Press.
Van Evera S (1997). Guide to Methods for Students of Political Science. Ithaca, NY: Cornell University Press.
Woodward J (2002). What Is a Mechanism? A Counterfactual Account. Philosophy of Science 69, S366–S377.
Ylikoski P (2009). The Illusion of Depth of Understanding in Science. In De Regt, Leonelli & Eigner
(eds.): Scientific Understanding: Philosophical Perspectives. Pittsburgh: Pittsburgh University Press: 100–119.
Ylikoski P (2013). Causal and Constitutive Explanation Compared. Erkenntnis 78, 277–97.
Ylikoski P (2014). Agent-Based Simulation and Sociological Understanding. Perspectives on Science 22,
318–35.
Ylikoski P (2016). Thinking with the Coleman Boat. The IAS Working Paper Series, 1, http://urn.kb.se/res
olve?urn=urn:nbn:se:liu:diva-132711.
Ylikoski P & Aydinonat E (2014). Understanding with Theoretical Models. Journal of Economic Methodology
21, 19–36.
412
31
DISAGGREGATING
HISTORICAL EXPLANATION
The move to social mechanisms
Daniel Little
1. The covering-law model

Carl Hempel published his chief contribution to the philosophy of history in 1942, nearly
75 years ago. The article is “The Function of General Laws in History” (Hempel 1942), and it
set the stage for several fruitless decades of debate within analytic philosophy about the nature
of historical explanation. Hempel argued that all scientific explanations have the same logi-
cal structure: a deductive (or probabilistic) derivation of the explanandum from one or more
general laws and one or more statements of fact. Explanation, in Hempel’s view, simply is
derivation of the explanandum from general laws. He is emphatic, moreover, in insisting that
valid explanations in history must have this form:
We have tried to show that in history no less than in any other branch of empiri-
cal inquiry, scientific explanation can be achieved only by means of suitable general
hypotheses, or by theories, which are bodies of systematically related hypotheses.
(1942: 44)
What kinds of general laws does Hempel think that historians have in the back of their minds
when they offer elliptical explanations? He refers to regularities of individual or social psychol-
ogy (40), regularities of collective behavior (“groups migrate to regions which offer better living
conditions”), or at the macro level, regularities linking growing discontent to the outbreak of
revolution (41).
Hempel concedes the point that few existing historical explanations actually look like this,
with explicit law statements embedded in a deductive argument; but he argues that this shows
only that existing explanations are elliptical, incomplete, or invalid. And often, he finds, what
is offered as a historical explanation is in fact no more than an “explanation sketch” (42), with
placeholders for the general laws.
This set of assumptions leads to insolvable problems for historical explanation if we accept
Hempel’s account, however, because it is hard to think of a real historical research question
where there might be a set of social or individual regularities sufficient to deductively entail the
outcome. Bluntly, the social and behavioral sciences have never produced theories of individual
or collective behavior that issue in statements of general laws that could be the foundation for a
413
Daniel Little
covering-law explanation of an historical event or pattern. And given that social phenomena are
formed by actors with a range of features of agency and decision-making, we have very good
reason to think that this lack of regularities is inherent in the social world. The social world is
simply not governed by a set of social or individual laws.
So the strong governing laws that would be needed for a covering-law explanation do
not exist in the social domain. There are regularities and generalizations, but they are small
in scope and do not aggregate through deduction to the derivation of regularities among
large events or structures. Social regularities are phenomenal, not governing; they reflect
characteristics of the actors rather than governing the behavior of the ensembles (Little 1993).
This does not this mean that historical explanation is impossible. But we need to turn our
attention from large social regularities to causal mechanisms and powers to see what a good
historical explanation looks like. (Philosophers of social science who continue to maintain
that social explanation depends upon the discovery of generalizations include (Kincaid 1990,
1996; McIntyre 1996).) More generally, the dominant quantitative paradigm for sociological
research depends crucially on the idea that social explanation requires the discovery of robust
correlations and associations (Abbott 1998).
Instead of imagining the social world in analogy with the law-governed world of natural
phenomena, some philosophers of history propose an approach to social science theorizing that
emphasizes agency, contingency, and plasticity in the makeup of social facts (Little 2010). This
approach recognizes that there is a degree of pattern in social life—but emphasizes that these
patterns fall far short of the regularities associated with laws of nature. It emphasizes contingency
of social processes and outcomes. It insists upon the importance and legitimacy of eclectic use
of social theories: the processes are heterogeneous, and therefore it is appropriate to appeal to
different types of social theories as we explain social processes. It emphasizes the importance of
path-dependence in social outcomes. It suggests that the most valid scientific statements in the
social sciences have to do with the discovery of concrete social-causal mechanisms, through
which some types of social outcomes come about.
McAdam, Tarrow, and Tilly illustrate this approach in their treatment of contentious poli-
tics in Dynamics of Contention (McAdam, Tarrow, and Tilly 2001), and the field of contentious
politics is in fact highly suitable to the mechanisms approach. There are numerous clear exam-
ples of social processes instantiated in groups and organizations that play into a wide range of
episodes of contention and resistance—the mechanics of mobilization, the processes that lead
to escalation, the communications mechanisms through which information and calls for action
are broadcast, the workings of organizations. So when we are interested in discovering explana-
tions of the emergence and course of various episodes of contention and resistance, it is both
plausible and helpful to seek out the specific mechanisms of mobilization and resistance that can
be discerned in the historical record.
These considerations suggest that a good historical explanation identifies a number of inde-
pendent mechanisms and processes that are at work in a particular circumstance, and then
demonstrates how these mechanisms, and the actions of the actors involved, lead to the outcome.
Historical explanations typically take the form of causal narratives linking actions, circumstances,
and consequences leading up to the outcome to be explained.
2. Social mechanisms
The social mechanisms approach to explanation (SM) has filled a very important gap in the
theory of social explanation in the past twenty years, between the covering-law model and
merely particularistic accounts of specific events. The SM approach is particularly prominent in
414
Disaggregating historical explanation
the emerging program of analytical sociology (Hedström and Swedberg 1996) and critical realism
(Bhaskar 1975), but has made its mark in comparative historical sociology and other areas of the
social sciences as well (Steinmetz 2004; Gorski 2013; Gross 2009; McAdam et al. 2001).
The core of this approach is the idea that social explanation requires discovery of the under-
lying causal mechanisms that give rise to outcomes of interest (Glennan 1996; Cartwright 1999;
Cummins 2000; Machamer et al. 2000; Woodward 2003; Hedström 2005; Craver and Darden
2013). On this perspective, explanation of a phenomenon or regularity involves identifying the
specific causal processes and causal relations that underlie this phenomenon or regularity, and
causal mechanisms are heterogeneous (see Chapter 12). Social mechanisms are concrete social
processes in which a set of social conditions, constraints, or circumstances combine to bring
about a given outcome (see Chapter 30). A strong example of this approach is to be found in
Dynamics of Contention where McAdam, Tarrow, and Tilly place primary emphasis on causal
mechanisms and processes as the underlying factors that serve to explain complex episodes of
social contention (McAdam et al. 2001: 30). On this approach, social explanation does not take
the form of inductive discovery of laws; rather, the generalizations that may be discovered in
the course of social science research are subordinate to the more fundamental search for causal
mechanisms and pathways in individual outcomes and sets of outcomes. This approach casts
doubt on the search for generalizable theories across numerous societies. It looks instead for
specific causal influence and variation. The approach emphasizes variety, contingency, and the
availability of alternative pathways leading to an outcome, rather than expecting to find a small
number of common patterns of development or change. The contingency of particular path-
ways derives from several factors, including the local circumstances of individual agency and
the across-case variation in the specifics of institutional arrangements—giving rise to significant
variation in higher-level processes and outcomes.
There is an important ontological assumption underlying the concept of a mechanism—the
idea that there is a substrate that makes the mechanism work. By referring to a nexus between
I and O as a “mechanism,” we presume that there is some underlying domain of things and
powers that makes the observed regularity a “necessary” one: given how the world works, the
input I brings about events that lead to output O. In evolutionary biology it is the specifics of
organisms within an ecology conjoined with natural selection. Crucial to a valid understand-
ing of social mechanisms is a general answer to the question: what ontological features of the
social world facilitate and empower the causal connection from one end of a social mechanism
to another? What stands in the place of “causal powers of natural entities” in the social world?
In the social world it is the concrete situation of the actor and the social and natural envi-
ronment in which he/she acts. The causal substrate of social mechanisms is the domain of facts
about intentional agents, socially situated in embodied social relations, that constitute the motive
power of social causation. This corresponds to the ontology of “methodological localism” and
actor-centered sociology I have developed elsewhere (Little 2006, 2009, 2014).
On this approach, the causal capacities of social entities are to be explained in terms of the
structuring of preferences, worldviews, information, emotions, incentives, and opportunities
for agents. The causal properties of a social entity inhere in its power to affect individuals’
behavior through incentives, preference-formation, belief-acquisition, or powers and oppor-
tunities. The micro-mechanism that conveys cause to effect is supplied by an account of the
actions of agents with specific goals, beliefs, and powers, within a specified set of institutions,
rules, and normative constraints.
The framework of social mechanisms as a basis for social explanation raises an important
question about the role and scope of generalizability that we can expect from a social explana-
tion. Briefly, the mechanisms identified by McAdam, Tarrow, and Tilly show a limited but real
415
Daniel Little
degree of generalizability; as they assert, social mechanisms can be expected to recur in other
circumstances and times. But the event itself is one of a kind. This is a familiar feature of Tilly’s
way of thinking about contentious episodes as well: the American Civil War was a singular
historical event. But a good explanation will invoke mechanisms that recur in other contentious
settings as well. We should not expect to find general theories of civil wars; but our explanations
of particular civil wars can invoke quasi-general theories of mid-level mechanisms of conflict,
mobilization, and escalation.
Are there any social mechanisms? In fact, a cursory survey of comparative sociology, political
science, and the new institutionalism provides a very large body of explanations that iden-
tify common social mechanisms available for historical explanations—for example, “collective
action problems often cause strikes to fail,” “increasing demand for a good causes prices to rise
for the good in a competitive market,” “transportation systems cause shifts of social activity and
habitation.” Here is a short but suggestive list: public goods problems (Hardin 1982); political
entrepreneurship (Bates 1981); principal-agent problems (Ensminger 1992); features of ethnic
or religious group mobilization (Hardin 1995); market mechanisms and failures (Akerlof 1970);
rent-seeking behavior (Seligson and Passé-Smith 1993); mechanisms of corruption (Klitgaard
1988); the social psychology of race (Steele and Aronson 1995). The literature now includes
examples of dozens of well-developed models of social mechanisms that can be theorized and
observed in concrete social settings.
These are all mechanisms that work at the level of socially situated actors within extended
institutions and organizations. They characterize one or more of the features of agency and
structure in concrete social circumstances. We understand how they work; individuals with
specified motivational and cognitive characteristics, placed within the context of the social set-
tings identified by the mechanisms, will behave in ways that bring about the outcome. And they
are abstract enough that they can be identified in a wide range of settings: the feudal manor,
the collective farm, the Wall Street law firm. In fact, we might say (along with Robert Merton
(1967)) that the most fundamental value of theories in the social sciences is the formulation of
models of mechanisms at this level—theories of the middle range.
We might attempt to classify mechanisms according to a variety of criteria, as Carl Craver
and Lindley Darden have done for biology (Craver and Darden 2013) and Andrew Bennett has
begun to do for international relations theory (Bennett 2014; see also Chapter 7, this volume).
One approach is to classify mechanisms according to what they do—the level of action and
process within which they operate and the kinds of things they influence or bring about (Little
2015). Some mechanisms influence individual behavior directly; others influence the formation
of the individual; others provide pathways of aggregation from individual to collective; and so
forth. These questions imply a handful of regions of activity for social mechanisms: for example,
mechanisms affecting the neuro-cognitive system; mechanisms affecting action and deliberation;
mechanisms affecting identity formation; mechanisms of institutional influence on individuals;
mechanisms of aggregation from individual to social; mechanisms occurring within the domain
of social action and collective action; mechanisms of hierarchy and control; and mechanisms of
higher-level causal influences. The key point is that mechanisms are heterogeneous, deriving
from different causal properties and powers, and playing a causal role in different parts of the
social whole. It is worthwhile to try to provide a preliminary way of classifying the mechanisms
that arise in social and historical explanations.
Emphasis on causal mechanisms for adequate social explanation has several beneficial effects
for historians. It provides a credible alternative to the covering-law model of explanation. It
takes us away from easy reliance on uncritical statistical models. It makes it apparent why his-
torians and social scientists benefit from collaboration. And it also leads us away from excessive
416
emphasis on large-scale classification of events into revolutions, democracies, or religions, and

toward more specific analysis of the processes and features that serve to discriminate among
instances of large social categories (Tilly 1995). This view parallels arguments offered by Stuart
Glennan on the “ephemeral” nature of historical causes (Glennan 2010).
3. Causal narratives
The idea of causal mechanisms fits very well into a characteristic mode of historical writing, the
form of a causal narrative. Essentially the idea is that a causal narrative of a complicated outcome
or occurrence is an orderly analysis of the sequence of events and the causal processes that con-
nected them, leading from a set of initial conditions to the outcome in question. The narrative
pulls together our best understanding of the causal relations, mechanisms, and conditions that
were involved in the process and arranges them in an appropriate temporal order. It is a series
of answers to “why and how did X occur?” designed to give us an understanding of the full
unfolding of the process. A narrative is more than an explanation; it is an attempt to “tell the
story” of a complicated outcome. So a causal narrative will include a number of causal claims,
intersecting in such a way as to explain the complex event or process that is of interest. A key
part of the justification of a causal narrative is the existence of well-developed empirical theories
of the social mechanisms that are invoked, and these theories emerge from several of the social
sciences. (There is a more extensive discussion of causal narratives in New Contributions to the
Philosophy of History (Little 2010).)1
We might illustrate this idea by looking at the approach taken to contentious episodes
and periods by McAdam, Tarrow, and Tilly (2001). In their treatment of various contentious
periods, they break the given complex period of contention into a number of mechanisms
and processes, conjoined with contingent and conjunctural occurrences that played a signifi-
cant causal role in the outcome. The explanatory work that their account provides occurs at
two levels: the discovery of a relatively small number of social mechanisms of contention that
recur across multiple cases, and the construction of complex narratives for particular episodes
that bring together their understanding of the mechanisms and processes that were in play in
this particular case.
The narrative for a particular case (the Mau Mau uprising, for example) takes the form of
a chronologically structured account of the mechanisms that their analysis identifies as having
been relevant in the unfolding of the insurgent movement and the government’s responses.
McAdam, Tarrow, and Tilly give attention to “episodes” within larger processes, with the clear
implication that the episodes are to some degree independent from each other and are amenable
to a mechanisms analysis themselves. So a narrative is both a concatenated series of episodes
and a nested set of mechanisms and processes. (Here again the idea of ephemeral mechanisms is
pertinent to the social case (Glennan 2010).)
In short, a narrative describes a particular process or event; but it does so by identifying
recurring processes, mechanisms, and forces that can be discerned within the unfolding of the
case. So generalizability comes into the story at the level of the components of the narrative—
the discovery of common social processes within the historically unique sequence of events.
4. Social variation and mechanisms

Variation within a social or historical phenomenon is all but ubiquitous in history and in the
contemporary world. Think of the Cultural Revolution in China, demographic transition in
early modern Europe, the ideology of a market society, or the experience of being black in the
417
Daniel Little
United States. We have the noun—“Cultural Revolution”—which can be explained or defined

in a sentence or two as an extended social phenomenon of mobilization and conflict that took
place in China from 1966–76; and we have the complex underlying social realities to which it
refers, spread out over many cities, villages, and communes across China (Esherick et al. 2006).
The Cultural Revolution consisted of a myriad of local and regional moments of political
mobilization and conflict, with complex relationships to national party organizations as well. So
the Cultural Revolution was in fact a highly heterogeneous social and political phenomenon.
This description focuses on locational variation in processes—village to village, country to
country. But social scientists often also highlight variations across social segments within a given
location: class, race, gender, religion, occupation. Do poor sharecroppers have a different fer-
tility profile over time than the wealthy in a particular region at a particular time? Are there
significant differences in survival strategies for distinct groups defined by race or ethnicity in a
city or a group of cities? Detailed investigations almost always reveal substantial variations across
social phenomena like these.
Careful study of the causal and social mechanisms giving rise to historical phenomena in
particular settings is a crucial means for understanding and explaining the facts of historical
variation. How did the activism and ideology of Cultural Revolution spread from Beijing to
Nanjing and other locations? How did activism spread from city to rural locations? How did
youth organizations mobilize their followers? How did local circumstances cause changes and
variations in the political movement? How much path dependency existed in the spread of
revolutionary ideas and strategies? The answers to all these questions most naturally take the
form of a sketch of one or more underlying social mechanisms.
This is another place where the appeal to social mechanisms can be seen once more to be
highly relevant and helpful to historical research. If we work on the assumption that any large
social process—the dispersed locations of contention associated with the Cultural Revolution,
say—is the compound result of a set of underlying causal social mechanisms, and if we hypoth-
esize that many of these mechanisms are in play in some places but not in others, then we can
explain both similarity and difference in the occurrence of the phenomenon across time and
place. Now the work of historical investigation can be put in these terms: identify some of the
social mechanisms that evidently recur in various locations; identify some of the mechanisms
that lead to significantly different results in some places; and identify some of the cross-location
mechanisms that are at work to secure a degree of synchrony and parallel in the developments
observed in different locations (communication systems, networks of leaders, dissemination of
activists). Case studies and comparative research permit both a degree of generalization and an
explanation of variation.
In other words, the intellectual strategy here is to disaggregate the large social factor into the
results of a larger number of underlying mechanisms; and then to attempt to discover how these
mechanisms played out differently in different settings throughout the range of the Cultural
Revolution, protoindustrialization, or ethnic conflict in South Asia. Significantly, as we have
seen above, this is exactly the strategy of research and explanation that Charles Tilly and his
colleagues were led to in their emphasis on discovering the component social mechanisms that
underlie social contention (McAdam et al. 2001).
5. Mechanisms and predictions?

To what extent is it possible to predict the course of large-scale history—the rise and fall
of empires, the occurrence of revolution, the crises of capitalism, or the ultimate failure of
twentieth-century Communism (see also Chapter 13)? One possible basis for predictions is
418
the availability of theories of underlying processes. To arrive at a supportable prediction about

a state of affairs, we might possess a theory of the dynamics of the situation, the mechanisms
and processes that interact to bring about subsequent states, and we might be able to model the
future effects of those mechanisms and processes. A biologist’s projection of the spread of a dis-
ease through an isolated population of birds is an example. Or, second, predictions might derive
from the discovery of robust trends of change in a given system, along with an argument about
how these trends will aggregate in the future. For example, we might observe that the popula-
tion density is rising in water-poor southern Arizona, and we might predict that there will be
severe water shortages in the region in a few decades. However, neither approach is promising
when it comes to large historical change. There are strong reasons for doubting the availability
of long-term predictions in history.
One reason for the failure of large-scale predictions about social systems is the complexity of
causal influences and interactions within the domain of social causation. We may be confident that
X causes Z when it occurs in isolated circumstances. But it may be that when U, V, and W are
present, the effect of X is unpredictable, because of the complex interactions and causal dynamics
of these other influences. This is one of the central findings of complexity studies—the unpredict-
ability of the interactions of multiple causal powers whose effects are non-linear. It is the feature of
social life that Roy Bhaskar designates as “open” rather than “closed” (Bhaskar 1975).
Another difficulty is the typical fact of path dependency of social processes. Outcomes are
importantly influenced by the particulars of the initial conditions, so simply having a good idea
of the forces and influences the system will experience over time does not tell us where it will
wind up. There are too many contingent causal influences along the way to allow us to argue
that a given future scenario is substantially more likely than dozens of others.
Third, social processes are sensitive to occurrences that are singular and idiosyncratic and not
themselves governed by systemic properties. If the winter of 1812 had not been exceptionally
cold, perhaps Napoleon’s march on Moscow might have succeeded, and the future political
course of Europe might have been substantially different. But variations in the weather are not
themselves systemically explicable—or at least not within the parameters of the social sciences.
Fourth, social events and outcomes are influenced by the actions of purposive actors. So it
is possible for a social group to undertake actions that avert the outcomes that are otherwise
predicted. Take climate change and rising ocean levels as an example. We may be able to predict
a substantial rise in ocean levels in the next fifty years, rendering existing coastal cities largely
uninhabitable. But what should we predict as a consequence of this fact? Societies may pursue
different strategies for evading the bad consequences of these climate changes—retreat, massive
water control projects, efforts at atmospheric engineering to reverse warming. And the social
consequences of each of these strategies are widely different. So the acknowledged fact of global
warming and rising ocean levels does not allow clear predictions about social development.
For these and other reasons, it is difficult to have any substantial confidence in predictions of
the large course of change that a society, cluster of institutions, or population will experience.
History is contingent, and there are always alternative pathways that might have been taken, and
history has no general plan. So, no grand predictions in history.
But then we have to ask a different question. What kinds of predictions or projections are
possible in history? Analysis of existing causal mechanisms does in fact provide a modest basis
for limited predictions about the future. Consider a few examples: labor unrest will intensify in
China in the next ten years; large technology disasters will occur in Europe; conflicts over land
use in East Jerusalem will continue to deepen in the coming decade. Each of these predictions
is credible because we can identify salient mechanisms leading to the outcome, and we can be
reasonably sure that they will play out as expected.
419
Daniel Little
Several things are apparent when we consider these predictions. First, they are limited in
scope; they have to do with small-scale features of the historical drama. Second, they depend
on specific and identifiable social circumstances, along with clear ideas about social mechanisms
connecting the present to the future. Third, they are at least by implication probabilistic; they
indicate likelihoods rather than inevitabilities. Fourth, they imply the existence of ceteris paribus
conditions: “Absent intervening factors, such-and-so is likely to occur.” But, finally, they all
appear to be intellectually justifiable. They may not be true, but they can be grounded in an
empirically and historically justified analysis of the mechanisms that produce social change, and
a model projecting the future effects of those mechanisms in combination.
The heart of prediction is our ability to identify dynamic processes and mechanisms that are
at work in the present, and our ability to project their effects into the future. Modest predictions
are those that single out fairly humdrum current processes in specific detail, and derive some
expectations about how these processes will play out in the relatively short run. Grand predic-
tions, on the other hand, purport to discover wide and encompassing patterns of development
and then to extrapolate their civilizational consequences over a very long period. A modest pre-
diction about China is the expectation that labor protest will intensify over the next ten years.
A grand prediction about China is that it will become the dominant economic and military
superpower of the late twenty-first century. We can have a fair degree of confidence in the first
type of prediction, whereas there are vastly too many possible branches in history, too many
countervailing tendencies, too many accidents and contingencies that may occur to give us any
confidence in the latter prediction.
Ceteris paribus conditions are unavoidable in formulating historical expectations about the
future, because social change is inherently complex and multi-causal. So even if it is the case
that a given process, accurately described in the present, creates a tendency for a certain kind
of result, it remains the case that there may well be other processes at work that will offset
this result. The tendency of powerful agents to seize opportunities for enhancing their wealth
through processes of urban development implies a certain kind of urban geography in the
future; but this outcome might be offset by a genuinely robust and sustained citizens’ movement
at the level of local politics.
The idea that historical predictions are generally probabilistic is partly a consequence of the
fact of the existence of unknown ceteris paribus conditions. But it is also, more fundamentally, a
consequence of the fact that social causation itself is almost always probabilistic. If we say that
rising conflict over important resources (X) is a cause of inter-group violence (Y), we don’t
mean that X is necessarily followed by Y; instead, we mean that X raises the likelihood of the
occurrence of Y (see Chapter 13).
So two conclusions seem justified. First, there is a perfectly valid intellectual role for mak-
ing historical predictions. Indeed, this is the basis of all policy reasoning. But these need to be
modest predictions: limited in scope, closely tied to theories of existing social mechanisms,
and accompanied by ceteris paribus conditions. And second, grand predictions should be treated
with great suspicion. At their best, they depend on identifying a few existing mechanisms and
processes; but the fact of multi-causal historical change, the fact of the compounding of uncer-
tainties, and the fact of the unpredictability of complex systems should all make us dubious
about large and immodest claims about the future.
6. Conclusion
Seventy years after Hempel’s classic article, the covering-law theory is now generally regarded as
a fundamentally wrong-headed way of thinking about historical (and social) explanation. Logical
420
positivism is not a convenient lens through which to examine the social and historical sciences.
There is too much contingency in the social world. Rather than being the result of law-governed
processes, social outcomes proceed from the contingent and historically variable features of the
actors and situations who make them. So the attention of many researchers interested in specify-
ing the nature of historical and social explanation has focused on social mechanisms constituted
and driven by common features of agency. This results in a different kind of explanation: accounts
of particular episodes that shed light on the causal processes that appear to have been involved in
their production, but no general accounts of large-scale historical patterns or outcomes.
Analysis of concrete causal mechanisms provides a basis for a degree of generality in histori-
cal inquiry. This is true in several ways. First, explanations based on social mechanisms can take
place in both a generalizing and a particular context. We can explain a group of similar social
outcomes by hypothesizing the workings of a common causal mechanism giving rise to them;
and we can explain a unique event by identifying the mechanisms that produced it in the given
unique circumstances. Second, a social-mechanism explanation relies on a degree of lawfulness;
but it refrains from the strong commitments of the deductive-nomological method. Mechanistic
causation is inherently conjunctural causation (Little 2000), in that a given outcome is almost
invariably the result of multiple causal influences and mechanisms. There are no high-level
social regularities. Third, we can refer both to particular individual mechanisms and a class of
similar mechanisms. For example, the situation of “easy access to valuable items along with low
probability of detection” constitutes a mechanism leading to pilferage and corruption. We can
invoke this mechanism to explain a particular instance of corrupt behavior—a specific group
of agents in a business who conspire to issue false invoices—or a mid-level general fact—the
logistics function of a large military organization which shows itself to be prone to repeated cor-
ruption. So mechanistic explanations support a degree of generalization across instances of social
activity while equally reflecting the facts of historical uniqueness and contingency.
Note
1 Robert Bates introduces a similar idea under the rubric of “analytic narrative” (1998). The chief differ-
ence between his notion and mine is that his account is limited to the use of game theory and rational
choice theory to provide the linkages within the chronological account, whereas I want to allow a plu-
ralistic understanding of the kinds and levels of causes that are relevant to social processes.
References
Abbott, A. 1998. “The causal devolution.” Sociological Methods and Research 27(2): 148–181.
Akerlof, G. 1970. “The market for ‘lemons’: quality, uncertainty and the market mechanism.” Quarterly
Journal of Economics 84: 488–500.
Bates, R. H. 1981. Markets and states in tropical Africa: the political basis of agricultural policies. Berkeley, CA:
University of California Press.
Bates, R. H. 1998. Analytic narratives. Princeton, NJ: Princeton University Press.
Bennett, A. 2014. “The mother of all isms: causal mechanisms and structured pluralism in international
relations theory.” European Journal of International Relations 9(3): 459–481.
Bhaskar, R. 1975. A realist theory of science. Leeds: Leeds Books.
Cartwright, N. 1999. The dappled world: a study of the boundaries of science. Cambridge: Cambridge University
Press.
Craver, C. F. and L. Darden. 2013. In search of mechanisms: discoveries across the life sciences. Chicago:
The University of Chicago Press.
Cummins, R. 2000. “‘How does it work?’ vs. ‘What are the laws?’: Two conceptions of psychologi-
cal explanation.” In Explanation and cognition. F. Keil and R. Wilson. Cambridge, MA: MIT Press:
117–144.
421
Daniel Little
Ensminger, J. 1992. Making a market: the institutional transformation of an African society. Cambridge:
Esherick, J., P. Pickowicz, and A. G. Walder. 2006. The Chinese cultural revolution as history: studies of the
Walter H. Shorenstein Asia-Pacific Research Center. Stanford, CA: Stanford University Press.
Glennan, S. 1996. “Mechanisms and the nature of causation.” Erkenntnis 44: 49–71.
Glennan, S. 2010. “Ephemeral mechanisms and historical explanation.” Erkenntnis 72: 251–266.
Gorski, P. 2013. “What is critical realism? And why should you care?” Contemporary Sociology 42(5):
658–670.
Gross, N. 2009. “A pragmatist theory of social mechanisms.” American Sociological Review 74(3): 358–379.
Hardin, R. 1982. Collective action. Baltimore: The Johns Hopkins University Press.
Hardin, R. 1995. One for all: the logic of group conflict. Princeton, NJ: Princeton University Press.
Hedström, P. 2005. Dissecting the social: on the principles of analytical sociology. Cambridge: Cambridge
University Press.
Hedström, P. and R. Swedberg. 1996 “Social mechanisms.” Acta Sociologica 39: 281–308.
Hempel, C. 1942. “The function of general laws in history.” Journal of Philosophy 39(2): 35–48.
Kincaid, H. 1990. “Defending laws in the social sciences.” Philosophy of the Social Sciences 20(1): 56–83.
Kincaid, H. 1996. Philosophical foundations of the social sciences: analyzing controversies in social research.
Klitgaard, R. E. 1988. Controlling corruption. Berkeley, CA: University of California Press.
Little, D. 1993. “On the scope and limits of generalizations in the social sciences.” Synthese 97: 183–207.
——. 2000. “Explaining large-scale historical change.” Philosophy of the Social Sciences 30(1): 89–112.
——. 2006. “Levels of the social.” In handbook for philosophy of anthropology and sociology. S. Turner and
M. Risjord. Amsterdam: Elsevier Publishing: 343–371.
——. 2009. “The heterogeneous social.” In Philosophy of the social sciences: philosophical theory and scientific
practice. C. Mantzavinos. Cambridge: Cambridge University Press: 154–178.
——. 2010. New contributions to the philosophy of history. Dordrecht: Springer Science.
——. 2011. “Causal mechanisms in the social realm.” In Causality in the sciences. P. Illari, F. Russo and
J. Williamson. Oxford: Oxford University Press. DOI: 10.1093/acprof:oso/9780199574131.003.0013.
——. 2014. “Actor-centered sociology and the new pragmatism.” In Individualism, holism, explanation and
emergence. J. Zahle and F. Collin. Dordrecht, London, New York: Springer: 55–75.
——. 2015. “Classifying mechanisms by location.” Understanding society. http://understandingsociety.
blogspot.com/2014/08/classifying-mechanisms-by-location.html (Accessed 29 March 2017).
Machamer, P., et al. 2000. “Thinking about mechanisms.” Philosophy of Science 67(1): 1–25.
Mahoney, J. 2001. “Beyond correlational analysis: recent innovations in theory and method.” Sociological
Forum 16(3): 575–593.
Mayntz, R. 2004. “Mechanisms in the analysis of social macro-phenomena.” Philosophy of the Social Sciences
34(2): 237–259.
McAdam, D., S. G. Tarrow, and C. Tilly. 2001. Dynamics of contention: Cambridge studies in contentious politics.
New York: Cambridge University Press.
McIntyre, L. C. 1996. Laws and explanation in the social sciences: defending a science of human behavior. Boulder,
CO: Westview Press.
Merton, R. K. 1967. On theoretical sociology. New York: Free Press.
Seligson, M. A. and J. T. Passé-Smith, Eds. 1993. Development and underdevelopment: the political economy of
inequality. Boulder, CO: L. Rienner Publishers.
Steel, D. 2004. “Social mechanisms and causal inference.” Philosophy of the Social Sciences 34(1): 55–78.
Steele, C. M. and J. Aronson. 1995. “Stereotype threat and the intellectual test performance of African
Americans.” Journal of Personality and Social Psychology 69(5): 797–811.
Steinmetz, G. 2004. “Odious comparisons: incommensurability, the case study, and ‘small N’s’ in sociology.”
Sociological Theory 22(3): 371–400.
Tilly, C. 1995. “To explain political processes.” American Journal of Sociology 100: 1594–1610.
Woodward, J. 2003. Making things happen: a theory of causal explanation. New York: Oxford University Press.
422
32
MECHANISMS IN
ECONOMICS1
Caterina Marchionni
1. Introduction
The market represents the paradigmatic example of an economic mechanism. Adam Smith
famously theorized it as functioning as if led by an invisible hand so as to satisfy the needs of
market participants. Over the years the market has been variously theorized as a mechanism
for resource allocation, price discovery, assignment of property rights, and many other things
besides (cf. Rosenbaum, 2000; Mirowski, 2007). At the same time, it has also come to be treated
more and more abstractedly and transported far from the economic domain to become a mecha-
nism for phenomena as diverse as mating behavior in animals, competition between churches,
and marriage choices—instances of a wider trend known as economics imperialism (Mäki, 2009a).
In spite of the centrality of the market mechanism, however, economics is not solely concerned
with market-related phenomena. In fact, a recurring theme of this chapter will be that econom-
ics is distinct from the other social sciences not so much by virtue of the kind of real-world
mechanisms (and phenomena) with which it deals, but because of the way in which mechanisms
are identified and analyzed. What I hope to highlight is the role that economists’ methodologi-
cal commitments play in defining what counts as a mechanism in economics.
In economics the term mechanism has various uses. Julian Reiss (2013, pp. 104–5) identifies
four different notions. The first is the econometricians’ notion of mechanisms as individual
causal relations. The contrast here is with mere correlation. The second refers to variables that
intervene between a cause and an effect and, as Reiss observes, it is often used in the context
of causal inference. The third takes mechanisms to be underlying structures or processes (for
example, the market), while the fourth takes mechanisms to be pieces of theory (for example,
a theoretical hypothesis showing the conditions under which the market clears). It is mainly
the last two notions, to which I will simply refer as mechanisms as underlying structures, that
come the closest to the conception of mechanisms advanced by current mechanistic philoso-
phers. It is also the one with which I will be mainly concerned in this chapter, even though,
as we will see, the notions of mechanisms as underlying structures and as intervening variables
are not always kept clearly separate in philosophical discussions about causal inference and
extrapolation (cf. Kincaid, 2004).
Mechanisms have been prominent in recent philosophical reflections on economics: they
have been claimed to provide justification for methodological individualism, to be necessary
423
Caterina Marchionni
for causal inference, and to aid extrapolation of causal claims from one context to another. In
what follows, after giving a characterization of how mechanisms are conceived and represented
in economics (sections 2 and 3), I discuss the alleged connection between mechanism and
methodological individualism (section 4), and the role of mechanisms in causal inference from
statistical data and in extrapolation (sections 5 and 6). Section 7 offers some concluding remarks.
2. What is a mechanism in economics?

Let us begin with a minimal definition proposed as a way of capturing the basic features of
mechanisms that contemporary mechanistic philosophers would agree on. I also take it to
characterize what Reiss calls mechanisms as underlying structures.
[M] A mechanism for a phenomenon consists of entities and activities organized in

such a way that they are responsible for the phenomenon.
(Illari and Williamson, 2012, p. 120; see also Chapter 1 and
Glennan, forthcoming)
Further features can be added to [M] to produce more specific accounts that restrict the scope
of what kinds of things qualify as mechanisms. There are two ways in which [M] can be
augmented to take into account the specificities of economics. The first concerns the kind of
mechanism economics deals with, whereas the second concerns how economists identify and
analyze mechanisms.
Dan Steel defines social mechanisms as follows:
[SM] Social mechanisms are complexes of interacting individuals, usually classified into
specific social categories, that generate causal relationships between aggregate-level
variables.
(Steel, 2004, p. 59)
Compared to [M], [SM] involves individuals as component parts, individuals that are
typically classified into social categories, such as buyers and sellers, fathers and daughters, and
who engage in certain kinds of activities (such as buying and selling, providing a dowry and
marrying) by virtue of the social roles they occupy. Starting from [SM], which is arguably a
general description including economic mechanisms as a subset, one obvious way to single
out economic mechanisms is by virtue of their being about particular kinds of social roles,
namely those pertaining to the market, or the economy more generally. This is, however, at
most only a tiny part of the story. Not only is economics concerned with phenomena that do
not clearly pertain to the economy, but the other social sciences are also interested in (say)
markets and market-related phenomena.
Another way of reformulating [SM] so as to take into account the specificity of econom-
ics is to include the type of assumptions economists make about the behavior of individuals,
namely those assumptions that derive from economists’ commitment to rational-choice theory.2
Although sometimes rational-choice theory is interpreted as being concerned exclusively with
individuals and their properties, it often presupposes structural and institutional facts, and its
“individuals” can also be firms, households, or organizations (see, for example, Kincaid, 1996;
Janssen, 1993). To capture the latter feature, let us replace individuals with rational agents in [SM].
This gives us the following characterization of mechanisms in economics:
424
Mechanisms in economics
[EM] Mechanisms in economics are complexes of rational agents, usually classified

into social categories, whose actions and interactions generate causal relationships
between aggregate-level variables.
[EM] defines mechanisms on the basis of the kind of entities that compose them, namely agents,
and the kind of properties ascribed to them, namely rational behavior. This is a descriptive (not
a normative) claim about what economists (typically) take mechanisms to be and does not entail
that this is what economic mechanisms really are. In what follows [EM] will be unpacked and
related to some of the main debates concerning mechanisms in economics.
3. Theoretical modeling of mechanisms

What distinguishes economics from other social sciences is not only the kind of mechanisms
economics deals in but also the way in which these mechanisms are studied. That is, mainly by
building and analyzing simple models of mechanisms described at a high level of abstraction.
This characteristic is captured by some of the most prominent accounts of models in econom-
ics. The connection between theoretical models and mechanisms features in Mäki’s account,
according to which economic models are means to isolate the operation of a mechanism from
the interference of other factors (see, for example, Mäki 1992, 2009b).3 Similarly, Cartwright
(2001, see also her 1989) claims that economic models are “blueprints for socio-economic
machines”: by theoretical means models create the right conditions for mechanisms to operate
unimpeded. Such conditions do not typically occur spontaneously in the real world, implying
that the disturbing factors that in the models were isolated away will affect the mechanism’s
operation. Of course, not all economic models aim at representing mechanisms; some are
better thought of as “phenomenological” models (see Chapter 17). Moreover, those models
that can be conceptualized as isolating mechanisms might not succeed in actually representing
any real-world mechanism.
In their modeling of mechanisms, economists also subscribe to a set of theoretical com-
mitments and desiderata, which contributes to setting the modeling approach of economics
apart from that of other sciences (Marchionni, 2013). First, the mechanistic requirement holds that
the phenomenon to be explained should be shown to result from a mechanism that fits [EM]
above. The legitimacy of the mechanistic requirement will be the topic of the next section. In
particular, we will see that different interpretations of this requirement have different degrees of
plausibility. A second desideratum economists emphasize relates to unification and requires that
the mechanism be derived from a unifying theory; that is, rational-choice theory. Economists’
commitment to rational-choice theory has been harshly criticized: rational-choice theory has
been found to be either empirically wanting, at least as a theory of individual behavior, or
empirically vacuous. Its credentials as a unifying theory are also suspicious. As Reiss (2013)
points out, its flexibility rather than its content account for its unifying power. This brings us to
the third desideratum that holds that, other things being equal, it is a good thing that the same
kind of mechanism is shown to account for many phenomena. Since scope is typically a posi-
tive function of the level of abstraction at which a mechanism is described, the desideratum of
generality leads to a preference for abstract descriptions of mechanisms.
Consider, for example, the application of Hotelling’s model—in which firms choose where
to locate spatially to maximize their market shares—to political parties choosing where to locate
themselves in the political space to maximize the number of votes (Kuorikoski, 2009; see also
Reiss, 2013). The main result of Hotelling’s model of spatial localization, according to which
425
Caterina Marchionni
firms will tend to locate close to one another, is also shown to account for the fact that political
parties tend toward the center of the political spectrum. It is the abstract “logic of the situation”
that is hypothesized to be similar, and hence to account for the similarity between the economic
and political phenomenon (Kuorikoski, 2009).4 This conception of mechanisms is compatible
with the characterization of mechanisms in economics [EM], which in turn is compatible with
the minimal definition [M]: in Hotelling’s model, the components are the firms (or the political
parties), who by virtue of their socio-economic roles perform activities that in interaction bring
about the phenomenon to be explained.
The ease with which abstract descriptions of mechanisms can be transferred across
domains has drawbacks. Very little might be inferred about the political market for votes on
the basis of the set of features it shares in common with standard markets, while at the same
time relevant features specific to each domain might be unduly ignored (Kuorikoski, 2009).
Furthermore, the similarity between “situations” does not automatically warrant infer-
ences about properties of the agents across domains (Kuorikoski, 2009). For example, if the
assumption of maximizing behavior might be justified by the selection pressures the market
exerts on firms, it is not necessarily the case that such behavior is legitimately attributed to
political parties if similar selection pressures are absent or are counteracted by other institu-
tional mechanisms. Finally, the strategy of building simple models of abstract mechanisms
is likely to pose limits on the kind of phenomena economists would succeed in explaining.
For some this is not a far-fetched possibility (see, for example, Lawson, 1997; Northcott and
Alexandrova, 2015).
4. Methodological individualism
Economists’ commitment to the doctrine of methodological individualism emerges with par-
ticular clarity from the belief that macroeconomics should be built on microeconomic foun-
dations (Janssen, 1993; Hoover, 2001). The philosophical debate on micro foundations and
methodological individualism has mainly concerned whether individual-level mechanisms are nec-
essary for economic explanation and/or whether explanations that do include individual-level
mechanisms are somehow better than purely macro-level explanations (Kincaid, 1996). To
make the discussion more concrete, let us use a stylized example originally presented in Jackson
and Pettit (1992), which I also discuss in Marchionni (2008).
Suppose that the phenomenon to be explained is an increase in the crime rate in a particular
neighborhood. Such an increase can be explained in two ways. An aggregate-level explanation
identifies a recent increase in the level of unemployment as the cause of the increase in crime
rate. An individual-level explanation instead would describe the changes in the opportunities
and motivations of particular individuals. As a thesis about explanation, methodological indi-
vidualism would hold either that the aggregate-level explanation alone does not explain or that
in any case the individual-level explanation is better.
There is a sense in which the aggregate-level explanation is deficient. What is missing
is a description of the mechanism connecting crime and unemployment. But does such a
mechanism always need to be at the individual level? Harold Kincaid (1996) has offered both
conceptual and empirical arguments against the claim that individual-level mechanisms are
necessary for explanation. In its strongest version, methodological individualism holds that
underlying mechanisms must only cite individuals and their properties. This is a non-starter,
however: as mentioned above, rational-choice theory, the allegedly individualist theory par
excellence, is not concerned only with individual behavior and often presupposes social kinds
(Janssen, 1993; Kincaid, 1996).
426
A weaker version of the argument linking methodological individualism, mechanism, and

explanation takes it that individual-level explanations are somehow better. This idea, too, has been
disputed. Compared to an aggregate-level explanation, the individual-level one, describing the
changes in opportunities and motivations of particular individuals, misses relevant information,
namely that irrespective of the behavior of particular individuals, an increase in unemployment
would have brought about an increase in the crime rate (Jackson and Pettit, 1992; Garfinkel,
1981; see also Kincaid, 1996). These arguments show that neither an exclusively individual-
level explanation nor an exclusively aggregate-level one is always to be preferred.5
Finally, an even weaker version of methodological individualism takes it that the compari-
son should not be between an explanation that simply relates the level of unemployment and
the crime rate and one that only cites individuals, their motivations, and actions (Coleman,
1990; Janssen, 1993, Chapter 30). Instead, the issue is whether the aggregate-level explanation
is improved by showing how an increase in the level of unemployment affects individuals and
how this in turn causes the crime rate to increase. For example, suppose now that the direction
of causality goes from crime rate to unemployment level and that the agents’ choice of whether
to engage in criminal activities as well as their opportunity to find a job is affected by the social
networks in which they interact (Granovetter, 1973). In particular, suppose that agents are more
likely to engage in criminal activities the more people around them do so, and are more likely to
find jobs the more people around them are actually employed. It follows that an increase in the
crime rate in a particular neighborhood makes it more likely for an individual to interact with
a criminal than with someone who is employed and therefore can provide information about
new jobs (Calvó-Armengol and Zenou, 2003, p. 71). This contributes to increasing the level
of unemployment, which in turn contributes to increasing the crime rate. This is a mechanistic
explanation, but the mechanism described is not a purely individual-level one. It describes how
an aggregate variable (crime rate) affects another aggregate variable (unemployment level) via
micro determinants (individuals’ job search) (see Figure 32.1 and also Chapter 30).6
This style of explanation is compatible with current mechanistic approaches to explana-
tion. Although (constitutive) mechanisms are at a lower level than the phenomenon to be
explained, levels of mechanisms do not map onto traditional compositional ones characterized
by mereological or aggregative relations (Chapter 14; Bechtel and Hamilton, 2007). Instead,
“X’s ϕ-ing is at a lower mechanistic level than S’s ψ-ing if and only if ϕ-ing is a component in
the mechanism for S’s ψ-ing” (Craver, 2007, p. 189). This means that philosophical accounts of
mechanistic explanation do not require that mechanisms in economics should be at the individ-
ual level. For some economic phenomena, the agents interacting in the mechanism can be firms,
organizations, whole countries, or even subpersonal entities (Kincaid, 2004). Furthermore, in a
mechanistic explanation the level of organization as well as the environment in which mecha-
nisms are embedded are also important—in our example, these are the networks of relations
Figure 32.1 Coleman’s boat. The mechanism connecting C (crime rate) to U (unemployment level),
where x and y represent individual-level variables and the arrows represent causal relations
427
Caterina Marchionni
in which individuals are embedded and changes in the level of unemployment. If economists’
mechanistic requirement is interpreted as demanding that economic phenomena be explained
by representing the multi-level mechanisms that bring them about, then such a requirement can
be justified along the lines proposed by current mechanistic philosophers.
5. Causal inference
Knowledge of mechanisms has been claimed to play a key role in making causal inferences from
statistical data more secure. It has even been suggested that to distinguish genuine causal relations
from mere correlations, knowledge of mechanisms is necessary (Elster, 1983). As an illustration,
let us return to the crime-unemployment example. The idea is that knowledge of a connecting
mechanism between the aggregate-level variables helps to identify whether the two variables
are in fact causally connected, and the direction of causality, or whether they are both effects of
a common unmeasured cause. What is under dispute is whether mechanisms are always neces-
sary to identify genuine causal relations in the context of non-experimental research. Note that
this claim has two interpretations: the quest for mechanisms can be interpreted as a quest for
individual-level mechanisms or for lower-level mechanisms more generally.
Kincaid (1996, pp. 179–82) advances two objections to the first interpretation of this claim;
that is, the necessity of individual-level mechanisms (see also Hoover, 2001; Reiss, 2008; Steel,
2004). The first is that there is no reason to stop at the level of individuals. If mechanisms are
necessary to confirm causal relations, why shouldn’t we go down the hierarchy of levels until
we reach the rock bottom? Clearly such regress does not help the claim that individual-level
mechanisms are necessary for causal inference in economics.7 The second objection is that
causal relationships can be identified with enough confidence through other means such as
randomized controlled trials or statistical techniques, and Kincaid offers a few examples of this.
Even if we agree with Kincaid and others that evidence of individual-level mechanisms is not
necessary for the confirmation of causal claims, it might still play a useful role. This is the posi-
tion Steel (2004, 2011) advocates. The key to appreciate how evidence of mechanisms can help
causal inference is to distinguish between direct and indirect casual inference (Steel, 2011). One of
Steel’s illustrations concerns the postulated causal relationship between the legalization of abortion
in 1973 in the US and the decline of the crime rate in the 1990s (Donohue and Levitt, 2001).
Information about the micro-level mechanism, from being an unwanted child to criminality, was
marshaled as further evidence of the causal link between legalization of abortion and decline of
the crime rate. Direct causal inference concerns the causal relationship between legalization of
abortion and decline of crime rates, whereas the indirect causal inference concerns the causal rela-
tionship between being an unwanted child and criminal behavior. The reason why evidence of
the causal relationship between unwanted childhood and criminal behavior is valuable is practical:
it concerns the fact that we might be in a position to make a stronger inference about the vari-
ables in the mechanism than about the variables in the original relation. In this example, the direct
casual inference concerns aggregate variables, whereas the indirect causal inference concerns the
individual-level variables, but presumably the same logic applies if the variables in the mechanism
were at the same level as those of the primary causal claim. Hence, I agree with Reiss that “it is
not necessary that the ‘mediating’ variable obtains at a lower level than the original cause and effect
variables” (2013, p. 104). I suspect the same logic applies more broadly in cases in which a primary
causal claim is supported by evidence of different kinds (Claveau, 2012; Staley, 2004). If so, then
it is unclear whether the relevant notion of mechanism here involves any form of reduction, not
even in the broad sense of underlying or constitutive structures.8
428
The strategy of abstraction and simple models discussed above constitutes the most common
source of mechanistic hypotheses in economics (see also Reiss, 2008, pp. 116–17).9 Assessing
the plausibility of these mechanistic hypotheses is ultimately an empirical matter. I agree with
Steel that there is no one set of methodologies that uniquely supplies mechanistic evidence,
which can be obtained by laboratory and field experimentation, or as in Donohue and Levitt’s
example, by correlational studies. In Donohue and Levitt’s study, evidence in favor of the
mechanistic hypothesis also came from studies in countries where for a time abortions had to
be approved by the government. In these studies, children born from women who were denied
the procedure were found to be more likely to engage in criminal behavior later on. Using such
evidence in support of the causal claim about legalization of abortion and decline of criminality
in the United States, however, involves a further inferential step—one concerning the relevant
similarity between the countries in which the evidence was obtained and the United States. This
is known as the problem of extrapolation, to which I now turn.
6. The problem of extrapolation

The problem of extrapolation concerns how to justify transporting causal claims from one
context, for example a laboratory experiment or one country, to another, the real world or
another country. Steel (2008) offers a comprehensive philosophical treatment of mechanism-
based extrapolation. Mechanism-based extrapolation crucially involves the deployment of a
methodology he calls comparative process tracing. The latter is a matter of comparing mechanisms
in the model and in the target by focusing on stages where background knowledge tells us there
are likely to be causally relevant differences and/or on those downstream stages where upstream
differences are likely to have left a mark (see Figure 32.2).
Steel is optimistic that comparative process tracing can be used to justify extrapolation in
biology, but he is more cautious with regard to economics (and social science more gener-
ally) for two reasons. The first is that for mechanism-based extrapolation to work, we need
mechanisms that are modular, but policy interventions on some part of a mechanism might
turn out to affect the mechanism’s overall structure. The second is that often there is uncer-
tainty about the mechanisms behind economic phenomena. Let us consider each problem in
turn, starting from the latter.
Figure 32.2 Comparative process tracing. Stages of the mechanisms leading from C (crime rate)
to U (unemployment) in the model and in the target. The arrows represent causal
relations, the dashed arrows represent relations to be inferred about the mechanism in
target, and the double lines represent differences and similarities. Adapted from Guala,
2010, p. 1074
429
Caterina Marchionni
First, the viability of mechanism-based extrapolation in economics, Steel (2008) argues, is

complicated by the uncertainty concerning what mechanisms are responsible for economic phe-
nomena. For example, in spite of sustained and systematic experimental study since its discovery
in the 1970s, preference reversal—“a behavioral tendency for the preference ordering of a pair
of alternatives to depend, in a predictable way, on the process used to elicit it” (Starmer, 2008,
p. 1)—still lacks a theoretical explanation. Yet, as Guala (2010) observes, uncertainty about the
cause of preference reversal only tells us that further work is needed, not that uncertainty is an
ineliminable and pervasive feature of economics experiments.
Second, that the effect of policy interventions on causal structure constitutes a problem for
economics is captured in the well-known Lucas critique, which states that since agents’ optimal
behavior often changes in response to policy changes, many macroeconomic forecasts, which
are based on assumptions about agents’ optimal behavior, are bound to fail (Lucas, 1976). More
generally, it might be the case that policy interventions on some part of a mechanism affect the
mechanism’s structure. There are two possible ways of addressing the difficulties posed by such
“structure-altering interventions”: an experimental and a theoretical one. At least in principle it
is possible to design an economic experiment in which the intervention would alter the struc-
ture in the same way as the policy intervention, if implemented, would (Guala, 2010, p. 1079).
The result of such an experiment could help us overcome the problem of structure-altering
interventions and hence deploy mechanism-based extrapolation. Since large-scale interventions
are hard to implement in the laboratory or in the field, and experimenting on a smaller scale
would still entail uncertainty about the effect of an implementation on a larger scale, Steel’s
concern with structure-altering interventions remains a practical problem.
The second route, suggested by Steel (2008, p. 158), is to rely on a more fundamental
theory to tell what kind of changes in causal structure the intervention is likely to produce. At
present the most likely candidate in economics for such a theory is rational-choice theory.10
Therefore, the issue turns on the appropriateness of rational-choice theory qua fundamental
theory. In particular, Steel calls upon recent results in experimental economics, which show
that small changes in variables that fall outside the domain of the theory have dramatic effects
on behavior, casting doubts about whether rational-choice theory can help in anticipating the
consequences of structure-altering interventions.
Steel’s last conclusion might be too hasty, however. It can be argued that experimental results
concerning individual behavior do not suffice to demonstrate that rational-choice theory cannot
be relied upon to anticipate changes in causal structure. Rational-choice theory indeed need
not be interpreted as a theory about individual behavior as such, but as a theory of individual
behavior in settings in which it is supported by the right institutional scaffolding (compare with
Satz and Ferejohn, 1994; Ross, 2014). Although the economics experiments that have attracted
the most attention are those that demonstrate the existence of behavioral anomalies, these are
not the only kind of experiments economists have been engaged with.
Santos (2007), for example, distinguishes between technological and behavioral experi-
ments. Behavioral experiments are aimed at investigating individual behavior and have often
been interpreted as yielding results that are at odds with rational-choice theory. By contrast,
technological experiments are aimed at investigating institutional (market) mechanisms. Used
as complements to the theoretical models developed in the field of mechanism design, tech-
nological experiments have guided many of the successful applications of game theory to the
design of real-world markets such as the Federal Communications Commission auctions for
the allocation of telecommunication licenses (Roth, 2002). Santos (2007) attributes the success
of technological experiments to the robustness of the relation between the designed institution
and aggregate outcomes to changes in the environment (most notably, preferences). In other
430
words, the market institution is so designed so as to ensure that the resulting actions are rational
and income maximizing.
Moreover, if what we need for mechanism-based extrapolation is information about the
causal structure in the source and some relevant information about the causal structure in the
target, there is no need for the mechanism to include individual-level variables—think, for
example, of a causal chain between aggregate-level variables. Hoover (2001, 2009) can be
interpreted as making a similar point when he argues that the Lucas critique does not neces-
sarily imply the necessity of individual-level mechanisms. The Lucas critique holds that for
estimated relationships to be stable across policy changes, those relationships need to capture
the deep parameters in the economy, where deep parameters refer to “the fundamental onto-
logical building blocks of the economy” (Hoover, 2009, p. 393). To go from here to the
indispensability of individual-level mechanisms requires the further assumption that those
parameters are necessarily micro, which is not obvious. As in the previous discussion about
causal inference, it is unclear whether reduction, or underlying constitutive structures, is
involved in the context of mechanism-based extrapolation. Rather, it seems that the broader
notions of mediating variables and causal chains might be sufficient. This is not a critique
of Steel’s account of extrapolation—its main insight still stands regardless of the notion of a
mechanism sufficient to get it off the ground. At this stage, the relevance of pointing at the
possibility that intervening steps need not be at a lower level than the phenomenon to be
explained only expands the range of cases in which extrapolation is legitimate.
7. Concluding remarks
Some of the main debates in the philosophy and methodology of economics are intertwined
with one or another use of the notion of mechanism. Specifically, we have examined two
such notions: mechanisms as (theoretical hypotheses about) underlying structures and mecha-
nisms as intervening variables. I have shown that at least some economic models aim at
capturing mechanisms conceived as abstract descriptions of how social-level phenomena
result from the actions and interactions of rational agents, paralleling the idea of mechanisms
as underlying structures. This modeling strategy is closely linked to economists’ commitment
to methodological individualism, even though on closer inspection mechanistic explanation
does not square well with strong versions of methodological individualism, nor are such
strong versions apt descriptions of the actual practice of economics. Furthermore, we have
seen that although mechanisms are held to aid both causal inference and extrapolation, it is
not always clear whether the relevant notion of mechanism at stake is that of mechanisms as
underlying causal structures or as intervening variables.
It might then be that the centrality of the notion of mechanism is a product of the flex-
ibility with which the notion itself is used rather than the role it actually plays in economics. If
so, the traction of mechanistic ideas would improve by further clarity about what notion is at
stake in a particular case. This is not to claim that the focus on mechanism in the philosophy of
economics has had no value. On the contrary, reframing, for example, the age-old issue about
methodological individualism in terms of mechanistic explanation contributed to make clear
that although explaining by mechanisms is valuable, there is no reason to suppose that such
mechanisms should be exclusively at the individual level. Similarly, attention to mechanisms has
brought to the fore the function that different kinds of evidence have in both causal inference
and extrapolation and, hence, the importance of different methods of generating evidence. This
is especially topical now that the toolkit of empirical methodologies at the disposal of econo-
mists has been expanding.
431
Caterina Marchionni
Notes
1 I would like to thank the editors, Phyllis Illari and Stuart Glennan, and my colleagues Emrah
Aydinonat, Jaakko Kuorikoski, Luis Mireles Flores, Samuli Pöyhönen, and Petri Ylikoski, for their
helpful suggestions. All remaining mistakes are obviously mine.
2 According to some, the assumption that rational agents respond to incentives suffices to characterize the
core idea behind economics. Landsburg (1993, p. 3), for example, writes: “Most of economics can be
summarized in four words: ‘People respond to incentives.’ The rest is commentary.”
3 Some commentators have claimed that economic model building is in fact at odds with a mechanistic
world picture (e.g. Lawson, 1997). In rough outline, the criticism is that mainstream economics is com-
mitted to a form of deductivism, which presupposes that only events and regularities between them
exist, and is unsuitable for the discovery of mechanisms.
4 “The logic of the situation” is part of Popper’s method of situational analysis (1994), according to which
explanation in social science proceeds by describing the situation the agent is in. Coupled with the
rationality principle, the logic of the situation “explains” the agent’s behavior.
5 Note that in this case the individual-level explanation involves particular individuals rather than types
of individuals (Mäki, 2002).
6 Some wonder whether the explanatory strategy exemplified by Coleman’s boat qualifies as a form
of methodological individualism in that it allows for the effect of social structure. To distinguish this
variety of methodological individualism from its stronger versions, some refer to it as “structural indi-
vidualism” (see, for example, Udéhn, 2002).
7 Kincaid’s objection succeeds only if mechanistic ideas are wedded to reductionism. Recent mecha-
nistic accounts, however, are not reductionistic in this sense. A mechanism is always a mechanism for a
phenomenon and that sets the boundaries of the relevant mechanism. Whether statistical evidence is
sufficient to establish causation, however, is an object of current philosophical debate (see, for example,
Russo and Williamson, 2007; Howick, 2011).
8 Kincaid (2004) distinguishes horizontal and vertical mechanisms, where the former are intervening vari-
ables, whereas the latter are the micro-processes that bring about the macro effect. Yet an underlying
causal structure need not be a micro-process but can also involve social entities and activities that
account for the system having a certain property or exhibiting a certain behavior. Similarly, Craver
(2007) distinguishes etiological and constitutive mechanisms. A complete description of an etiological
mechanism represents the entities, their activities and form of organization, a requirement that does not
apply in the case of mechanisms as intervening variables. A clarification of the relation between these
distinct yet similar notions is a topic for further research.
9 Steel (2008) advocates the method of process tracing, which consists in presenting evidence for the
existence of several social practices that, when linked together, produce a chain of causation from one
variable to another. The relation between process tracing and the type of mechanistic evidence impli-
cated in indirect causal inference is not fully spelled out. However, since I take mechanistic evidence to
be the more general category, in this chapter I only use this terminology.
10 Rational-choice theory is not the only possible fundamental theory for the social sciences. As I interpret
it, however, Steel’s point is that current alternatives have not yet reached the status currently enjoyed by
rational-choice theory.This does not rule out the possibility that an alternative micro-theory may be devel-
oped that performs better in predicting the effects of structure-altering interventions (Steel, 2008, p. 160).
References
Bechtel, W & Hamilton, A 2007, “Reduction, integration, and the unity of science: natural, behavioral,
and social sciences and the humanities”, in TAF Kuipers (ed.), General philosophy of science: focal issues,
Elsevier, New York, pp. 377–430.
Calvó-Armengol, A & Zenou, Y 2003, “Does crime affect unemployment? The role of social networks”,
Annales d’Economie et de Statistique vol. 71–2, pp. 173–88.
Cartwright, N 1989, Nature’s capacities and their measurement, Clarendon Press, Oxford.
Cartwright, N 2001, “Ceteris paribus laws and socio-economic machines”, in U Mäki (ed.), The economic
world view, Cambridge University Press, Cambridge, pp. 275–92.
Claveau, F 2012, “The Russo-Williamson Theses in the social sciences: Casual inference drawing on
two types of evidence”, Studies in History and Philosophy of Biological and Biomedical Sciences vol. 43,
pp. 806–13.
432
Coleman, JS 1990, Foundations of social theory, Harvard University Press, Cambridge, MA.
Craver, C 2007, Explaining the brain, Oxford University Press, Oxford.
Donohue, J & Levitt, S 2001, “The impact of legalized abortion on crime”, Quarterly Journal of Economics,
vol. 116, no. 2, pp. 379–420.
Elster, J 1983, Explaining technical change: a case study in the philosophy of science. Cambridge University Press,
Cambridge.
Garfinkel, A 1981, Forms of explanation, Yale University Press, New Haven, CT.
Glennan, S forthcoming, The new mechanical philosophy, Oxford University Press, Oxford.
Granovetter, M 1973, “The strength of weak ties”, American Journal of Sociology vol. 78, no. 6, pp. 1360–80.
Guala, F 2010, “Extrapolation, analogy, and comparative process tracing”, Philosophy of Science vol. 77,
no. 5, pp. 1070–82.
Hoover, K 2001, Causality in macroeconomics, Cambridge University Press, Cambridge.
Hoover, K 2009, “Microfoundations and the ontology of macroeconomics”, in H Kincaid &
D Ross (eds) The Oxford handbook of philosophy of economics, Oxford University Press, New York,
pp. 386–409.
Howick, J 2011, “Exposing the vanities—and a qualified defense—of mechanistic reasoning in health care
decision making”, Philosophy of Science vol. 78, pp. 926–40.
Illari, P & Williamson, J 2012, “What is a mechanism? Thinking about mechanisms across the sciences”,
European Journal for Philosophy of Science, vol. 2, pp. 119–35.
Jackson, F & Pettit, P 1992, “In defense of explanatory ecumenism”, Economics and Philosophy, vol. 8, 1–21.
Janssen, M 1993, Micro-foundations: a critical inquiry, Routledge, London.
Kincaid, H 1996, Philosophical foundations of the social sciences, Cambridge University Press, New York.
Kincaid, H 2004, “Contextualism, explanation and the social sciences”, Philosophical Explorations, vol. 7,
no. 3, pp. 201–18.
Kuorikoski, J 2009, “Two concepts of mechanism: componential causal system and abstract form of inter-
action”, International Studies in the Philosophy of Science, vol. 23, no. 2, 143–60.
Landsburg, S 1993, The armchair economist: economics and everyday life, Free Press, New York.
Lawson, T 1997, Economics and reality, Routledge, London.
Lucas, RE 1976, “Econometric policy evaluation: a critique”, in K Brunner & A Meltzer (eds.) The Phillips
curve and the labor market, vol. 1, Carnegie-Rochester Conference on Public Policy, Amsterdam, North
Holland, pp. 19–46.
Mäki, U 1992, “On the method of isolation in economics”, Poznan Studies in the Philosophy of the Sciences
and the Humanities, vol. 38, pp. 147–68.
Mäki, U 2002, “Explanatory ecumenism and economics imperialism”, Economics and Philosophy, vol. 18,
pp. 235–57.
Mäki, U 2009a, “Economics imperialism: concept and constraints”, Philosophy of the Social Sciences, vol. 39,
no. 3, pp. 351–80.
Mäki, U 2009b, “MISSing the world: models as isolations and credible surrogate systems”, Erkenntnis
vol. 70, no. 1, pp. 29–43.
Marchionni, C 2008, “Explanatory pluralism and complementarity: from autonomy to integration”,
Philosophy of the Social Sciences, vol. 38, pp. 314–33.
Marchionni, C 2013, “Playing with networks: how economists explain”, European Journal for Philosophy of
Science, vol. 3, no. 3, pp. 331–52.
Mirowski, P 2007, “Markets come to bits: evolution, computation and markomata in economic science”,
Journal of Economic Behavior and Organization, vol. 63, pp. 209–42.
Northcott, R & Alexandrova, A 2015, “Prisoner’s dilemma doesn’t explain much”, in M Peterson (ed.),
The prisoner’s dilemma, Cambridge University Press, Cambridge, pp. 64–84.
Popper, K 1994, The myth of the framework, Routledge, London.
Reiss, J 2008, Error in economics, Routledge, London.
Reiss, J 2013, Philosophy of economics: a contemporary introduction, Routledge, London.
Rosenbaum, E-F 2000, “What is a market? On the methodology of a contested concept”, Review of Social
Economy, vol. 58, no. 4, pp. 455–82.
Ross, D 2014, Philosophy of economics, Palgrave Macmillan, New York.
Roth, A 2002, “The economist as engineer: game theory, experimentation, and computation as tools for
design economics”, Econometrica, vol. 70, no. 4, pp. 1341–78.
Russo, F & Williamson, J 2007, “Interpreting causality in the health sciences”, International Studies in the
433
Caterina Marchionni
Santos, A 2007, “The ‘materials’ of experimental economics: technological versus behavioral experiments”,
Journal of Economic Methodology, vol. 14, no. 3, pp. 311–37.
Satz, D & Ferejohn, J 1994, “Rational choice and social theory”, Journal of Philosophy, vol. 91, no. 2,
pp. 71–87.
Staley, K 2004, “Robust evidence and secure evidence claims”, Philosophy of Science, vol. 71, no. 4,
pp. 467–88.
Starmer, C 2008, “Preference reversals”, in SN Durlauf & LE Blume (eds.), The new Palgrave dictionary of
economics online, 2nd edn, Palgrave Macmillan, London.
Steel, D 2004, “Social mechanisms and causal inference”, Philosophy of the Social Sciences, vol. 34, pp. 55–78.
Steel, D 2008, Across the boundaries: extrapolation in biology and the social sciences, Oxford University Press,
New York.
Steel, D 2011, “Causality, causal models, and social mechanisms”, in IC Jarvie & J Zamora-Bonilla (eds),
The SAGE handbook of the philosophy of the social sciences, SAGE Publications, London, pp. 288–304.
Steel, D 2013, “Mechanisms and extrapolation in the abortion-crime controversy”, in H-K Chao,
S-T Chen & R Millstein (eds.), Mechanisms and causality in biology and economics, Springer, New York,
pp. 185–206.
Udéhn, L 2002, “The changing face of methodological individualism”, Annual Review of Sociology, vol. 28,
pp. 479–507.
Ylikoski, P 2012, “Micro, macro, and mechanisms”, in H Kincaid (ed.), The Oxford handbook of philosophy
of the social sciences, Oxford University Press, Oxford, pp. 21–45.
434
33
COMPUTATIONAL
MECHANISMS1
Gualtiero Piccinini
For a long time, computation was seen as making the notion of mechanism precise. Computation
explicates mechanism. More recently, the notion of mechanism articulated by philosophers of
science has been applied to shed light on computational systems. Mechanism explicates compu-
tation. This chapter recounts how this shift came about.
1. The mathematical theory of computation

The notion of computation comes from mathematics. To a first approximation, math-
ematical computation is the solving of mathematical problems by following an algorithm.
A classic example is Euclid’s algorithm for finding the greatest common divisor to two
numbers. Because algorithms solve problems automatically within finitely many steps, they
are sometimes called “mechanical” procedures. This is a first hint that computation has
something to do with mechanisms.
In the early twentieth century, mathematicians and logicians developed formal logic to
answer questions about the foundations of mathematics. One of these questions was the
decision problem for first-order logic; namely, the question of whether an algorithm can
determine, for any given statement written in a first-order logical system, whether that state-
ment is provable within the system.2 In 1936, Alonzo Church and Alan Turing proved that
the answer is negative—there is no algorithm solving the decision problem for first-order
logic (Church 1936; Turing 1936–7).
Turing’s proof is especially relevant because it appeals to machines. To show that no algo-
rithm can solve the decision problem for first-order logic, Turing needed to make precise what
counts as an algorithm. He did so in terms of simple machines for manipulating symbols on
an unbounded tape divided into squares. These Turing machines, as they are now called, have a
mobile control device that travels along the tape and reads, writes, and erases symbols on each
square of the tape—one square at a time. The control device can be in one of a finite number
of states. It determines what to do based on its state as well as what’s on the tape, in accordance
with a finite set of instructions.
Turing established three striking conclusions. First, he argued persuasively that any algorithm,
suitably encoded, can be followed by one of his machines—this conclusion is now known as the
Church–Turing thesis. Second, Turing showed how to construct universal Turing machines—special
435
Gualtiero Piccinini
Turing machines that can be used to simulate any other Turing machine by executing the relevant
instructions, which are written on their tape along with data for the computation. Finally, Turing
proved that none of his machines solves the decision problem for first-order logic.
Turing’s last result may be generalized as follows. There are denumerably many Turing
machines (and therefore algorithms). That is to say, you can list all the Turing machines one
after the other and put a natural number next to each. You can do this because you can enu-
merate all the finite lists of instructions that determine the behavior of each Turing machine.
Each Turing machine computes a function from a denumerable domain to a denumerable range
of values—for instance, from (numerals representing) natural numbers to (numerals represent-
ing) natural numbers. It turns out that there are undenumerably many such functions. Since an
undenumerable infinity is much larger than a denumerable one, there are many more functions
than Turing machines. Therefore, most functions are not computable by Turing machines. The
decision problem for first-order logic is just one of these functions, which cannot be computed
by Turing machines. The functions computable by Turing machines are called Turing-computable;
the rest are called Turing-uncomputable. The existence of Turing-uncomputable functions over
denumerable domains is theoretically very important. We will run into it again.
2. Mechanisms as Turing machines?

A few years after Turing’s paper came out, Warren McCulloch and Walter Pitts argued that
brains were a kind of Turing machine (without tapes). More specifically, McCulloch and Pitts
defined circuits of simplified neurons for performing logical operations such as AND, OR, and
NOT on digital inputs and outputs (strings of “ones” and “zeros”). They showed how to build
neural networks out of such circuits and argued that their networks are equivalent to Turing
machines; they also suggested that brains work similarly enough to their artificial neural net-
works, and therefore to Turing machines (McCulloch and Pitts 1943).3
Shortly after that, John von Neumann offered a sweeping interpretation of McCulloch and
Pitts’s work:
The McCulloch–Pitts result . . . proves that anything that can be exhaustively and
unambiguously described, anything that can be completely and unambiguously put
into words, is ipso facto realizable by a suitable finite neural network.
(von Neumann 1951: 22)
Since McCulloch–Pitts networks are computationally equivalent to Turing machines (with

bounded tapes), von Neumann’s statement entails that anything that can be exhaustively and
unambiguously described is realizable by a Turing machine. It expands the scope of the Church–
Turing thesis from covering mathematical algorithms to covering anything that can be described
“exhaustively and unambiguously.”
Thus was born the idea that Turing machines are not just a special kind of machine for per-
forming computations. According to von Neumann’s broader interpretation, Turing machines
are a model of anything exhaustively and unambiguously described. Insofar as mechanisms can
be exhaustively and unambiguously described, Turing machines are a model of mechanisms.
Insofar as physical systems can be exhaustively and unambiguously described, Turing machines
are a model of physical systems. (Insofar as physical systems are mechanisms, the previous two
statements are equivalent.)
Around the time that von Neumann was making his bold statement, the first digital com-
puters were being designed, built, and put to use. Digital computers are computationally
436
Computational mechanisms
equivalent to universal Turing machines (until they run out of memory). The main use of early
computers was to run computational simulations of physical systems whose dynamics was too
complex to be solved analytically. This sort of computational model became ubiquitous in many
sciences. Computers became a tool for simulating just about any physical system—provided,
of course, that their behavior could be described—in von Neumann’s words—“exhaustively
and unambiguously.”
Aided by the spread and popularity of computational models, views along the lines of von
Neumann’s hype about what McCulloch and Pitts had allegedly shown became influential.
One development was the widespread impression that minds must be computable in
Turing’s sense. After all, McCulloch and Pitts had allegedly shown that brains are basically
Turing machines. The theory that minds are computer programs running on neural software
was soon to emerge from this milieu (Putnam 1967; Chapter 6, this volume).
Closely related is the idea that explaining clearly how a system produces a behavior requires
providing a computer program for that behavior or, at least, some kind of computational
model. Sometimes, this view was explicitly framed in terms of mechanisms—to the effect that
explaining a behavior mechanistically requires providing a computer program for that behavior
(e.g., Dennett 1978: 83). Conversely, some theorists argued that, if human beings can behave in
a way that is Turing-uncomputable, then the mind is not a machine (e.g., Lucas 1961).
Other theorists made a similar point using different terminology. They distinguished
between the functions being computed and the mechanisms implementing the computation.
They argued that the functions performed by a system ought to be explained by functional
analysis—a kind of explanation that describes the functions and sub-functions being performed
by a system while abstracting away from the implementing mechanisms. For example, multi-
plying numbers (by the method of partial products) is functionally analyzed into performing
single-digit multiplications, writing the results so as to form partial products, and then adding
the partial products. Since functional analysis abstracts away from mechanisms, allegedly it
is distinct and autonomous from mechanistic explanation. These theorists often maintained
that psychology provides functional analyses, while neuroscience studies the implementing
mechanisms—therefore, psychological explanations are distinct and autonomous from neuro-
scientific explanations (see Chapter 29). These same theorists also maintained that functional
analyses consist in a computer program or other computational model (Fodor 1968; Cummins
1983). As a result, again, explaining a behavior requires providing a computer program or
other computational model for that behavior.
Even more grandly, von Neumann’s statement prefigured pancomputationalism—the view
that every physical system is computational (e.g., Putnam 1967), that the entire universe is a giant
computing machine (Zuse 1970), or that computation is somehow the building block of the
physical world (e.g., Fredkin 1990; Wheeler 1982). Pancomputationalism took on a life of its
own within certain physics circles, where it became known as digital physics or “it from bit”
(Wheeler 1990) and—after the birth of quantum computation—transmuted itself into the view
that the universe is a quantum computer (Lloyd 2006).4
3. Physical computation
All this talk of computation, especially in psychology and philosophy of mind, raised the ques-
tion of what it takes for a physical system to perform a computation. Is the brain a computing
system? Is the mind the software of the brain? What about other physical systems: which of them
perform computations? For these questions to be answerable, some account must be given of
what it takes for a brain or other physical system to perform a computation.
437
Gualtiero Piccinini
One early and influential account of physical computation asserted that a physical system
performs a computation just in case there is a mapping between the computational state transi-
tions that constitute that computation and the physical state transitions that the system undergoes
(Putnam 1967). The main appeal of this simple mapping account is that it’s simple and intuitive.
Eventually it became clear that it’s far too weak.
The main problem with the simple mapping account is that, without further constraints,
it’s too easy to map a computation to a series of physical state transitions. Ordinary physical
descriptions define trajectories with undenumerably many state transitions, whereas classical
computational descriptions such as Turing machine programs define trajectories with denumer-
ably many state transitions. If nothing more than a mapping is required, any denumerable series
of state transition can be mapped onto any undenumerable series of state transitions. Thus, by
the simple mapping account, any computation is implemented by any ordinary physical system
(Putnam 1988; Searle 1992). This result is now known as unlimited pancomputationalism; it is
widely seen as a reductio ad absurdum of the simple mapping account of physical computation.
To avoid unlimited pancomputationalism, several theorists introduced constraints on which
mappings are acceptable. Perhaps the most popular constraint is a causal one. According to
the causal account of physical computation, only mappings that respect the causal relations
between the computational state transitions—including those that are not instantiated in a
given computation—are acceptable (Chalmers 2011; Chrisley 1994; Scheutz 1999). In other
words, mapping individual computations onto individual physical state space trajectories is
insufficient. What is also needed is that the mapping be set up so that the physical states that
correspond to the computational states stand in appropriate causal relations with one another,
mirroring the causal relations implicitly defined by the computational description.
The main advantage of sophisticated mapping accounts, such as the causal account, is that
they remain close to the simplicity of the mapping account while avoiding unlimited pan-
computationalism. The main disadvantage of these accounts is that they remain committed to
limited pancomputationalism—the view that everything performs some computation or another.
Limited pancomputationalism is in line with von Neumann’s sweeping interpretation of the
Church–Turing thesis, but it is in tension with a common way of understanding the computa-
tional theory of mind.
According to many theorists (following Fodor 1968, 1975), the mind (or at least cognition)
is explained computationally in a way that other physical processes are not. Specifically, the
computational explanation of mind has to do with the manipulation of representations. Another
popular family of accounts of physical computation—semantic accounts—is tailored to the per-
ceived needs of this representational version of the computational theory of mind. According
to semantic accounts, physical computation is the manipulation of representations—or, more
precisely, the manipulation of certain kinds of representation in appropriate ways (Cummins
1983; Fodor 1975; Pylyshyn 1984; Shagrir 2006).
The greatest advantage of semantic accounts is that they can avoid all forms of pancompu-
tationalism. In fact, semantic accounts restrict physical computation to those physical systems
that manipulate the right kinds of representation in the right way—presumably, only minds and
artificial computing systems qualify. Semantic accounts also come with a disadvantage, though:
they make no sense of any (alleged) computational system that does not manipulate repre-
sentations. It turns out that, following computability theory, there is no difficulty in defining
computations that manipulate meaningless letters. It is just as easy to define a Turing machine
that manipulates uninterpreted marks as it is to define one that manipulates meaningful symbols.
It is equally easy to program a digital computer to sort meaningless marks into meaningless
orders as it is to program it to manipulate meaningful symbols in meaningful ways. Semantic
438
accounts cannot account for this. Later we will see that the difficulties of both mapping and
semantic accounts led to the most recent accounts of physical computation: the mechanistic
accounts. But first we need to prepare the terrain by questioning some of the assumptions that
are built into the traditional understanding of physical computation.
4. The scope of the Church–Turing thesis

Given the hype about computability that we reviewed, eventually scholars began to
investigate its history and proper scope. One of the earliest, rigorous investigations of this
sort was Guglielmo Tamburrini’s PhD dissertation (Tamburrini 1988), which was soon fol-
lowed by a series of incisive articles by his advisor, Wilfried Sieg (e.g., Sieg 1994, 1999).
Such investigations reconstructed the intellectual context of Church and Turing’s work—
which, as I mentioned above, was the foundations of mathematics. They also clarified
the interpretation and conceptual foundation of Turing’s results. They showed that the
Church–Turing thesis, properly understood, is about what mathematical problems can be
solved by following algorithms—not about what the mind in particular or physical systems
in general can do. The latter questions are independent of the original Church–Turing
thesis and must be answered by other means.
With this more rigorous and restrictive understanding of the Church–Turing thesis in the
background, Jack Copeland (2000) argued that we should distinguish between a mechanism
in the broad, generic sense and a mechanism in the sense of a procedure that computes a
Turing-computable function. A mechanism in the broad, generic sense is a system of organ-
ized components—this is the notion behind the new mechanism in philosophy of science,
which we will discuss in the next section. Whether a mechanism in the broad, generic sense
computes a Turing-computable function or a Turing-uncomputable function has nothing to
do with the original Church–Turing thesis.
Copeland pointed out that, at least in mathematics, there are hypothetical machines that
are more powerful than Turing machines—they compute Turing-uncomputable functions
(Turing 1939). Copeland called a machine that is computationally more powerful than a Turing
machine a hypercomputer. Finally, Copeland suggested that whether the brain or another physical
system is a regular computing system or a hypercomputer is an empirical question—it is possible
that brains are hypercomputers or that artificial hypercomputers can be built. Given Copeland’s
argument, the mind may be mechanistic in the generic sense even if it does something that is
Turing-uncomputable.
What about the view that something is mechanical if and only if it is computable by some
Turing machine? This kind of view relies on the notion of “mechanical” used in logic, whereby
algorithms are said to be mechanical procedures. Since Turing machines are formal counter-
parts to algorithms, any procedure that is not computable by Turing machines is ipso facto
not mechanical in that sense. As we’ve seen, this reasonable conclusion was fallaciously used
to imply that the human mind is mechanical or mechanistic (in an unspecified sense) if and
only if it is equivalent to a Turing machine. Copeland dubbed this kind of inference—from
the Church–Turing thesis to conclusions about whether minds are mechanistic or Turing-
computable—the Church–Turing fallacy (Copeland 2000, Piccinini 2007a).
Parallel to and independently of Copeland’s argument, a literature on physical hypercompu-
tation began to emerge within the foundations of physics. Various philosophers and physicists
proposed hypothetical means by which exotic physical systems may be able to compute
Turing-uncomputable functions. A prominent example involves a receiver system traveling
toward a huge, rotating black hole while receiving signals on the results of a Turing machine’s
439
Gualtiero Piccinini
computation while the Turing machine orbits the black hole. Because of the way the black
hole distorts space-time, the Turing machine may be able to complete an infinite number of
operations while the receiver system that launched the Turing machine traverses a finite time-
like trajectory; this allows the receiver to compute a Turing-uncomputable function, at least
for one value of the function (Hogarth 1994; Etesi and Németi 2002). Even though there is
little indication that hypercomputation is feasible, this literature as well as Copeland’s argument
undermine the idea that Turing machines are a general model of mechanisms. Still, none of this
directly challenges (limited) pancomputationalism in a broader sense. It may still be that every
physical system is computational—only now some physical systems may compute functions
that no Turing machine can compute.
5. The rise of the new mechanism

Around the same time that Copeland and others were defending the physical possibility of
hypercomputation, another group of philosophers of science revived the view that constitu-
tive explanation—explanation of what a system does in terms of what its organized subsystems
do—is provided by mechanisms (Chapter 1). These philosophers became known as the new
mechanists. They argued that many special sciences, including biology and neuroscience, explain
phenomena by uncovering mechanisms. What they meant by “mechanism” is what Copeland
meant: an organized system of components and activities such that the components and activi-
ties, organized the way they are, produce the phenomenon (Bechtel and Richardson 1993;
Glennan 1996; Machamer, Darden, and Craver 2000).
A few points about the new mechanism are relevant here.
First, the notion of mechanism articulated by the new mechanists is grounded in the expla-
nations provided by the special sciences. The new mechanists scrutinized scientific practices in
sciences such as molecular biology and neuroscience. They argued that such sciences explain
phenomena mechanistically. This raises the question of whether other sciences—including, say,
computer science and computer engineering—explain mechanistically.
Second, the new mechanists differed slightly in the way they conceptualized mechanisms. For
example, some talked about mechanisms performing operations, others of mechanisms perform-
ing activities, yet others of mechanisms performing functions. Those differences don’t matter for
our purposes. What matters is that none of these theorists conceptualized mechanisms in terms
of computation. The view examined in previous sections, whereby a mechanism is something
whose behavior can be modeled by a Turing machine, was absent from the new mechanism.
Mechanisms were systems of concrete components organized to perform operations, activities,
or functions. (Elsewhere, I argue that these formulations are essentially equivalent anyway;
Piccinini forthcoming, cf. Illari and Williamson 2012.)
Third, even when mechanists talked about functions, they were not talking about the
kind of mathematical functions of a denumerable domain that may or may not be Turing-
computable. Instead, they were talking about functions as a kind of activity. In this sense of
function, the function of the heart is to pump blood, and the function of a drill is to make
holes. Even with respect to functions, the new mechanist notion of mechanism is independent
of the notion of computation.
Fourth, many new mechanists were hostile to teleology. When they talked about functions,
they typically understood them in a nonteleological way, as something a mechanism does rather
than something a mechanism is for (cf. Craver 2001). But the functions of mechanisms can also
be understood teleologically, as what something is for. There is a large literature, independent
of the new mechanist literature, devoted to explicating teleological functions in nonteleological
440
terms; it may be combined with the new mechanist notion of mechanism to yield functional
mechanisms; that is, mechanisms with teleological functions (Garson 2013). Functional mecha-
nisms in this teleological sense are a subset of all mechanisms (Chapter 8).
6. The mechanistic account of digital computation

The new mechanists argued forcefully that many special sciences explain by uncovering mecha-
nisms. The most obvious candidate sciences were molecular biology, physiology, neuroscience,
and perhaps the study of engineered artifacts.
If all of these sciences explain mechanistically, the question arises of how computer science
and computer engineering explain. A closely related question is how computational explanation
works—not only in computer science and engineering but also in psychology and neuroscience
(Chapter 29). A mechanistic answer to these questions is the starting point of the mechanistic
account of physical computation.
I began defending a version of the mechanistic account in the early 2000s (Piccinini 2003;
although at that time I called it the “functional account”) and spent much of the subsequent
decade developing and extending it (2015). Others followed suit with their own versions,
sometimes tailoring it to specific sciences or varieties of computation (Kaplan 2011; Fresco
2014; Milkowski 2013). The mechanistic account explicates physical computation in terms of
mechanisms and has become a prominent account of physical computation. In explicating the
mechanistic account, I will focus on the version I know best.
A first observation is that computer science and computer engineering have their own
special domain. They study computational systems. By contrast, other sciences—with the excep-
tion of cognitive science and neuroscience—generally do not focus on computational systems.
Thus, if computing systems are mechanisms, there seems to be something distinctive about
them—something that makes them the proper domain of a specific science. Accordingly, an
important task of the mechanistic account is to say what distinguishes computing mechanisms
from other mechanisms.
This task is made difficult by the contested boundaries between computing systems and
non-computing systems. Some theorists take a very restrictive attitude: only systems that
manipulate representations according to rules and such that their manipulations are caused by
representations of the rules count as genuinely computing systems (Fodor 1975; Pylyshyn 1984).
Other theorists take a very liberal attitude: every physical system performs some computation
or another (Chalmers 2011; Scheutz 1999; Shagrir 2006). Yet other theorists fall somewhere
between these two extremes.
To establish reasonable boundaries, I used the following strategy. First, identify a set of para-
digmatic examples of physical computing systems (digital computers, calculators, etc.) and a set
of relatively uncontroversial examples of non-computing systems (digestive systems, hurricanes,
etc.). Second, identify what distinguishes the computing systems from the non-computing
systems. Finally, use the resulting account to decide any boundary cases. My first stab at a
mechanistic account covered only digital computation (including hypothetical hypercomput-
ers), precisely because digital computing systems such as computers and calculators are generally
accepted as computational and because they fall under Turing’s mathematical theory of compu-
tation (Piccinini, 2007b, 2008a).
To construct a mechanistic account of digital computation, the key notion was that of a
digit, which is a concrete counterpart of the abstract notion of letter (or symbol) employed by
computability theorists. As we saw above, a Turing machine is said to manipulate strings of letters
(symbols) according to a set of instructions. A digit is a physical macrostate that corresponds to
441
Gualtiero Piccinini
the mathematical notion of letter (or symbol). It comes in finitely many types that can be reliably
distinguished and manipulated by a physical system. Like a set of letters, it can be concatenated
into strings such that the physical system that manipulates the digits can differentiate between
digits based on their place within a string and treat each digit differently depending on its place
within the string. Given these preliminaries, I argued that physical computing systems are physi-
cal systems whose function is manipulating strings of digits according to a rule defined over the
digits. The rule is just the mapping from input strings of digits to output strings of digits.
This account has several virtues.
First, it makes computation an objective property of physical systems. Whether a physical
system has functions, whether it manipulates strings of digits, and whether it does so in accord-
ance with rules defined over the digits are objective properties of the system, which can be
empirically investigated. Unlimited pancomputationalism is thus avoided.
Second, only a fairly restricted class of physical systems counts as computational. Thus, lim-
ited pancomputationalism is avoided.
Third, the mechanistic account does not assume that computation requires representation. It
shares this virtue with mapping accounts. This is a virtue because many paradigmatic examples
of computing systems do not manipulate representations in any interesting sense. In addition,
the notion of representation is somewhat murky and in need of a clear foundation, so we should
avoid representation, if possible, in explicating computation (Chalmers 2011).
Fourth, the mechanistic account comes with a clear and compelling notion of computational
explanation, which fits the practices of the relevant sciences. That is, computational explanation
is a species of mechanistic explanation. It is the species of mechanistic explanation that appeals
to components that manipulate strings of digits in accordance with rules defined over the digits.
Fifth, the mechanistic account is the first to explain miscomputation. Computer scientists
and engineers work hard to test the reliability of computer circuits, debug computer programs,
and fix computers. In other words, they fight miscomputation. Mapping accounts have a hard
time making any sense of that—after all, there is always a mapping between any physical state
transition and some computation or another. Semantic accounts are not much better off. In
any case, no one attempted to make sense of miscomputation until mechanistic accounts came
along. According to mechanistic accounts, there are several possible sources of miscomputation.
A system may miscompute because it was poorly designed (relative to the goals of the designers),
or poorly built (relative to the design plan), or misused. If a system is programmable, it can also
miscompute because it was programmed incorrectly (relative to the programmer’s intentions).5
Sixth, the mechanistic account can be used to provide an illuminating taxonomy of comput-
ing systems (Piccinini 2008b, 2008c).
Based on this initial mechanistic account, I argued that so-called analog computers and neural
networks that have no digital inputs and outputs are not computing mechanisms properly so
called (Piccinini 2008b, 2008c). This turned out to be overly restrictive.
7. The mechanistic account of generic computation

Analog computers in the strict sense (Pour-El 1974) are physical systems that are used to solve
systems of differential equations. They solve systems of equations by embodying the mathemati-
cal relations between the variables in the equations within the physical variables in the computer
and by physically manipulating the physical variables in a way that corresponds to the relations
defined by the equations. Perhaps the most important operation performed by analog computers
is integration, which allows the computer to output a value corresponding to the integral of a
function over a time interval. Notice that systems of differential equations are defined over real
442
(continuous) variables, which can take any real number as a value. To embody such variables,
analog computers must manipulate physical variables that (are assumed to) take any real values,
at least within reasonable intervals. (Outside of such reasonable intervals, an analog computer
would malfunction; for example, overly high voltage levels within a circuit would fry the cir-
cuit.) Of course, physical variables that (are assumed to) take any real value can only be meas-
ured and manipulated to a finite degree of approximation, which means that analog computers
produce outputs that are approximations of the desired result.
Clearly, there are important differences between digital computers and analog computers.
Whereas digital computers manipulate strings of digits, analog computers manipulate continu-
ous variables. Whereas digital computers can follow precise instructions defined over digits,
analog computers follow systems of differential equations. Whereas digital computers compute
functions defined over denumerable domains, analog computers solve systems of differential
equations. Whereas digital computers can be universal in Turing’s sense, Turing’s notion of uni-
versality does not apply to analog computers. Whereas digital computers produce exact results,
analog computers produce approximate results. These differences may be seen as evidence that
analog “computers” are so different from digital computers that they shouldn’t even be consid-
ered computers at all (Piccinini 2008b).
And yet, those same differences also point at deep similarities. Both digital and analog
computers are used to solve mathematical problems. Both digital and analog computers solve
mathematical problems encoded in the values of physical variables. Both digital and analog
computers solve mathematical problems by following some sort of procedure or rule defined
over the variables they manipulate. Both digital and analog computers used to compete within
the same market—the market for machines that solve systems of equations. (Digital comput-
ers eventually won that competition, in part because they do much more than solving systems
of equations.) These similarities are serious enough that many scientific communities found it
appropriate to use the same word—“computer”—for both classes of systems. Maybe they are
onto something; philosophers may be better off figuring out what that is rather than asserting
that scientists are misguided.
A similar point applies to other unconventional models of computation, such as DNA
computing and quantum computing. In recent decades, computer scientists have investigated
models of computation that depart in various ways from mainstream electronic digital comput-
ing. These models employ chemical reactions, DNA molecules, or quantum systems to perform
computations—or so the people who study them assert. It would be valuable to have an account
of physical computation general enough to cover them all. Notice that neither supporters of
mapping accounts nor supporters of semantic accounts ever addressed this sort of problem.
In light of the above, we need a notion of physical computation broader than that of digi-
tal computation—a notion of computation in a generic sense that covers digital computation,
analog computation, and other unconventional kinds of computation. One option would be to
declare that everything computes (limited pancomputationalism). But that erases all boundaries
between the domain of computer science and every other scientific domain. It would be bet-
ter if there were a way to characterize what digital computing systems have in common with
analog computers and other unconventional computing systems as well as what distinguishes all
of them from other physical systems.
A helpful notion is medium independence, which was used by Justin Garson (2003) to charac-
terize the notion of information introduced by Edgar Adrian (1928) in neurobiology. Garson
argued that information is medium-independent in the sense that it can be carried by a physical
variable (such as spike trains) regardless of the physical origin of the information (which may be
light, sound waves, pressure, etc.).
443
Gualtiero Piccinini
Medium independence in this sense is stronger than multiple realizability. If a property is

medium-independent, it is also multiply realizable, because it’s realizable in different media.
The converse does not hold: a multiply realizable property may or may not be medium-
independent. For example, the property of being a mousetrap is multiply realizable but not
medium-independent: any mousetrap must handle the same medium—mice!
A similar notion of medium independence can be used to characterize all the vehicles of com-
putation, whether digital or not. Specifically, all physical computing systems—digital, analog, or
what have you—can be implemented using physically different variables so long as the variables
possess the right degrees of freedom and the implementing mechanisms can manipulate those
degrees of freedom in the right way. In other words, computational vehicles are macroscopic
variables defined in a medium-independent way. This, then, became my mechanistic account
of computation in the generic sense: a physical computing system is a mechanism whose func-
tion is to manipulate medium-independent variables in accordance with a rule defined over the
variables (Piccinini and Scarantino 2011; Piccinini 2015).
This more general mechanistic account of physical computation inherits the virtues of the
mechanistic account of digital computation. In addition, it covers analog computers and other
unconventional models of computation. In virtue of its generality, the account covers the
notion of computation used by computational neuroscientists without assuming that it must be
digital—or analog, for that matter. In fact, by relying on this generalized mechanistic account,
Sonya Bahar and I have argued on empirical grounds that neural computation is neither digital
nor analog—it is a third, sui generis, notion of computation (Piccinini and Bahar 2013).
Perhaps the biggest point of contention surrounding the mechanistic account of physical com-
putation is whether it adequately subsumes all forms of computational explanation. Carl Craver
and I (2011) argued that it does because functional analysis—the traditional alternative to mecha-
nistic explanation—is just a sketch of a mechanism. Others responded that, at least in cognitive
science or cognitive neuroscience, there are forms of computational explanation that are not
mechanistic (Chirimuuta 2014; Shagrir 2010; Shapiro 2016; Weiskopf 2011). The most com-
mon reason they give is that such forms of computational explanation abstract away from some
aspects of the mechanism that carries out the computation. I believe that when the proper roles of
abstraction within mechanistic explanation are appreciated, constitutive explanation—including
computational explanation—remains mechanistic (Boone and Piccinini 2016, forthcoming). This
debate is likely to continue (see Chapters 16, 17, and 20).
Notes
1 Thanks to Ken Aizawa, Giovanni Camardi, Joe Dewhurst, Anne Jacobson, Jack Mallah, Marcin
Milkowski, Alessio Plebe, Charles Rathkopf, Oron Shagrir, Wilfried Sieg, Guglielmo Tamburrini, my
audience at IACAP 2016, and especially Stuart Glennan and Phyllis Illari for helpful comments on
previous versions. This project was supported in part by International Studies and Programs at the
University of Missouri—St. Louis.
2 This formulation is due to Church (1936). Hilbert and Ackerman (1928) had asked whether some
algorithm could determine whether any given statement written in first-order logic is universally valid—
valid in every structure satisfying the axioms. In more modern terms, Hilbert and Ackerman’s question
is whether any given statement written in first-order logic is logically valid—true under every possible
interpretation. Church’s formulation is equivalent to Hilbert and Ackerman’s because of the complete-
ness of first-order logic (Church 1936, fn. 6).
3 Stephen Kleene (1956) confirmed their assertions by proving that McCulloch–Pitts networks are com-
putationally equivalent to finite state automata, which are computationally equivalent to Turing machines
without tapes or with bounded tapes (strictly speaking, Turing machines have an unbounded tape, which
increases their computational power).
444
4 Independently, some logicians and mathematicians showed that large classes of physical or computational
systems could be simulated more or less exactly by Turing machines (e.g., Gandy 1980; Rubel 1989);
others looked for possible exceptions (e.g., Pour-El and Richards 1989).
5 For more on miscomputation, see Piccinini 2015: 148–150.
References
Adrian, E. D. (1928). The Basis of Sensation: The Action of the Sense Organs. New York: Norton.
Bechtel, W. and R. C. Richardson (1993). Discovering Complexity: Decomposition and Localization as Scientific
Research Strategies. Princeton, NJ: Princeton University Press.
Boone, W. and G. Piccinini (2016). “The Cognitive Neuroscience Revolution,” Synthese (in press).
10.1007/s11229-015-0783-4
Boone, W. and G. Piccinini (forthcoming). “Mechanistic Abstraction,” Philosophy of Science.
Chalmers, D. J. (2011). “A Computational Foundation for the Study of Cognition,” Journal of Cognitive
Science 12(4): 323–57.
Chirimuuta, M. (2014). “Minimal Models and Canonical Neural Computations: The Distinctness of
Computational Explanation in Neuroscience,” Synthese 191(2): 127–154.
Chrisley, R. L. (1994). “Why Everything Doesn’t Realize Every Computation,” Minds and Machines 4:
403–430.
Church, A. (1936). “An Unsolvable Problem in Elementary Number Theory,” The American Journal of
Mathematics 58: 345–363.
Copeland, B. J. (2000). “Narrow Versus Wide Mechanism: Including a Re-Examination of Turing’s
Views on the Mind-Machine Issue.” The Journal of Philosophy XCVI(1): 5–33.
Craver, C. (2001). “Role Functions, Mechanisms, and Hierarchy,” Philosophy of Science 68: 53–74.
Cummins, R. (1983). The Nature of Psychological Explanation. Cambridge, MA: MIT Press.
Dennett, D. C. (1978). Brainstorms. Cambridge, MA: MIT Press.
Etesi, G. and I. Németi (2002). “Non-Turing Computations via Malament-Hogarth Spacetimes,”
International Journal of Theoretical Physics 41: 342–370.
Fodor, J. A. (1968). Psychological Explanation. New York: Random House.
Fodor, J. A. (1975). The Language of Thought. Cambridge, MA: Harvard University Press.
Fredkin, E. (1990). “Digital Mechanics: An Information Process Based on Reversible Universal Cellular
Automata,” Physica D 45: 254–270.
Fresco, N. (2014). Physical Computation and Cognitive Science. New York: Springer.
Gandy, R. (1980). “Church’s Thesis and Principles for Mechanism.” In J. Barwise, H. J. Keisler, and
K. Kuhnen (Eds.), The Kleene Symposium (123–148). Amsterdam: North-Holland.
Garson, J. (2003). “The Introduction of Information into Neurobiology,” Philosophy of Science 70: 926–936.
Garson, J. (2013). “The Functional Sense of Mechanism,” Philosophy of Science 80: 317–333.
Glennan, S. (1996). “Mechanisms and the Nature of Causation,” Erkenntnis 44: 49–71.
Hilbert, D. and W. Ackermann (1928). Grundzüge der theoretischen Logik. Berlin: Springer.
Hogarth, M. L. (1994). “Non-Turing Computers and Non-Turing Computability,” PSA 1: 126–138.
Illari, P. M. and J. Williamson (2012). “What Is a Mechanism? Thinking about Mechanisms across the
Sciences,” European Journal of Philosophy of Science 2: 119–135.
Kaplan, D. M. (2011). “Explanation and Description in Computational Neuroscience,” Synthese 183 (3):
339–373.
Kleene, S. C. (1956). “Representation of Events in Nerve Nets and Finite Automata.” In C. E. Shannon
and J. McCarthy (Eds.), Automata Studies (3–42). Princeton, NJ: Princeton University Press.
Lloyd, S. (2006). Programming the Universe: A Quantum Computer Scientist Takes on the Cosmos. New York:
Knopf.
Lucas, J. R. (1961). “Minds, Machines, and Gödel,” Philosophy 36: 112–137.
Machamer, P. K., Darden, L., and C. Craver (2000). “Thinking about Mechanisms,” Philosophy of Science
67: 1–25.
McCulloch, W. S. and W. H. Pitts (1943). “A Logical Calculus of the Ideas Immanent in Nervous
Activity,” Bulletin of Mathematical Biophysics 7: 115–33.
Milkowski, M. (2013). Explaining the Computational Mind. Cambridge, MA: MIT Press.
Piccinini, G. (2003). Computations and Computers in the Sciences of Mind and Brain.
Doctoral Dissertation, Pittsburgh, PA: University of Pittsburgh. URL = <http://etd.library. pitt.edu/
ETD/available/etd-08132003-155121/>.
445
Gualtiero Piccinini
Piccinini, G. (2007a). “Computationalism, the Church-Turing Thesis, and the Church-Turing Fallacy,”
Synthese 154(1): 97–120.
Piccinini, G. (2007b). “Computing Mechanisms,” Philosophy of Science 74(4): 501–526.
Piccinini, G. (2008a). “Computation without Representation,” Philosophical Studies 137(2): 205–241.
Piccinini, G. (2008b). “Computers,” Pacific Philosophical Quarterly 89(1): 32–73.
Piccinini, G. (2008c). “Some Neural Networks Compute, Others Don’t,” Neural Networks 21(2–3): 311–321.
Piccinini, G. (2015). Physical Computation: A Mechanistic Account. Oxford: Oxford University Press.
Piccinini, G. (forthcoming). “Activities Are Manifestations of Causal Powers.” In M. Adams, Z. Biener,
U. Feest, and J. Sullivan (Eds.), Eppur Si Muove: Doing History and Philosophy of Science with Peter
Machamer. Berlin: Springer.
Piccinini, G. and S. Bahar (2013). “Neural Computation and the Computational Theory of Cognition,”
Cognitive Science 34: 453–488.
Piccinini, G. and C. Craver (2011). “Integrating Psychology and Neuroscience: Functional Analyses as
Mechanism Sketches,” Synthese 183: 283–311.
Piccinini, G. and A. Scarantino (2011). “Information Processing, Computation, and Cognition,” Journal of
Biological Physics 37(1): 1–38.
Pour-El, M. B. (1974). “Abstract Computability and Its Relation to the General Purpose Analog Computer
(Some Connections Between Logic, Differential Equations and Analog Computers),” Transactions of the
American Mathematical Society 199: 1–28.
Pour-El, M. B. and J. I. Richards (1989). Computability in Analysis and Physics. Berlin: Springer-Verlag.
Putnam, H. (1967). “Psychological Predicates.” In W. H. Capitan and D. D. Merrill (Eds.), Art, Mind, and
Religion (37–48). Pittsburgh, PA: University of Pittsburgh Press.
Putnam, H. (1988). Representation and Reality. Cambridge, MA: MIT Press.
Pylyshyn, Z. W. (1984). Computation and Cognition. Cambridge, MA: MIT Press.
Rubel, L. A. (1989). “Digital Simulation of Analog Computation and Church’s Thesis,” Journal of Symbolic
Logic 54(3): 1011–1017.
Scheutz, M. (1999). “When Physical Systems Realize Functions . . . ,” Minds and Machines 9(2): 161–196.
Searle, J. R. (1992). The Rediscovery of the Mind. Cambridge, MA: MIT Press.
Shagrir, O. (2006). “Why We View the Brain as a Computer,” Synthese 153(3): 393–416.
Shagrir, O. (2010). “Brains as Analog-Model Computers,” Studies in History and Philosophy of Science 41:
271–279.
Shapiro, L. A. 2016. “Mechanism or Bust? Explanation in Psychology.” British Journal for the Philosophy of
Science 10.1093/bjps/axv062.
Sieg, W. (1994). “Mechanical Procedures and Mathematical Experience.” In A. George (Ed.), Mathematics
and Mind (71–117). New York: Oxford University Press.
Sieg, W. (1999). “On Computability.” In A. Irvine (Ed.), Philosophy of Mathematics (Handbook of the
Philosophy of Science) (535–630). Amsterdam: North-Holland.
Tamburrini, G. (1988). Reflections on Mechanism, unpublished Ph.D. dissertation, Columbia University.
Turing, A. M. (1936–7). “On Computable Numbers, with an Application to the Entscheidungsproblem,”
Proceeding of the London Mathematical Society 42(1): 230–265.
Turing, A. M. (1939). “Systems of Logic Based on Ordinals,” Proceedings of the London Mathematical Society,
Ser. 2 45: 161–228.
von Neumann, J. (1951). “The General and Logical Theory of Automata.” In L. A. Jeffress (Ed.), Cerebral
Mechanisms in Behavior (1–41). New York: Wiley.
Weiskopf, D. (2011). “Models and Mechanisms in Psychological Explanation,” Synthese 183(3): 313–338.
Wheeler, J. A. (1982). “The Computer and the Universe,” International Journal of Theoretical Physics 21(6–7):
557–572.
Wheeler, J. A. (1990). “Information, Physics, Quantum: The Search for Links.” In W. H. Zurek (Ed.),
Complexity, Entropy, and the Physics of Information (3–28). Redwood City, CA: Addison-Wesley.
Zuse, K. (1970). Calculating Space. Cambridge, MA: MIT Press.
446
34
MECHANISMS AND
ENGINEERING SCIENCE1
Dingmar van Eck
1. Introduction
Use of “mechanism talk” is ubiquitous in engineering science (e.g., Chandrasekaran and
Josephson 2000; Goel 2013). Philosophical discussions of mechanisms also frequently invoke
engineered systems, such as pumps, car engines, mouse traps, toilets, soda vending machines,
and the like in illustrating various aspects of mechanisms and mechanistic explanation (see Levy
2014). Nevertheless, focused philosophical analyses of the structure of mechanistic explanations in
engineering science are scarce (see van Eck 2015). Reference to engineered systems in discussions
of mechanisms and mechanistic explanation is often a loose metaphor, not a conceptualization
that offers sophisticated understanding of what mechanistic explanation looks like in engineer-
ing practice. Moreover, philosophical work that aims to elucidate the connection(s) between
engineering and systems biology—connections that practicing engineers and biologists have been
stressing for more than a decade (e.g. Csete and Doyle 2002)—is also few and far between, in
particular with respect to the use of engineering principles in the construction of mechanistic
explanations in systems biology (see Braillard 2015). In this chapter I address both these issues.
In this chapter I give an outline of the structure of mechanistic explanation in engineering
science, and organize this discussion around two features that extend the mechanistic program
toward explanation when applied to engineering science. First, in section 2, I show that in engi-
neering, two distinct sub-types of role function—“behavior function” and “effect function”—are
employed in the functional individuation of mechanisms, rather than role function simpliciter.
Empirically informed understanding of mechanistic explanation in engineering science requires
sensitivity to this distinction (van Eck 2015). I illustrate this point in terms of reverse engineering
and malfunction explanations in engineering science.
Second, in section 3, I discuss connections between (control) engineering and systems
biology, focusing on the usage of engineering principles in the construction of mechanistic
explanations in systems biology. Systems biology has adopted engineering tools and princi-
ples, in particular from control engineering, to model and explain complex biological systems.
These tools are often in the service of characterizing the organization of mechanisms in
abstract, truncated fashion. I discuss a case of heat shock response in Escherichia coli to illustrate
the role of engineering principles in mechanistic explanation in systems biology (see El-Samad
et al. 2005; Braillard 2015).
447
Dingmar van Eck
This case again shows the relevance of distinguishing behavior from effect function and,
moreover, gives means to elaborate a key issue in a recent and general discussion on the
explanatory power of mechanistic explanations, viz. to flesh out the distinctions between the
explanatory desiderata of “completeness and specificity” (Craver 2007) and “abstraction” (Levy
and Bechtel 2013). Rather than being in competition, as some authors have it, I argue that these
desiderata are suitable for different explanation-seeking contexts.
2. Mechanistic explanation in engineering science

By now, quite a few accounts of mechanistic explanation are on offer in the literature. Although
they come in different flavors, there is broad consensus on a number of key features:
All mechanistic explanations begin with (a) the identification of a phenomenon or

some phenomena to be explained, (b) proceed by decomposition into the entities and
activities relevant to the phenomenon, and (c) give the organization of entities and
activities by which they produce the phenomenon.
(Illari and Williamson 2012: 123; see also Chapter 1)
Mechanistic explanations thus explain how mechanisms, i.e. organized collections of entities and
activities, produce phenomena (Machamer et al. 2000; Glennan 2005; Bechtel and Abrahamsen
2005; Craver 2007).
Role function ascription is considered crucial for (b) decomposition and (c) the elucidation
of mechanisms’ organization (Machamer et al. 2000; Craver 2001; Illari and Williamson 2010;
see Chapter 8). As Machamer et al. (2000) write:
Mechanisms are identified and individuated by the activities and entities that constitute
them, by their start and finish conditions, and by their functional roles. Functions are
the roles played by entities and activities in a mechanism. To see an activity as a func-
tion is to see it as a component in some mechanism, that is, to see it in a context that
is taken to be important, vital, or otherwise significant.
(Machamer et al. 2000: 6)
Mechanistic role functions thus refer to activities that make a contribution to the workings of
mechanisms of which they are a part, and mechanistic organization is key for the ascription of
functions. For instance, in the context of explaining the circulatory system’s activity of “deliver-
ing goods to tissues,” the heart’s “pumping blood through the circulatory system” is ascribed a
function relative to organizational features such as the availability of blood, and the manner in
which veins and arteries are spatially organized (Craver 2001: 64).
This perspective on the general structure of mechanistic explanation, and the importance
of (role) functional individuation of mechanisms, finds widespread support in the literature on
mechanistic explanation in the life sciences. Frequently, in this literature, mechanisms of techni-
cal artifacts, such as clocks, mousetraps, and car engines, are invoked as metaphors to elucidate
features of biological mechanisms (Craver 2001) and features of mechanisms in general (Glennan
2005; Darden 2006; Illari and Williamson 2012). The mechanistic concept of role function, and
its utility in the functional individuation of mechanisms, has likewise been explicated in terms
of mechanisms of technical artifacts such as car engines (Craver 2001).
However, as mentioned in section 1, reference to such technical mechanisms must not be
understood as providing insight into mechanistic explanation in engineering science per se since
448
Mechanisms and engineering science
engineers use multiple notions of function in the functional individuation of technical mechanisms,
rather than the concept of role function simpliciter (van Eck 2015). (And, as we will see, the distinc-
tion between “contextual” and “isolated” descriptions of an entity’s activity (Craver 2001) also does
not capture the distinction in engineering concepts of function.)
Function is a key term in engineering and an ambiguous one (e.g. Chandrasekaran and
Josephson 2000). A variety of function notions are used in mechanism individuation and
explanation in engineering science, and the precise notion of function invoked depends on
the explanatory and design task at hand. Contrary to explanation in other sciences, expla-
nation in engineering science cannot be seen in isolation from design. For instance, failure
analysis—malfunction explanation—is an important type of explanation (Bell et al. 2007)
that has as its ultimate aim the improvement of technical systems; also reverse engineering
explanation is not “merely” mechanistic explanation since its ultimate aim is the redesign
and subsequent improvement of extant technical systems (Otto and Wood 2001). Similarly,
explanation is in the service of conceptual design in knowledge base-assisted designing in
which (mechanistic) explanations of the workings of extant technical systems and their com-
ponents are archived and put to use to develop novel design specifications (Stone and Wood
2000). Below I zoom in on two such contexts and the relevance of different notions of engi-
neering function for the functional individuation of technical mechanisms in these contexts,
viz. reverse engineering explanation and malfunction explanation. As we will see, specific
notions of function are optimally “engineered” for specific explanatory settings (I focus here
on explanatory contexts, not the design contexts to which they are related).
Function has no uniform meaning in engineering: different approaches advance different
conceptualizations (Erden et al. 2008), and some researchers use the term with more than one
meaning simultaneously (Chandrasekaran and Josephson 2000). This ambiguity led to philo-
sophical analysis of the precise meanings of function involved. Vermaas (2009) regimented the
spectrum of available function meanings into three “archetypical” engineering conceptualiza-
tions of function: behavior function—function as the desired behavior of a technical artifact; effect
function—function as the desired effect of behavior of a technical artifact; and purpose function—
function as the purpose for which a technical artifact is designed.2 In the ensuing discussion on
reverse engineering explanation and malfunction explanation, the notions of behavior function
and effect function are most relevant.
Behavior functions are typically modeled as conversions of flows of materials, energy, and
signals, where input flows and output flows in the conversion (are assumed to) match in terms
of physical conservation laws (Stone and Wood 2000; Otto and Wood 2001). For instance, the
function “loosen/tighten screws” of an electric screwdriver is then represented as a conver-
sion of input flows of “screws” and “electricity” into corresponding output flows of “screws,”
“torque,” “heat,” and “noise” (see Stone and Wood 2000: 364). Since these descriptions of
functions are specified such that input and output flows match in terms of physical conserva-
tion laws—here, the conservation of energy through the conversion of electrical energy into
rotational, thermal, and acoustic energy—they are taken to refer to specific physical behaviors
of technical artifacts (Vermaas 2009).
Effect function descriptions refer to only the technologically relevant effects of the physical
behaviors of technical artifacts: the requirements are dropped that descriptions of these effects
meet conservation laws and that matching input and output flows are specified (Vermaas 2009).
The function of an electric screwdriver is then described simply as, say, “loosen/tighten screws,”
leaving the physical antecedents of this effect unmentioned. Behavior function descriptions thus
refer to the “complete” behaviors involved, including features like thermal and acoustic energy
flows, whereas effect functions refer to subsets of these behaviors, i.e. desired effects.3
449
Dingmar van Eck
Engineering descriptions and explanations of the workings of extant technical artifacts and
artifact designs are often constructed by functionally decomposing functions into a number of
sub-functions. The relationships between functions and sets of their sub-functions are often
graphically represented in functional decomposition models. Like the concept of function,
such models come in a variety of “archetypical” flavors (van Eck 2011). In the context of
reverse engineering explanation and malfunction explanation, the relevant ones are behavior
functional decomposition—a model of an organized set of behavior functions, and effect functional
decomposition—a model of an organized set of effect functions.
In reverse engineering explanations, elaborate behavior functions and functional decom-
positions are used; in malfunction explanations, less detailed effect functions and functional
decompositions are employed.
In engineering science, reverse engineering and engineering design go hand in glove
(e.g. Otto and Wood 2001; Stone and Wood 2000). Consider Otto and Wood’s (2001) reverse
engineering and redesign method, in which a reverse engineering phase in which reverse
engineering explanations are developed for existing artifacts precedes and drives a subsequent
redesign phase of those artifacts. The goal of the reverse engineering phase is to explain how
existing artifacts produce their overall (behavior) functions in terms of underlying mechanisms,
i.e. organized components and sub-functions (behaviors) by which overall (behavior) func-
tions are produced. These explanations are subsequently used in the redesign phase to identify
components that function sub-optimally and to either improve them or replace them with
better-functioning ones.
In the reverse engineering phase, an artifact is first broken down component-by-component,
and hypotheses are formulated concerning the functions of those components. In this method,
functions are behavior functions and are represented by conversions of flows of materials,
energy, and signals. Since the aim of the reverse engineering phase is to understand in detail the
manner in which an extant technical system operates, elaborate behavior function descriptions
are used. Descriptions of input–output conversions give more relevant details than descriptions
of effects. After this analysis, a different reverse engineering analysis commences in which com-
ponents are removed, one at a time, and the effects are assessed of removing single components
on the overall functioning of the artifact. Such single-component removals are used to detail the
behavior functions of the (removed) components further. The idea behind this latter analysis is
to compare the results from the first and second reverse engineering analysis to gain a potentially
more nuanced understanding of the functions of the components of the (reverse engineered)
artifact. Using these two reverse engineering analyses, a behavior functional decomposition
of the artifact is then constructed in which the behavior functions of the components are
specified and interconnected by their input and output flows of materials, energy, and signals
Electricity, human force, relative Torque, heat, noise, human force,

rotation, weight weight
Hand, bit, screw Hand, bit, screw

Loosen/tighten
screws
Direction, on/off, manual use Looseness (or tightness)
Figure 34.1 Overall behavior function of an electric power screwdriver. Thin arrows represent energy
flows; thick arrows represent material flows, dashed arrows represent signal flows (adapted
from Stone and Wood 2000: 363, figure 2)
450
bit Human Human
force force Human force
hand import couple hand secure hand separate hand
hand solid solid solid
bit
bit
hand hand
hand
secure Human dissipate allow rot.
rotation force torque DOF
Heat, noise
Human hand
import hand
force regulate Human regulate
human
rotation force translation Human force
force
Human
Direction on/off force heat
Elect.
convert
store supply actuate regulate
elect. to
electricity electricity electricity Elect. electricity Elect. torque
Elect.
torque bit torque

H.f. H.f. Human force
change transmit torque rotate torque dissipate Heat, noise
torque torque torque solid bit torque bit
bit
Figure 34.2 Behavior functional decomposition of an electric power screwdriver. Thin arrows represent energy flows; thick arrows represent material flows, dashed
arrows represent signal flows (adapted from Stone and Wood 2000: 364, figure 4)
Dingmar van Eck
(Otto and Wood 2001). Such models represent parts of the mechanisms by which technical
systems operate, to wit: causally connected behaviors of components. Examples of an over-
all behavior function and behavior functional decomposition of a reverse engineered electric
screwdriver are given in Figures 34.1 and 34.2, respectively.
In the model in Figure 34.2, temporally organized and interconnected behaviors are
described. Components of artifacts are described in Otto and Wood’s method in tables, what in
engineering are called “bills of materials,” together with a model, called an “exploded view,”
of the components composing the artifacts. Taken together, these component and behavior
functional decomposition models provide functional individuations and representations of
mechanisms of artifacts.
Such (behavior functional decomposition) models are subsequently used to identify sub-
optimally functioning components and so drive succeeding redesign phases (Otto and Wood
2001). The focus here is on the reverse engineering explanation part of the methodology.
In malfunction explanation, this detail in mechanistic models is, however, not required:
engineers take it that less detailed effect functions and functional decompositions there do a
better explanatory job.
In malfunction analysis, explanation-seeking questions of the following format arise:
Why does artifact x not serve the expected function to ϕ?
Such questions are contrastive: why malfunction, rather than normal function? In the engineer-
ing literature, malfunction explanations that answer contrastive questions list different and
fewer mechanistic features than reverse engineering explanations which answer questions about
normal behavior or function.4 Such explanations are constructed using effect functions and
functional decompositions.
Malfunction explanations in engineering pick out only a few features of mechanisms, i.e.
those causal factors—failing components or sub-mechanisms—that are taken to make a differ-
ence to the occurrence of a specific malfunction, as well as some coarse-grained details of the
containing mechanism to understand where the fault is located. Yet most information about
structural and behavioral specifics of malfunctioning components/sub-mechanisms, and their
containing mechanisms, is left out (Hawkins and Woollons 1998; Bell et al. 2007).5,6
Consider, by way of example, the Functional Interpretation Language (FIL) methodology
for malfunction analysis and explanation (Bell et al. 2007). In FIL, functions are effect functions
and represented in terms of their triggers and effects. Triggers describe input states that actuate
physical behaviors which result in certain (expected) effects. So triggers are the input conditions
for effects, i.e. functions, to be achieved but they are not the immediate physical antecedents
of these effects. These physical antecedents, i.e. input flows, are not referred to in trigger-effect
descriptions and neither do these descriptions meet conservation laws. For instance, consider
the function description “depress_brake_pedal”—“red_stop_lamps_lit” of a car’s stop light (Bell
et al. 2007: 400), in which electromagnetic radiation (“light”) is “created” rather than being
converted from other energies. This description, rather, is a summary of some salient features of
(manipulating) such artifacts; depressing the brake pedal will, if the system functions properly,
result in the lighting of the stop lamps, i.e. the effect function.
According to Bell et al. (2007), such trigger and effect representations serve two explanatory
ends in malfunction analyses: first, they highlight relevant behavioral features of a given artifact,
i.e. effects, and, simultaneously, provide the means to ignore less relevant or irrelevant behavioral
features, i.e. physical behaviors underlying these effects; second, they support assessing which
components are malfunctioning (Bell et al. 2007: 400–1).
452
For instance, the trigger-effect representation “depress_brake_pedal”—“red_stop_lamps_lit”

highlights the input condition of a pedal being depressed, and the resulting desired effect
of lighted lamps, yet ignores the structural and behavioral specifics of the brake pedal and
stop lamps, such as the pedal lever and electrical circuit mechanisms, as well as the energy
conversions—e.g., mechanical energy conversions into electricity—that are needed to achieve
this effect. Such representations only highlight those features that are considered explanatorily
relevant to assess malfunctioning systems, and omit reference to physical behaviors/energy
conversions by which the desired effects are achieved.
Second, such trigger-effect descriptions support comparing normally functioning technical
systems with malfunctioning ones (Bell et al. 2007). Trigger-effect descriptions support assess-
ing whether the expected effects in fact obtain, and, if not, which and how components are
malfunctioning (Bell et al. 2007). A normally functioning artifact, say the car’s stop lights, has
both a trigger and an effect occurring; the brake pedal is depressed and the stop lights are lit.
Trigger-effect descriptions support analysis of two varieties of malfunction. First, a trigger may
occur, yet fail to result in the intended effect. Say, the brake pedal is depressed, yet the stoplights
are not on. Second, a trigger may not be occurring, yet the effect is nevertheless present. Say,
the brake pedal is not depressed, yet the stoplights are on (see Bell et al. 2007). Such analysis of
the actual states of triggers and effects allows one to focus on the most likely causes of failure
(Bell et al. 2007). Say, if the pedal is depressed and the lights fail to ignite, the first likely causes
to investigate may be whether the electrical circuits in the lights are broken or the “on/off”
connection between the brake and electrical circuitry (connected to the lamp) is damaged. On
the other hand, if the pedal is not depressed and the lights are lit, a first likely cause to investi-
gate may be whether the “on/off “connection between the brake and the electrical circuitry is
damaged. To support more detailed malfunction analyses, functions are often decomposed into
sub-functions in FIL. An example of a functional decomposition of a two-ring cooking hotplate
is given in Figure 34.3.
The usage of effect functions and functional decompositions in FIL is the optimal choice given
that function descriptions are used to black-box or suppress reference to unwanted behavioral
function purpose
cook on
cook food
hotplate
OR
function function purpose
cook on right cook on left cook on ring
triggers triggers
switch on heat right ring switch on heat left ring
trigger effect trigger effect
Figure 34.3 Effect functional decomposition of a two-ring cooking hotplate (adapted from
Bell et al. 2007)
453
Dingmar van Eck
and structural details. Effect function descriptions only highlight the relevant difference-making
properties with respect to malfunctioning artifacts, whereas more elaborate behavior function
descriptions include irrelevant details such as, say, the thermal energy generated when lamps are lit.
The upshot of these two cases is that explanations in engineering are furnished relative to
explanatory objectives and, importantly, the level of detail included in these explanations hinges
on specific concepts of technical function. Engineering scientists simplify or increase the details
of explanations—functional decompositions—depending on the explanatory purpose at hand,
and these adjustments are made using specific concepts of technical function (compare e.g.
Figures 34.2 and 34.3). In reverse engineering explanation, elaborate or “complete” descriptions
of mechanisms are provided, in terms of behavior functions and functional decompositions, to
answer the question of how a technical system exhibits a given overall behavior. In malfunction
explanation, less elaborate “sketches” of mechanisms are provided in terms of effect functions
and functional decompositions, referring only to some mechanistic features, namely those
difference-making factors that mark the contrast between normal functioning and malfunction-
ing technical systems. So, depending upon explanatory context, mechanisms are individuated
in different ways using different conceptualizations of function in engineering science. Neither
function conceptualization in itself accommodates both ways in which mechanisms are function-
ally individuated in engineering science. Behavior and effect function ascriptions are invoked to
individuate mechanisms in different ways depending on the task at hand.
However, this distinction in functional individuation, and its reliance on different function
concepts, remains opaque when seen from a perspective that conceives of mechanism individu-
ation and mechanistic explanation in terms of mechanistic role function ascription simpliciter.
The concept of mechanistic role function, an activity that makes a contribution to the workings
of a mechanism of which it is a part, admits of two interpretations in the context of engineering
science: behavior function on the one hand and effect function on the other. Note that behavior
and effect descriptions of function describe, in different ways, the contributions of components
to mechanisms of which they are a part. The distinction between behavior and effect function
thus is not to be conflated with the distinction between a mechanism description and a descrip-
tion of a mechanism’s overall activity. Neither is the behavior-effect function distinction to be
conflated with the distinction between “isolated” and “contextual” descriptions of an entity’s
activity (Craver 2001): isolated descriptions describe activities without taking into account the
mechanisms in which they are situated; contextual descriptions describe activities in terms of the
mechanistic contexts in which they are situated and to which they contribute. Both behavior
and effect functions are of the contextual variety, describing contributions of components to the
mechanisms of which they are a part. So to arrive at an empirically informed understanding of
explanatory practices in engineering, and at consistency of the general structure of mechanistic
explanation with these practices, regimenting the concept of role function into domain-specific
engineering concepts of behavior and effect function, i.e. sub-types of role function, is needed.
I now turn to another facet of the relationship between mechanistic explanation and engi-
neering that has received little sustained analysis: the usage of engineering principles in the
construction of mechanistic explanations in systems biology. Here we see again the relevance
of knowing the ways in which mechanisms are functionally individuated in engineering: the
manner in which biological mechanisms are individuated in engineering terms also hinges on
specific engineering conceptualizations of function. To better understand the specifics of mecha-
nism individuation along engineering lines in systems biology, sensitivity to the varieties of
engineering function and functional individuation is thus called for. In the case I discuss below,
mechanism individuation hinges on the use of effect function descriptions and ascriptions.
454
3. Explanation and systems thinking: where engineering and

systems biology meet
Although philosophy, it seems, is only recently picking up on the fruitful cross-talk between
engineering and systems biology (see Braillard 2015), with Wimsatt’s (e.g. 2006) work being a
notable exception, engineers and systems biologists alike have been stressing the conceptual ties
for more than a decade (Hartwell et al. 1999; Lazebnik 2002). With biological data about com-
plex biological systems exploding during the last 20 years or so because of (functional) genomics
projects and the like, opportunities to understand complex biological systems in far greater detail
became available. Yet cashing out that promise also signaled the need for new tools that enabled
massive data analysis and integration to build explanatory models of these complex systems with
a scale and complexity hitherto unknown (see Chapter 27). Here is where, amongst others,
engineering tools came in.
As Hartwell et al. (1999), for instance, commented with respect to engineering representa-
tional schemes:
In our opinion, the most effective language to describe functional modules and their
interactions [in systems biology] will be derived from the synthetic sciences, such as
computer science or engineering, in which function appears naturally.
(Hartwell et al. 1999: C49)
Or as Csete and Doyle (2002) commented with respect to systemic organization:
Advanced technologies [like cars and airplanes] and biology have extremely different
physical implementations, but they are far more alike in systems-level organization
than is widely appreciated
(Csete and Doyle 2002: 1664)
Functional engineering parlance and systemic organization come together in the connection
between control engineering and systems biology, in which (effect functional) decomposition
and control principles governing (the construction of) engineering systems are used to charac-
terize complex biological systems:
Some insightful recent papers advocate a similar modular decomposition of biological

systems according to the well defined functional parts used in engineering and, specifi-
cally, engineering control theory.
(Tomlin and Axelrod 2005: 4219)
A case in point is research by El-Samad et al. on the mechanism(s) to counter heat shock in
Escherichia coli (El-Samad et al. 2005; see Tomlin and Axelrod 2005; Braillard 2015). Heat
shock response is a widely conserved response of cells to cope with environmental stress
brought about by unusual increases in temperature, involving the induced expression of heat
shock proteins. Such temperature increases can damage proteins by breaking down their
tertiary structures. Heat shock proteins come in two varieties and mitigate this effect in two
different ways: molecular chaperones do so by refolding denatured proteins and proteases
by degrading denatured proteins. If the response is sufficiently swift and massive, cell death
can be prevented by protein repair and/or removal of damaged proteins. The response
455
Dingmar van Eck
needs to be tightly controlled in the sense that it is only activated in the case of heat shock,
since the response is highly energy-consuming and would make too high energy demands
if heat shock proteins would be produced all the time. Cells thus must maintain a delicate
balance between the protective effect of heat shock protein production and the metabolic
cost of overproducing these proteins. In E. coli, the RNA polymerase cofactor ø32 promotes
the transcription of heat shock proteins. After heat shock stress—temperature increase—
ø32 activity increases, resulting in the transcription of specific heat shock gene promoters
which initiate the transcription of genes that in turn encode specific heat shock proteins—
chaperones and proteases. This heat shock protein expression, when appropriate, prevents
cell death. This mechanism uses both feed-forward and feedback loops that process informa-
tion about temperature and the folding state of proteins in the cell. ø32 activity is crucial in
all this and depends on a feed-forward mechanism that senses temperature and controls ø32
transcription, and feedback regulatory mechanisms that register the folding levels of proteins
(levels of denatured cellular protein) and degrade ø32. These regulatory feedback mechanisms
are crucial to ensure that ø32 synthesis, activity, and stability are brought back to normal
levels after a sufficient number of heat shock proteins have been produced and the threat to
cell death is averted.
The above qualitative information on the heat shock response system is well known.
El-Samad and his group (2005) went further and constructed a quantitative, mathemati-
cal model of the heat shock response to “use this description to pose questions about
the regulatory architecture of the system” (El-Samad et al. 2005: 2737), i.e. the dynami-
cal, mechanistic organization that sustains the heat shock response. They came up with
an elaborate mathematical model consisting of 31 equations and seven parameters. Now,
to make the model computationally tractable and pose and answer questions about the
dynamical, mechanistic organization of the system, the original model had to be trimmed
down. This model reduction was effected by various simplifications and, importantly, the
salient modularity of the system which made it possible to decompose the system into func-
tional modules, described in terms of effect functions, along control engineering lines. The
resulting model was taken to be a “simplified yet reasonably accurate version of the original
model” (El-Samad et al. 2005: 2737).
As Braillard (2015) stressed, control engineering principles played an important heuristic
role in this model reduction, and thus in the discovery of the mechanism’s core organizational
features that sub-serve its overall regulatory behavior. The close analogy between engineered
systems and biological ones with respect to functional modular organizations sub-serving
regulatory processes makes this possible. As El-Samad et al. (2005) explain:
Control and dynamical systems theory is a discipline that uses modular decompositions
extensively to make modeling and model reduction more tractable. Because biological
networks are themselves complex regulation systems, it is reasonable to expect that
seeking similarities with the functional modules traditionally identified in engineering
schemes can be particularly useful.
(El-Samad et al. 2005: 2737)
In control engineering, decomposition into functional modules (modules defined in terms of

their effect-role functions) often begins with identification of the process to be regulated called
the “plant” (see Lind 1994), for instance altitude regulation of an airplane or temperature regu-
lation of a thermostat. Modules of the system that contribute to the regulation are described in
456
terms of their contributing functions, the most common of which are “sensors,” “detectors,”
“controllers,” “actuators,” and “feed-forward” and “feedback” signals. For instance, in a simple
heating system, the plant is the temperature regulation process, which is achieved, inter alia, by
a sensor module which measures ambient temperature, calculates the deviation from the desired
temperature, and feeds this information into the thermostat (controller). The thermostat then
outputs signals that are sent to an actuator (heat fuel valve) that generates an actuation signal
(e.g. fuel to furnace) that corrects deviation from the desired temperature. The sensor module
again measures the ambient temperature and, if needed, feeds back information on temperature
deviations to the controller, and so on.
El-Samad et al. (2005) applied this control engineering perspective to the E. coli heat shock
response system. In this application, the protein folding task (the refolding of denatured pro-
teins) is taken to be the process to be regulated (plant), the feed-forward signal (sent by a sensor)
is the temperature-dependent translational efficiency of ø32 synthesis, the controller is the level
of ø32 activity, chaperones function as actuators of the plant (the actuated plant input is the
number of molecular chaperones), and sensors measure plant output (the amount of denatured
protein), which in turn is fed back to the controller.
This decomposition allowed El-Samad et al. (2005) to construct a simplified model consist-
ing of just six equations and 11 parameters in which each equation describes the behavior of a
module. They remark:
This model provides useful insight into the heat shock system design architecture.
It also suggests a mathematical and conceptual modular decomposition that defines
the functional blocks or submodules of the heat shock system. This decomposition is
drawn by analogy to manmade control systems and is found to constitute a canonical
blueprint representation for the heat shock network.
(El-Samad et al. 2005: 2736)
What we here thus see is that analogical reasoning with respect to regulation processes, and the
functional architecture sub-serving these processes in engineered and biological systems, led to
a functional modular decomposition of a biological system that laid bare core organizational
features of the system by which it produces regulatory behavior. Engineering tools here serve
as a discovery heuristic for a mechanism’s core organizational features that sub-serve its overall
regulatory behavior (see Braillard 2015; Chapter 19, this volume). This usefulness of engineer-
ing concepts, i.e. modular decompositions in terms of effect functions, is not specific to the
E. coli case, and generalizes to a variety of cases (see Tomlin and Axelrod 2005) and suggests a
general discovery heuristic:
If the heat shock mechanism can be described and understood in terms of engineer-
ing control principles, it will surely be informative to apply these principles to a broad
array of cellular regulatory mechanisms and thereby reveal the control architecture
under which they operate.
(Tomlin and Axelrod 2005: 4220)7
In concluding this chapter, I suggest that this case gives relevant insights into a general discussion
on explanatory power in the recent mechanisms literature by providing an empirical illustration
of the complementarity of two allegedly competing perspectives on the explanatory power of
mechanistic explanations.
457
Dingmar van Eck
I have argued elsewhere that differences between two main (allegedly) competing
perspectives on the explanatory power of mechanistic explanations, “completeness and
specificity” (Craver 2007) and “abstraction” (Levy and Bechtel 2013), essentially boil down
to differences in the notions of difference making endorsed in these accounts and that
they are in fact not in competition (van Eck 2015). They are rather suitable for different
explanation-seeking contexts. Whereas abstraction dictates that mechanistic explanations
should only list the “primary factors” responsible for the occurrence of system function,
“completeness and specificity” prescribes that in addition to primary ones, “higher-order
factors” should also be described, which concern factors influencing the precise manner in
which a system function occurs or those sub-serving the primary factors. The E. coli case
gives an empirical illustration of this view.
The notion of “robustness” looms large in the E. coli case, as well as in systems biology and
engineering in general. Robust systems—ones resilient to perturbations to parts of the mecha-
nism or the environment in which it functions—require complex sub-systems dedicated to
counteracting perturbations (Kitano 2004). This holds both for complex biological systems and
(most) engineered systems. Think, for instance, of all the sub-systems of an airplane dedicated
to counteracting changes to make it fly in the appropriate manner, or the sub-systems in E. coli
that play a role in counteracting the effects of heat shock on protein deformation—chaperons
and proteases. As El-Samad et al. (2005) elaborate:
The modular decomposition of the hsr [heat shock response] shows a level of com-
plexity not justified by the basic functionality demanded from an operational heat
shock system. A simple and operational heat-shock system would consist solely of
a temperature sensor . . . and a transcriptional/translational apparatus that responds
appropriately to temperature changes.
(El-Samad 2015: 2738)
Why, then, is additional complexity present? Computational modeling indicated that “complex-
ity is indeed necessary to achieve robustness, noise rejection, speed of response, and economical
use of cellular resources, much like engineering systems” (El-Samad 2015: 2738).
Complexity and robustness provide an interesting slant on “abstraction” and “complete-
ness and specificity.” Depending on the questions one asks with respect to complex, robust
systems, either “completeness and specificity” or “abstraction” are better suited. For instance,
one may address the question: “How does the mechanism execute its regulatory function?”,
or the joint questions: “How does the mechanism execute its regulatory function?” and “Why
does it (execute its) function in a robust manner?” If one is interested in the key organizational
details that enable complex systems to function, abstract description suffices. If, on the other
hand, one is also interested in the mechanistic details that enable a mechanism to function in a
robust fashion, more specific and elaborate descriptions are called for. In the latter case, one is
not only interested in the “primary factors” responsible for the occurrence of system function,
but also in the “higher-order factors” influencing the precise manner in which it occurs or those
sub-serving the primary factors (see Weisberg 2007).
To round up, the functional individuation of mechanisms—in terms of behavior and effect
function ascription and decomposition—proceeds differently in engineering science than the
manner in which it is taken to work in the life sciences. Understanding these specifics is required
to understand the structure of mechanistic explanation in engineering science and, moreover,
adds to our understanding of the ways in which tools and insights from engineering are used in
458
mechanism individuation and explanation in systems biology. Finally, cases from engineering
and systems biology give general insights into the explanatory power of mechanistic models in
specific explanation-seeking contexts.
Notes
1 It is a pleasure thanking Phyllis Illari and Stuart Glennan for the opportunity to write this chapter and
for their helpful comments on a previous version.
2 The term “archetypical” here refers to “most common”; the three conceptualizations of function are not
meant to be exhaustive. For instance, some engineers use “function” to refer to intentional behaviors of
agents (see van Eck 2010). In reverse engineering analyses, “function” refers to actual or expected behav-
ior, without the normative connotation “desired.”
3 Behavior and effect functions thus have a partly common semantic structure: certain aspects or features of
behaviors that they both refer to.They are dissimilar in the sense that behavior function descriptions refer to
additional behavioral aspects, not referred to in effect function descriptions, so as to make these descriptions
accord with physical conservation laws. The relation between behavior and effect function is asymmetrical
in the sense that effects, being subsets of behaviors, are straightforwardly derivable from behaviors, but not
vice versa. From a given effect one cannot automatically derive the behavior of which the effect is a part.
Cars that run on gasoline operate by means of different energy conversions than cars that run on electricity,
yet both display the same effects; say, delivering acceleration. The semantic structure that they partly have
in common creates the possibility and need to be pluralist about mechanistic role functions, i.e. different
ways to conceive of the role functions of mechanisms, in the context of engineering science. I defend this
pluralism about mechanistic role functions later on in this section. To be sure, I am thus not advocating a
pluralist view about functions of mechanisms with a completely dissimilar semantic structure.
4 Reverse engineering explanation, like mechanistic explanation in general, is not contrastive, whereas
malfunction explanation is. The role of contrasts, essential to counterfactual accounts of explanation,
seems not vital to most accounts of mechanistic explanation. Mechanistic explanations are often
taken to track mechanisms that exhibit productive continuity, and are typically not construed in
counterfactual fashion. Counterfactual reasoning, rather, is often invoked in analyses of mechanism
discovery and in explanatory relevance assessments where interventions on putative components are
stressed (Craver 2007). Malfunction explanations thus make for an interesting extension of mecha-
nistic conceptions of explanation, since they are both mechanistic and contrastive.
5 That is, structural and behavioral characteristics are considered irrelevant in a first-round functional
analysis of malfunction. After this analysis, more detailed behavioral models of components and their
behaviors are used for identifying specific explanatorily relevant structural and behavioral characteristics
of malfunctioning components/sub-mechanisms (Bell et al. 2007). However, immediately specifying
these details in functional models is taken to result in listing a lot of irrelevant details.
6 Malfunction explanations in engineering thus exemplify Garson’s (2013) “functional sense of mecha-
nism” (see Chapter 8); a malfunction is seen as a breakdown of a mechanism, not as the result of a specific
mechanism for malfunction.
7 The analysis I gave in this section illustrates what Glennan and Illari call “methodological mechanism”
(see Chapter 1): seen from an epistemological and methodological perspective, biology and engineering
have much in common and tools and insights from the latter can help address explanatory issues in the
former (and vice versa). I thus do not address the metaphysical nature of mechanisms and how differences
play out in this regard between biology and engineering. So, for instance, whether biological mecha-
nisms are modular like (most) technical ones or whether we impose modularity on them to understand
how they work is, although a very intriguing question, not one I am concerned with here. Neither do I
focus here on possible differences that may emerge between the life sciences and engineering when we
consider normative conceptions of function. It might transpire, however, that these differences are not so
great as some might suspect, at least in the context of explanation: of course, engineers design and build
technical systems-mechanisms with desired (“proper”) functions, yet a “functional sense of mechanism”
(Garson 2013; see Chapter 8) suffices to account for a token mechanism with a function that it fails to
perform, both in engineering and biological contexts. Furthermore, in engineering contexts of explana-
tion, engineering sub-types of role function do the explanatory work, not an etiological conception of
function (van Eck and Weber 2014).
459
Dingmar van Eck
References
Bechtel, W., and A. Abrahamsen. (2005). “Explanation: A Mechanist Alternative”, Studies in History and
Bell, J., Snooke, N., and C. Price. (2007). “A Language for Functional Interpretation of Model Based
Simulation”, Advanced Engineering Informatics 21:398–409.
Braillard, P.A. (2015). “Prospects and Limits of Explaining Biological Systems in Engineering Terms”. In:
Explanation in Biology (Eds. P.A. Braillard & C. Malaterre), Dordrecht: Springer, 319–44.
Chandrasekaran, B., and J.R. Josephson. (2000). “Function in Device Representation”, Engineering with
Computers 16:162–77.
Craver, C.F. (2001). “Role Functions, Mechanisms, and Hierarchy”, Philosophy of Science 68:53–74.
Craver, C.F. (2007). Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience. New York:
Csete, M.E., and J.C. Doyle. (2002). “Reverse Engineering of Biological Complexity”, Science 295:
1664–9.
Darden, L. (2006). Reasoning in Biological Discoveries. Cambridge: Cambridge University Press.
El-Samad, H., Kurata, H., Doyle, J.C., Gross, C.A., and M. Khammash. (2005). “Surviving Heat Shock:
Control Strategies for Robustness and Performance”, PNAS 102(8):736–41.
Erden, M.S., Komoto, H., Van Beek, T.J., D’Amelio, V., Echavarria, E., and T. Tomiyama. (2008).
“A Review of Function Modeling: Approaches and Applications”, Artificial Intelligence for Engineering
Design, Analysis, and Manufacturing 22:147–69.
Garson, J. (2013). “The Functional Sense of Mechanism”, Philosophy of Science 80:317–33.
Glennan, S. (2005). “Modeling Mechanisms”, Studies in the History and Philosophy of the Biological and
Biomedical Sciences 36(2):375–88.
Goel, A.K. (2013). “A 30-Year Case Study and 15 Principles: Implications of an Artificial Intelligence
Methodology for Functional Modeling”, AIEDAM 27(3):203–15.
Hartwell, L.H., Hopfield, J.J., Leibner, S., and A.W. Murray. (1999). “From Molecular to Modular Cell
Biology”, Nature 402:C47–C52.
Hawkins, P.G., and D.J. Woollons. (1998). “Failure Modes and Effects Analysis of Complex Engineering
Systems using Functional Models”, Artificial Intelligence in Engineering 12(4):375–97.
Illari, P., and J. Williamson. (2010). “Function and Organization: Comparing the Mechanisms of Protein
Synthesis and Natural Selection”, Studies in History and Philosophy of Biological and Biomedical Sciences
41:279–91.
Illari, P., and J. Williamson. (2012). “What Is a Mechanism? Thinking about Mechanisms Across the
Sciences”, European Journal for Philosophy of Science 2:119–35.
Kitano, H. (2004). “Biological Robustness”, Nature 5:826–37.
Lazebnik, Y. (2002). “Can a Biologist Fix a Radio?—Or, What I Learned While Studying Apoptosis”,
Cancer Cell 2:179–82.
Levy, A. (2014). “Machine-Likeness and Explanation by Decomposition”, Philosopher’s Imprint 6:1–15.
Levy, A. and W. Bechtel. (2013). “Abstraction and the Organization of Mechanisms,” Philosophy of
Science 80:241–61.
Lind, M. (1994). “Modeling Goals and Functions of Complex Industrial Plants”, Applied Artificial Intelligence
8:259–83.
Machamer, P.K., Darden, L., and C.F. Craver. (2000). “Thinking about Mechanisms”, Philosophy of Science
57:1–25.
Otto, K.N., and K.L. Wood. (2001). Product Design: Techniques in Reverse Engineering and New Product
Development. Upper Saddle River, NJ: Prentice Hall.
Stone, R.B., and K.L. Wood. (2000). “Development of a Functional Basis for Design”, Journal of Mechanical
Design 122:359–70.
Tomlin, C.J., and Axelrod, J.D. (2005). “Understanding Biology by Reverse Engineering the Control”,
PNAS 102(12):4219–20.
Van Eck, D. (2010). “On the Conversion of Functional Models: Bridging Differences Between Functional
Taxonomies in the Modeling of User Actions”, Research in Engineering Design 21(2):99–111.
Van Eck, D. (2011). “Supporting Design Knowledge Exchange by Converting Models of Functional
Decomposition”, Journal of Engineering Design 22(11–12):839–58.
460
Van Eck, D. (2015). “Mechanistic Explanation in Engineering Science”, European Journal for Philosophy of
Science 5(3):349–75.
Van Eck, D., and E. Weber. (2014). “Function Ascription and Explanation: Elaborating an Explanatory
Utility Desideratum for Ascriptions of Technical Functions”, Erkenntnis 79:1367–89.
Vermaas, P.E. (2009). “The Flexible Meaning of Function in Engineering”, Proceedings of the 17th
International Conference on Engineering Design (ICED 09) 2:113–24.
Weisberg, M. (2007). “Three Kinds of Idealization”, The Journal of Philosophy 104(12):639–59.
Wimsatt, W. (2006). Re-Engineering Philosophy for Limited Beings: Piecewise Approximations to Reality.
Cambridge, MA: Harvard University Press.
461
INDEX
abiotic factors 349 ancient sources of mechanistic thought 3, 13–25;

abortions 429 Aristotle see Aristotle; early atomism 14–16, 17;
above-ground competition 349–52, 354 Epicureans 16–17; The Mechanics 20–1; On the
Abrahamsen, A. 35–6, 37, 298–9, 363 Cosmos 21–2
absences 352 Andersen, H. 100, 101, 102
abstract explanatory models 219–21 Anderson, C.H. 277
abstract forms of interaction 100–1 anencephaly 113
abstraction 278, 458; degree of 161, 257; models animations 247–8
of mechanisms 227, 232, 234 Anscombe, E. 131, 133, 135–6, 136–7, 140, 142
abundance 349 aortic arch asymmetry 340
accidental generalizations 272 apical constriction 335
action-formation mechanisms 403 Aplysia californica (sea mollusc) 379–80, 385–6
action potential 276–7 appropriateness 391–2
actions 419 aquatic sediments, biodiversity in 357–8
activities 116, 117, 118, 119–20, 151, 298; Aristotle 3, 17–20, 21, 23, 31, 33, 144; biology
activity-based approach to causation 136–7, 18–20; and Democritus 14, 15; the divinity 22;
140; by-what-activity experiments 355; relation two exhalations 17–18
with entities 120–4; varieties of 92, 94–6; see Armstrong, D.M. 149
also components Arrhenius, S. 66
activity-enabling properties 262 arrows 242–3, 247
activity probabilities 171–3, 179–80; as causal artificial intelligence (AI) 82–4
probabilities 175–6 artificial neural networks (ANNs) 82, 83, 275–6
activity signatures 262 asbestos 321–2, 323, 327
actualism desideratum of causation 140–1 Ashby, W. 81
additivity 277 aspirin 320–1, 323
affinitive organization 97, 353–5 associationistic models 78–9
agent-based simulation modeling (ABM) 406–7 atomism, early 14–16, 17
aggregates 186, 188–9 attractors 269, 270
aggregation explanations 292 automata 19–20
algorithmic level of analysis 389, 390, 392–4, 398 Axelrod, J.D. 455, 457
allelopathy 350, 351–2 axons 378–9
alternative medicine 322–3, 325
analog computers 442–3 Babbage, C. 79, 80
analysis, levels of see levels of analysis Bachelard, G. 32
Analytic Engine 79 backpropagation learning 394
analytical sociology 402 ‘backside attack’ mechanism 56
462
Index
backward chaining 262–3 bottom-up experiments 355–6

bacterial chemotaxis 169, 170–3, 175–6, 177, boundaries 4, 117, 124–8; mechanisms and natural
179–80, 369–70 kinds 199, 202–7; neuroscience 386; reality of
Barter, P.J. 321 126–8
Barthes, R. 180 bowtie networks 367
Batterman, R.W. 218, 293 Boyd, R. 198, 200–1
Bayesian approach 389, 390, 391, 392, 398 Boyle, R. 8, 26, 27–30, 59
Bayesian network models 177–8 brain: representations of 38, 39, 40; see also
Bechtel, W. 101, 127; abstraction 232; circadian neuroscience
rhythms 263, 363; complex localization 304; branching morphogenesis 335
defining mechanism 35; design principles Bray, D. 178, 179–80
396; diagrams 35–6, 37; dynamic mechanistic Brentano, F. 83
explanation 363; mechanistic explanation 35, broken mechanisms 105, 111–14
382, 392; metaphysical features of mechanisms Brücke, E. von 64
298–9; network models 275 Buckner, C. 394
Beebee, H. 164 Burian, R. 70
behavior functional decomposition 450–2 Butlerov, A. 49
behavior functions 449–52, 453, 454, 459 by-what-activity experiments 355
behavioral experiments 430 by-what-entity experiments 355
behavioral similarity 228, 230–1 by-what-property experiments 355
behaviorism 78–9 byproduct 105, 108
Bell, J. 452–3
Belle, M.D. 245–7 catalysts 47
below-ground competition 350, 351–2 causal account of physical computation 438
Bergson, H. 68 causal exclusion 396
Bernard, C. 64 causal imprecise probability 177
Berryman, S. 13, 24 causal inference 408, 428–9
Bertalanffy, L. von 68 causal-mechanical account of explanation 215,
Bhaskar, R. 401 216, 217
Bichat, X. 67–8 causal mechanism schemes 405–6
Biederman, I. 394 causal mechanisms 100, 402–3, 415; and statistical
Bier, E. 333 methodology 408–10
bimanual coordination 269–70, 271 causal narratives 417
biodiversity in aquatic sediments 357–8 causal powers 194
biology 3, 59–73; Aristotle’s 18–20; developmental causal probabilities 175–80
mechanisms 7, 98, 102, 332–47; ecology causal processes 134–5, 138, 139–40, 146
7, 348–61; evolutionary 6–7, 67, 296–307; causal role (CR) theory 108–10
holism and organicism 60, 61–2, 67–70; later causal scenarios 405–6
20th century 67; Loeb and the mechanistic causation 4–5, 94, 131–43, 163, 319; biomedicine
conception of life 65–6; mechanism and and contrastive focus 326–8; counterfactual
materialism 60–3; mechanistic philosophy in difference-making 134, 137–8, 140–1;
the 19th and 20th centuries 63–5; molecular Democritus 15; derivativeness 133, 136–7;
7, 67, 308–18; reevaluating mechanical evaluation of the contemporary debate
explanations 33–41; systemic turn 312; systems 139–41; Hume and 131, 132–3, 142, 144;
biology see systems biology interlevel 190–1; interventionist theory of 126;
biomedicine 7, 319–31; evidence and mechanical 135, 138–9, 141, 148; mechanisms,
contrastive focus 325–8; examples of counterfactuals and laws 5, 144–56; new
mechanisms 319–23; mechanisms’ mechanical philosophy 136–9; responses to
contribution to inferences 323–5 Hume 133–6, 142; taxonomy of tests for causal
biorobotics 263–4 hypotheses 410
biostatistical theory (BST) 108–9 CD4+ T-cells 367
biotic factors 349 cells 311–12
Boas, M. 30 cellular-physical mechanisms (CPMs) 7, 332–3,
Bogen, J. 136–7, 140, 152, 153, 161–2 335; integration of mechanisms 339–43, 344;
bonding, chemical 48, 53 not conserved 338–9
Borkovich, K.A. 179 ceteris paribus conditions 420
463
Index
chaining 262–3 computational explanation 442

Chalmers, A. 32 computational level of analysis 389, 390–2, 398
chance 5, 169–84 computational machines 79–81; Turing machines
characterization of mechanisms 2, 256–8 8, 79, 435–7, 439–40, 444, 445
Charleton, W. 33 computational mechanisms 8, 435–46; Church–
Chemero, A. 275, 276, 277 Turing thesis 435, 439–40; digital computation
chemical bonding 48, 53 441–2; generic computation 442–4;
chemical equations 46–7 mathematical theory of computation 435–6;
chemical reactions: classification of 206; origins of physical computation 437–9; rise of the new
the reaction mechanism 3, 46–58 mechanism 440–1
chemical thermodynamics 53–4 computational modeling: linking to mechanism
chemotaxis: bacterial 169, 170–3, 175–6, 177, diagrams 250–1; mental processes 82–4; systems
179–80, 369–70; in lobsters 264 biology 363
China 420; Cultural Revolution 417–18 computationally annotated mechanism diagram
Chinese medicine 322–3, 325 250–1
Chinese Room thought experiment 83 computer simulations: activity probabilities
Chirimuuta, M. 219 179–80; in mechanism discovery
cholesteryl ester transfer protein (CETP) 263–4
inhibitors 321 conceptual generality 140
Church, A. 435, 444 conditioned reflexes 377
Church–Turing thesis 435; scope of 439–40 connectionism 78, 389, 390
Cipollini, D. 355 connectionist machine learning 82–3
circadian rhythms 240–52, 263, 363 connectionist networks 275–6
circadian time (CT) 248 conserved MGMs 333, 336–9, 344
circularity 199, 207 conserved quantity theory 146–7, 149
Clerselier, C. 38–41 constitutive explanations 106–7, 216
cognitive modeling 392–4, 396–7 constitutive mechanisms 94, 124
cognitive science 7, 82–4, 239–40, 389–400; constitutive relationships 94; modeling 178–9
algorithmic level 389, 390, 392–4, 398; constitutive relevance 124, 125–6, 137
computational level 389, 390–2, 398; construction of a schema 258–9
implementational level 389, 390, 395–7, 398 contention 414, 417
Coleman’s boat diagram 252, 402–4, 409, 427 contextual desciptions 454
collision theory 53 contrastive focus 326–8
colors 243, 249–50 contrastive questions 452
comets 18 control engineering 455–7
communication 239 control parameters 269
co-morbidity 323 conventional complexes 244–5
comparative process tracing 429 conventionalism 199, 200
comparisons: across species 385–6; approaches Copeland, J. 439
to health and disease 322–3, 325; mechanistic corpuscles 28–9, 37–8
levels with other senses of 188–9; using corpuscularian philosophy 27–30, 31
diagrams 248–50 cosmos 21–2
competitive success 349–52 counterfactual difference-making 134, 137–8,
complementary medicine 322–3, 325 140–1
completeness, explanatory 219–20, 314–15, 458 counterfactuals 5, 144–56, 163; dynamical
complex causal structure (CCS) interpretations of generalizations and supporting 271–2
probability 174, 181 covering law (CL) model 191, 270, 405; approach
complex chance setups 176 to dynamical explanation 270–1; historical
complex localization 304 explanation 413–14
complexity 419, 458 Craver, C.F. 59, 116, 136, 138, 152, 171,
components 4, 116–30, 186–7, 298; evolutionary 353; constitutive relevance 125–6, 137;
biology 300, 301, 303; individuating 116–17, explanatory models 220; functions 109–10,
124–6; ontological nature of 117–20; relation 111; integration of mechanisms 340–1;
between 120–4; see also activities, entities, integrative unity 384; levels of analysis
interactions 392, 393–4, 396; metaphysical features of
composite systems 286–7, 291–3 mechanisms 298–9; model-to-mechanism
computational causal systems 100–1 mapping constraint 231, 393–4; natural kinds
464
Index
202–3, 204; phenomena 106; 339–43, 344; philosophical background and

relation between mechanistic examples 334–6
components 121–2, 123 DeWoskin, D. 245–7
crime rate 426–8 diagrams 6, 35–6, 238–54; circadian mechanisms
critical realism 401 240–52; comparing two or more mechanisms
crowding 359 248–50; components of 241–5; Descartes’ use
crystallins 205–6 of 37–41; functions served by 239–40; linking
Csete, M.E. 455 computational models to 250–1; representing
Cultural Revolution 417–18 time 247–8; space in 245–7
Cummins, R. 152–3, 393 dialectical materialism 61–2, 63, 71
cumulative advantage 405 Difference Engine 79, 80
cyanohydrin formation 50–2 difference-making, counterfactual 134, 137–8,
cybernetics 81 140–1
cystic fibrosis 106 differential equations 178
Digesting Duck (Vaucanson’s) 63, 76
DAGs (Directed Acyclic Graphs) 409 digital computation 441–2
Darden, L. 59, 116, 171, 255, 353; abstraction digital specificity 310–11
232; Anscombe’s approach to causation 136–7; digits 441–2
functions 110; integration of mechanisms Dijksterhuis, E.J. 30–1
340–1; metaphysical features of mechanisms dimensional realization 187
298–9; phenomena 106; relation between direct causal inference 428
mechanistic components 121–2, 123 discovery of mechanisms 6, 35, 239, 255–66, 304;
Darwin, C. 71 characterizing mechanisms and mechanism
Darwinism 63 schemas 256–8; computer simulations and
data graphics 252 biorobotics in 263–4; ecological mechanisms
date hubs 368 355–7; employing a schema type 260–1;
Datteri, E. 264 forward/backward chaining 262–3; guidance
Davidson, D. 98 from the phenomenon 259; localization 260;
Davies, J.A. 332 modular subassembly 261–2; neuroscience
De Caus, S. 36, 37 382–3, 383–6; stages in 258–9
de facto criticisms 32 diseases 113; see also biomedicine
de jure criticisms 32, 33 distribution 349
decoherence 289–90 divinity 21–2
decomposition 274–6, 382; functional divisibility 27, 28
decomposition models 450–4 DNA 260, 261–2, 263; mutations in replication
deductive-nomological (DN) model of 302–3
explanation 157–8, 215–16, 217 domain choice 205
degree distribution 366 Donohue, J. 429
democratization 405 dopamine homeostasis 106–7
Democritus of Abdera 14–15, 23 Doubly Decisive tests 410
dendrites 378–9 Dowe, P. 134–5, 140, 146, 148, 149
dense overlapping regulons 365 Doyle, J.C. 455
derivativeness 133, 136–7 Dretske, F.I. 149
Des Chene, D. 36–7 Driesch, H. 64–5, 68–9
Descartes, R. 26–7, 28–9, 30, 59; biological drift 301–2
writings 36–41; neuroscience 75–6, 376–7, Drosophila 334–5, 338, 342
381, 382, 383; reflex arc 75, 76, 376–7 Du Bois-Reymond, E. 64
Desch, C.H. 49 dualism 116, 121–4; arguments for 122–4
descriptions 226–7 Dupré, J. 96
descriptive adequacy 122 Dutrochet, H. 64
design principles 396; systems biology 365, dynamic polarization, law of 378
368–71 dynamic reorganization 371–2
designed-and-built etiologies 97–8 dynamical covering law (DCL) approach 270–1
determinism 173–4 dynamical explanation 217–18, 251, 267–80;
developmental mechanisms 7, 98, 102, 332–47; challenges for the mechanistic approach 274–7;
conserved genetic mechanisms 333, 336–9, 344; systems biology 362–4; varieties of approaches
integrating genetic and physical mechanisms to 270–4
465
Index
dynamical systems theory (DST) 6, 252, 267–80; evaluation of a schema 258–9

basic constructs 268–70 evidence-based medicine (EBM) 328
dynamicism 389, 390 evidence of mechanism 325–6
evolution 261
ecology 7, 348–61; defining 349; experimental evolutionarily conserved mechanisms 221–2
methods for mechanism discovery 355–7; evolutionary biology 6–7, 67, 296–307; central
group-level mechanisms 357–9; individual-level metaphysical features of mechanisms
mechanisms 349–57 298–300; drift 301–2; mechanistic framework’s
economics 8, 423–34; causal inference 428–9; contribution to understanding evolution
defining mechanisms in 424–5; methodological 303–4; mutation 302–3; natural selection
individualism 426–8; problem of extrapolation 233–4, 300–1
429–31; theoretical modeling of mechanisms evolutionary systems biology 370
425–6 evolved mechanisms 98–9, 102
effect functional decomposition 450, 452–4 experimentation 61, 382–3; ecological mechanism
effect functions 449, 450, 452–4, 459 discovery 355–7; orientation experiments 259;
effective field theories 286, 290–1 technological and behavioral experiments 430–1
Einstein, A. 287 explanandum phenomenon 391
electric power screwdriver 450, 451, 452 explanation 213; dynamical see dynamical
electronic theory of valence 52 explanation; epistemic and ontic 215, 234–5,
electrophiles 53 312–13; general philosophical account of
Eliasmith, C. 277 214–15; mechanistic see mechanistic explanation;
El-Samad, H. 455–7, 458 precursors to mechanistic explanation 215–16
Elster, J. 401–2 explanatory demarcation 214, 217
embedded data graphs 244 explanatory mechanisms 100
embryology 64–5 explanatory normativity 214, 217
embryonic induction 69–70 extrapolation, problem of 429–31
emergence 190–1
emergent behavior 292–3 fact, matters of 165–6
enantiomers 55, 56 failure analysis (malfunction explanation) 449, 450,
Enemy Release Hypothesis 352 452–4
engineering science 8, 447–61; mechanistic falling bodies 29, 159–60
explanation in 448–54; and systems biology Fast Enabling Links (FELs) 394
447, 454–9 feedback 81
enion probability analysis 292 feedforward loops 364–5
enslavement 292–3 Felline, L. 288
entities 116, 117–19, 298; activity-enabling fertilization 65–6
properties 262; by-what-entity experiments Firth, C.A.J.M. 178, 179–80
355; relation to activities 120–4; varieties Fisher, R.A. 67
of 92, 94–6; see also components flagella 169, 170
Entwicklungsmechanik 64, 65 flat realization 187
ephemeral mechanisms 98–9, 180, 299 flow diagram 268
Epicureans 16–17 Fontenelle, B. Le Bovier de 30
Epicurus 15, 16, 23 Forel, A. 378
epigenetics 327 Forger, D.B. 245–7
epistemic adequacy 122–3 forward chaining 262–3
epistemic emergence 190 France 64
epistemic explanation 215, 234–5, 312–13 free energy 53–4
epithelial folding 335 Freese, J. 407
Epstein, J. 407 Freud, S. 78
erraticity 176–7, 180 Froidmont, L. 26–7
Escherichia coli (E. coli): chemotaxis 169, Function Interpreting Language (FIL)
170–3, 175–6, 177, 179–80, 369–70; heat methodology 452–4
shock response system 455–7, 458 functional analysis 437
essentialism 199–200 functional biology 296–7
etiological explanations 106–7, 216 functional decomposition models 450–4
etiological mechanisms 124 functional homology 337
etiology, varieties of 92, 97–9 functional organization 245–7
466
Index
functionally indistinguishable component parts graphic symbols 244

275–6 Grew, N. 34
functions 4, 104–15, 152–3, 299, 440–1; behavior group-level ecological mechanisms 357–9
functions 449–52, 453, 454, 459; broken gut formation 342
mechanisms 105, 111–14; effect functions 449, Gutschoven, G. von 38
450, 452–4, 459; evolutionary biology 301,
302, 303; mechanisms and 107–11; mechanistic Haken–Kelso–Bunz (HKB) model 269–70, 271
role functions 111, 447, 448–54 Hall, J.C. 240, 241
fundamental laws 285, 286 Hamburger, V. 70
fundamental physics 283, 284–91 Hammett, L.P. 48, 50, 52, 55
Futuyma, D.J. 297 Hanson, N.R. 255–6
Hardin, P.E. 240, 241
Gabbey, A. 32–3 Harmer, S.L. 249–50
Galilean idealization 221, 227–8, 234 harmonious equipotential system 64–5, 68–9, 70
Galileo 159–60 Harré, R. 147–8, 401
Garson, J. 105, 443 Harrington, A. 68
Gassendi, P. 27, 28–9, 59 Hartwell, L.H. 455
gastric ulcer 322, 324 health 407–8; see also biomedicine
gastrulation 338 heart 37, 348
Gause, G.F. 351 heat shock response system 455–7, 458
gene expression 342; segment polarity network Hebb, D. 379, 381, 382
of 334–5 Hedgehog signaling pathway 334–5, 338, 343
gene–phenotype connections 257–8, 259, 260–1 Hedström, P. 402, 403, 408–9
General Relativity Theory (GRT) 288, 290 Helicobacter pylori 322, 324
generalization 208; accidental generalizations Helmholtz, L. von 64
272; dynamical generalizations 270–2; Hempel, C. 215, 413
generalizability of mechanistic explanations Henry, D. 19, 20
221–2; generalizability of social mechanisms hierarchy thesis for natural kinds 203
415–16; integrated developmental mechanisms Higgs mechanism 294
342–3, 344; molecular biology 313–14; high-density lipoprotein (HDL)-raising drugs 321,
regularities, mechanisms and laws 160–2 323, 327
generalized Law of Mass Action 50, 51 higher-level phenomena 192–4
generative explanation 406–7 His, W. 64, 378
generative mechanisms 147–8 historical explanation 8, 413–22; causal narratives
generic computation 442–4 417; covering law model 413–14; mechanisms
genetic determinism 309–10 and predictions 418–20; social mechanisms 8,
genetics 67, 70 414–17; social variation and mechanisms 417–18
Germany 64, 68 Hobbes, T. 59, 76
Gervais, R. 259 holism 60, 61–2, 67–70
Giere, R. 226 holistic materialism 61–3, 69
Gilbert, S. 62, 70 homeostasis 81
gill-siphon withdrawal reflex 380 homeostatic property cluster (HPC) 198–9, 200–1,
Glennan, S. 111, 353; counterfactuals 149–50, 202, 203, 207, 208
152; Democritus 14; dualism vs monism homogeneity 277
121–2; lowest level 194–5; mechanical homogeneous reference classes 327
theory of causation 131, 135–6, 138–9, 141, homology reasoning 336–9
148; metaphysical features of mechanisms Hooke, R. 30, 34
298–9; minimal mechanism 2, 105; modeling Hoop tests 410
mechanisms 225, 229–30; non-recurrent Hoover, K. 431
mechanisms 180; physics 285, 291 Hotelling’s model of spatial localization 425–6
global properties of networks 365–6 how-questions 392–4, 398
glyphs 241–3 Howick, J. 325
goal-directedness 110 hubs 366, 368
Godfrey-Smith, P. 227, 229 Hull, C. 79
Goldbeter, A. 250–1 Hume, D. 78; causation 131, 132–3, 142, 144;
Golgi, C. 78, 378 Humean roles for laws 164–6; responses to
grand predictions 418–19, 420 133–6, 142; skeptical realism 157, 164–5
467
Index
Hummel, J.E. 394 Jones, N. 367

Huneman, P. 366–7 Josephson, J. 256
Hutchins, B.R. 37
Huygens, C. 37 Kaiser, M. 152
hypercomputation 439–40 Kandel, E. 379–80, 381, 382, 383, 385
Kaplan, D.M. 220, 231, 393
iconic symbols (icons) 243–4 Kay, S.A. 249–50
idealization 278, 313, 394; Galilean 221, Kim, J. 396
227–8, 234; idealized explanatory models Kincaid, H. 426, 428
219–21; minimal 221, 227–8, 234; models of Kistler, M. 141
mechanisms 227–8, 232–3, 234 Knight, C.R. 409
ideas, relations of 165–6 Kuhlmann, M. 194–5, 291
Illari, P. 123–4, 323, 325, 353, 448 Kuorikoski, J. 100–1
ILLUMINATE trial 321
‘imagined concrete’ objects 229, 230 La Forge, L. de 38
implementational level of analysis 389, 390, La Mettrie, J. de 77
395–7, 398 Lange, M. 288
imprecise probability 177 LaPorte, J. 200
increased population size mechanism 358–9 Lapworth, A. 49; kinetic studies of reaction
indirect causal inference 428 mechanism 50–2
individual-level ecological mechanisms 349–57; Law of Mass Action 50, 51
experimental methods for discovery 355–7; and Law of Odd Numbers 159–60
the minimal account 351–5 laws 94; mechanisms, counterfactuals and 5,
individuation of components 116–17, 124–6 144–56; physics 283, 285–6; recent history of
induced organization 97, 353–5 laws and mechanisms 157–9; regularities, laws
industrial capitalism 63 and mechanisms 5, 157–68; two Humean roles
inference: causal 408, 428–9; medical inferences for 164–6
323–5 learning 341, 385–6; backpropagation learning
information-processing functions 384, 390–1, 393 394; history of neural mechanisms of 375–80;
information theory 81, 82 see also cognitive science
Ingold, C.K. 53, 54–5, 56 Leibniz, G.W. 30, 34, 76
inheritance 260 length contraction 287–8
innate reflexes 377 Leuridan, B. 152
instantiation 257, 395 levels, mechanistic see mechanistic levels
integral feedback control 369 levels of analysis 7, 389–400; algorithmic 389, 390,
integration of mechanisms: developmental 392–4, 398; computational 389, 390–2, 398;
mechanisms 339–43, 344; neuroscience 384–6 implementational 389, 390, 395–7, 398
interactions 116, 117, 118, 298; counterfactuals levels of organization 62–3, 398
149–51; strength of 125; varieties of 92, 94–6; Leverere, M. 292
see also components levers 20–1
interdependency thesis 121–2 Levins, R. 71
interlevel causation 190–1 Levitt, S. 429
interlevel integration 341 Levy, A. 100, 101, 232, 304
intermediates 47, 49, 56 Lewis, D. 131, 133–4, 135–6, 138, 190
intermolecular interactions 48, 53 Lewontin, R. 71
interventionist theory of causation 126 light, competition for 349–52, 354
intramolecular interactions 48, 53 lignin 364
intuitive understanding 313 limit sets 268–9
INUS conditions 126, 187 limited pancomputationalism 438
invisible hand 405 lines 242–3
irreducibility thesis 121–2 Link, B.G. 407
irregular mechanisms 152–3 lobster chemotaxis 264
isolated descriptions 454 localization 397; complex 304; Hotelling’s model
of spatial localization 425–6; strategy for
jade 200 discovering mechanisms 260, 274–6
James, W. 78 Locke, J. 78
Janssen, M. 287 Loeb, J. 65–7
468
Index
logic 158 Mechanics, the 20–1

Long-Term Potentiation (LTP) 191 Mechanism 2.0 101
Lonicera maackii (Amur honeysuckle) 7, Mechanism1 and Mechanism2 100
349–52, 353, 354–6 mechanism diagrams see diagrams
Lorentz, H. 287 mechanism schemas see schemas
Lotka, A. 79 mechanism sketches 239, 394
Love, A. 233 mechanisms-for 145–8, 149–54
lowest level 194–5 mechanisms-of 145–51, 154
Lucas critique 430, 431 mechanistic explanation 5, 35, 213–24; and
Lucretius 16–17 cognitive science 389–400; competing
Ludwig, C. 64 perspectives on the explanatory power of
Luken, J.O. 355–6 457–8; DST and 273–8; dynamic see
lump and split methodology 202, 203–5 dynamical explanation; in engineering science
Lutfey, K. 407 448–54; limits of 217–22; in molecular biology
Lüthy, C. 37 312–15; nature of 216–17; neuroscience
381–3; precursors to 215–16; social
Machamer, P.K. 59, 116, 136–7, 171, 353, sciences 404–5
448; functions 110; metaphysical features mechanistic levels 5, 185–97; comparisons and
of mechanisms 298–9; relation between contrasts 188–9; emergence and interlevel
mechanistic components 121–2, 123 causation 190–1; explication of 186–8; lowest
Mackie, J. 146 level 194–5; ontological status of higher-level
MacLeod, M. 364 phenomena 192–4; reduction 191–2
macro behavior, micro dynamics and 291–3 mechanistic requirement 425
Magendie, F. 64 mechanistic role functions 111, 447, 448–54
maintaining relationship 106–7 medicine see biomedicine
malfunction explanation 449, 450, 452–4 medium independence 443–4
mammalian circadian clock mechanism 240–50 memory 341; history of neural mechanisms of
Mangold, H. 69 375–80
manipulation 304 mental, the 3–4, 74–88, 437, 438; associationistic
manipulationism 137–8 models and early robotic models 78–9;
mapping accounts of physical computation 438 cognitive science see cognitive science;
mark-transmission 134–5, 146 computation and cybernetics 79–81;
market mechanism 423 computational models and artificial intelligence
Markov process models 178 82–4; early mechanists 75–7; neuroscience see
Marr, D. 7, 389–400 neuroscience
Mass Action, Law of 50, 51 metamechanisms 407–8
materialism 3, 59–73; dialectical 61–2, 63, 71; metaphysical intuitions 123
holistic 61–3, 69 metaphysical mechanists 2
mathematical objects 229 Méteores (Descartes) 37–8
mathematics 435; see also computational meteorology 17–18
mechanisms methodological individualism 426–8
matters of fact 165–6 methodological mechanists 2
Maxwell’s laws 161 micro dynamics 291–3
Mayr, E. 296–7 Millstein, R.L. 233, 357
McAdam, D. 414, 415–16, 417 mind–body problem 74
McCulloch, W. 436 minimal idealization 221, 227–8, 234
McGinnis, W. 333 minimal mechanisms 2, 92, 105, 201–2, 323;
McKusick, V. 204 individual-level ecological mechanisms 351–5
McMullin, E. 32 minimal model explanations 218
mechanical philosophy 3, 26–45, 59; challenging miscomputation 442
31–3; classifying 2–3; establishing 26–31; new Mitchell, R.J. 356, 357
mechanistic philosophy 1, 59, 308, 440–1; Mitchell, S.D. 99
re-emergence 1; reevaluating mechanical mixed mechanisms 323–4, 328
explanations 33–41; scope of mechanical model organisms 221
philosophies 2–3 model population 226
mechanical theory of causation 135, 138–9, model-to-mechanism mapping (3M) constraint
141, 148 231, 393–4
469
Index
models/modeling 6, 225–37; cognitive neural networks 206; artificial (ANNs) 82, 83,
modeling 392–4, 396–7; computational 275–6
see computational modeling; constitutive neuron doctrine 382
relationships 178–9; mechanisms in economics neuroscience 7, 67, 375–88; discovery of
425–6; in mechanistic science 229–33; mechanisms 382–3, 383–6; history of neural
the mental 3–4, 74–88; network models mechanisms of learning and memory 375–80;
364–8; phenomenological models 228, philosophical accounts of mechanisms in 381–3
230–1; probabilistic mechanisms 177–8; new mechanistic philosophy 1, 59, 308, 440–1
relevance for contemporary debates 233–5; Newell, A. 82
and scientific representation 226–9; systems Newton, I. 30, 31, 32
biology 362–74 Newton’s laws 160
modest predictions 419–20 Nicholson, D.J. 100
modular subassembly 261–2 NMDA receptors 191
modularity 185, 312, 372 no-overlap principle 203
molecular biology 7, 67, 308–18; defining 308–9; noise-free specificity 310–11
mechanistic explanations in 312–15; nature of nomological explanations 32–3
molecular mechanisms 309–12 nomologicity 157–8, 215–16
molecular clock mechanisms 240–52 noncanonical Wnt signaling pathway 343–4
molecular genetic mechanisms (MGMs) 7, 332–3, non-causal explanations 284, 288
334–5; conserved 333, 336–9, 344; dynamic non-constitutive mechanisms 94
constitution and organization of 343–4; nonlinearity 276–7
integration of mechanisms 339–43, 344 non-recurrent mechanisms 180
MONIAC 231 non-reductionistic account of causation 139–40
monism 116, 121–2 non-strict laws 272
moonlighting proteins 205 normativity of function 113
More, H. 27, 34 nucleophiles 53
Morgan, G.T. 49 number of parts 96–7, 99
Morgan, T.H. 67 Nussbaum, M. 24
morphogenetic mechanisms 332–3 nutrients, competition for 350, 351–2
mosaic view of causality 319
Moss, L. 110 Odd Numbers, Law of 159–60
movement: Aristotle 18–20; Democritus 15 On the Cosmos 21–2
multifield explanations 383, 384–6 ontic emergence 190
multi-level approach 383–6 ontic explanation 215, 234–5, 312–13
multiple desiderata 220–1 ontological program 27–30
multiple realization 194 operations 116, 117, 298; evolutionary biology
mutation 302–3 300, 301, 303; see also activities, interactions
Myung, J. 245–7 Oppenheim, P. 189
order-generating information 309–12
narratives 325–6; causal 417 organicism 60, 61–2, 67–70
Nathan, M. 233 organization 299; ecological mechanisms
National Institute of Dental and Craniofacial 353–5; evolutionary biology 300–1, 302, 303;
Research 343 functional vs physical in mechanism diagrams
natural boundaries 124–5 245–7; levels of 62–3, 398; temporal 273–4;
natural kinds 5, 198–210, 386; circularity 199, varieties of 92, 96–7
207; overlapping problem 199, 203–7; organizational emergence 190
philosophical approaches to 199–203 orientation experiments 259
natural selection 233–4, 300–1 Otto, K.N. 450–2
Naturphilosophie movement 68 overlapping problem 199, 203–7
necessity 132, 133
negative feedback 81, 96 Palmer, M.A. 357–8
Nelson, C.M. 335, 338 pancomputationalism 437; limited 438;
Nersessian, N.J. 364 unlimited 438
nerves 322–3 Panda, S. 249–50
nervous system 75, 76, 78–9, 376–80 parity thesis 121–2
network models 364–8 Parker, S. 30
network motifs 364–5, 370, 372 Parmenides 14
470
Index
parsimony 123–4 producing relation 106

parthogenesis 65–6 productivity of mechanisms 123
party hubs 368 properties 353; by-what-property experiments
patellar reflex 106 355; homeostatic property cluster (HPC)
path analysis 356, 357 198–9, 200–1, 202, 203, 207, 208
path dependency 419 property correlations 200–1
Pavlov, I. 377, 381, 382, 383, 386 protein classification 204–6
Perceptron 82 protein synthesis 93, 94, 204, 262–3, 340
Perini, L. 239 proximate causes 296–7
Perrault, C. 29 Psillos, S. 140
Phelan, J. 407 puerperal fever 322, 324
phenomena 4, 104–15, 256; broken mechanisms purpose functions 449
105, 111–14; characterizing the phenomenon pushy motion 169, 170
258; guidance from in discovery 259; higher- Putnam, H. 189
level 192–4; how mechanisms have phenomena
106–7; relationships between components, qualia 84
mechanisms and 124–8; varieties of 92–4 Quantitative Structure Activity Relations
phenomenological models 228, 230–1 (QSAR) 206
phenotype–gene connections 257–8, 259, 260–1 quantum holism 291
phylogenetic reconstruction 336 quantum mechanics 194–5, 288–90, 291
physical computation 437–9
physical organization 245–7 Ramón y Cajal, S. 78, 378–9, 382, 383, 386
physicalism 61 random networks 366, 368
physics 6, 283–95; of composite systems 286–7, Ratcliffe, S. 206
291–3; effective field theories 286, 290–1; rational analysis 391
fundamental 283, 284–91; limitations to rational-choice theory 424, 425, 426, 430, 432
mechanistic reasoning in 283–4; quantum reaction coordinate 47, 57
mechanics 194–5, 288–90, 291; relativity reaction mechanism 3, 46–58; Lapworth’s kinetic
theory 287–8, 290 studies 50–2; structure theory 48–9; theories of
Piggins, H.D. 245–7 reaction rates 53–5; Walden inversion 55–7
pineal gland 38, 40, 41 realism 99; and boundaries 126–8; about natural
Pitts, W. 436 kinds 200; pluralist 99; skeptical 157, 164–5
platelet function 320–1 realization: higher-level phenomena 192–4; levels
Plato 23 of 187; multiple 194
pluralism 114 reasoning: in discovering mechanisms 255–66;
pluralist realism 99 evidence, narratives and in medicine 325–6
policy interventions 429–30 reconstitution of a phenomenon 107
political science 402, 409–10 recurrence 93–4, 162
population regulation mechanisms 359 reductionism 1, 61, 66
population size increase mechanism 358–9 reductive explanation 191–2, 304, 383
positive feedback 81, 96 reflexes 78; Descartes and the reflex arc 75, 76,
Power, H. 30, 34 376–7; patellar reflex 106
power-ascription 147 regularity 93–4, 299–300; evolutionary biology
prediction 215–16; grand and modest predictions 301, 302, 303; laws, mechanisms and
418–20 regularities 5, 157–68; molecular biology
predictivism 272–3 313–14; view of causation 132–3
preference reversal 430 Reisel, S. 77
probability 5, 169–84; activity probabilities Reiss, J. 423, 424
171–3, 175–6, 179–80; chemotaxis in E. coli relations of ideas 165–6
169, 170–3, 175–6, 177, 179–80; determinism relativity theory 287–8, 290
and interpretations of 173–4; epistemology renormalization group explanations 293
177–80; erraticity and imprecise probability Reppert, S.M. 241, 242
176–7; metaphysics 173–7; modeling representation 442; mental 83–4; models and
probabilistic mechanisms 177–8; non- scientific representation 226–9
recurrent mechanisms 180 response nonlinearity 277
process nonlinearity 276–7 retroviruses 261
process tracing 409–10, 429 reverse engineering 185–6, 449, 450–2, 454
471
Index
reverse transcripterase 261–2 Simon, H. 82, 185

revision of a schema 258–9 simple mapping account 438
Rice, C.C. 218 simple mechanistic integration 340, 341
Richardson, R.C. 35, 275, 304, 382 single-input network motifs 365
RNA 261–2, 263, 340 singular causation 151
robotic models 79 situational mechanisms 403
robots 81 size levels 189
robust perfect adaptation (RPA) 368–70 skeptical realism 157, 164–5
robustness 366, 458 Skinner, B.F. 67
role functions 111, 447, 448–54 Skipper, R. 233
Rosbash, M. 240, 241 small-world networks 365–6
Ross, L. 218 Smith, A. 423
Roux, W. 64 Smoking-Gun tests 410
Royal Society 30, 34 Sober, E. 339
rule-based approach to artificial intelligence 82–3 social mechanisms 8, 401–12, 424; agent-based
Russell, B. 79 simulation and generative explanation 406–7;
causal scenarios and causal mechanism schemes
Sachs, J. 65, 66 405–6; Coleman’s diagram 402–4; historical
Salmon, W. 131, 134–5, 135–6, 138, 139–40, explanation 8, 414–17; mechanism-based
146, 148 explanations 404–5; metamechanisms 407–8;
Santos, A. 430 statistical methodology 408–10
Sarkar, S. 62, 70 Social Mechanisms (ed. Hedström and
scale-free networks 366–8 Swedberg) 402
Schelling, T. 404 social scientific mechanical philosophy 1
schema types 239; employing in discovery 260–1 social variation 417–18
schemas, mechanical 239, 256–8; constructing socioeconomic status (SES) 407–8
258–9; evaluating 258–9; revising 258–9 soul 15
Schooten, F., the Younger 37 space: in mechanism diagrams 245–7; state
Schrödinger, E. 309 space 268
Schrödinger equation 288–9 Special Relativity Theory (SRT) 287–8
Schuyl, F. 38–41 specification 257
scientific representation 226–9 specificity 458; of binding 310–11; digital 310–11
Scientific Revolution 26, 30 Spemann, H. 69–70
scientific unity 189 spontaneous symmetry breaking 290
scope of mechanical philosophies 2–3 St. Germain Royal Gardens 376
sea urchins 64–5, 65–6 stability 160–1
Searle, J. 83 Stahl, G.E. 34
seed dispersal 354 standard model of particle physics 286, 294
segment polarity network of gene expression standardized periodizations 342–3, 344
334–5 state space 268
selected effects (SE) theory 108–10, 113 state-space trajectories 390–1, 393
semantic accounts of physical computation 438–9 statistical methodology 408–10
semen 19 Statua Humana Circulatoria 77
Semmelweis, I. 322, 324 Steel, D. 424, 428, 429–30
sensitization 380 Stent, G. 67
sequential intertemporal integration 341 stereochemistry 55–7
sequential substitution mechanism 55–7 stimulation experiments 355–6
Shagrir, O. 391, 396 stochastic fluctuations 310–11
Shannon, C. 81 stochastic mechanisms 171; see also probability
shapes 242, 249 StochSim 178
Shapiro, H. 322–3 strategic mechanism 100, 304
Sherrington, C.S. 78, 379, 381, 382, 383 Straw-in-the-Wind tests 410
shooting stars 18 strength of interactions 125
Sieg, W. 439 Strevens, M. 292
Silberstein, M. 275, 276, 277 string theory 166
similarity 221–2, 228, 230–2; behavioral 228, structural equation modeling (SEM) 252, 356, 357
230–1; structural 228, 230–2, 234 structural explanations 32–3
472
Index
structural formulas 46–7, 49 Turing, A. 79–81, 435–6

structural similarity 228, 230–2, 234 Turing-computable functions 436, 439
structure-altering interventions 429–30 Turing machines 8, 79, 435–7, 439–40, 444, 445;
structure theory 48–9 universal 435–6, 436–7
structure-transference 146 Turing test 80
sub-reactions 46–7, 48 Turing-uncomputable functions 436, 437, 439–40
substitution reactions 55–7 twist gene 342
succession 354 two exhalations 17–18
Swampman 98 two-ring cooking hotplate 453
Swedberg, R. 402, 403 type identity 193
swerves 15, 16
synapses 379 Uexküll, J. von 68–9
synchronous substitution mechanism 55–7 ultimate causes 296–7
synthetic biology 371 uncertainty 429–30
systemic turn 312 underlying relation 106
systems biology 7, 362–74; design principles 365, unemployment 426–8
368–71; dynamic mechanistic explanation uniformity of parts 97
362–4; engineering science and 447, 454–9; universal behavior 292–3
network models 364–8 universal Turing machines 435–6, 436–7
universe 21–2
Tabery, J. 140 unlimited pancomputationalism 438
Takumi, T. 245–7
Tamburrini, G. 264, 439 valence 52
Tar receptors 170 variation, social 417–18
target 105, 108 varieties of mechanisms 4, 91–103; dimensions 92;
Tarrow, S.G. 414, 415–16, 417 entities, activities and interactions 92,
taxonomy: of tests for causal hypotheses 410; 94–6; etiology 92, 97–9; organization 92, 96–7;
varieties of mechanisms see varieties of and other taxonomies 100–1; phenomena
mechanisms 92–4
technological experiments 430–1 Varner, V.D. 335, 338
temporal organization 273–4 Vaucanson, J. de 76
text 243 Vaucanson’s Duck 63, 76
Thagard, P. 260–1 vector field 268
theories of reaction rates 53–5 Vermaas, P.E. 449
thermodynamics 292 Virchow, R. 65
Thorndike, E. 78 visual representations 238–54; see also diagrams
Tilly, C. 414, 415–16, 417 vitalism 3, 59–73
time: circadian 248; integrated developmental Vogt, K. 64
mechanisms 341–3; and network models volition 323
368; representing in static diagrams 247–8; Von Neumann, J. 436–7
representations and developmental mechanisms vulnerability 367
339–40; standardized periodizations 342–3,
344; temporal organization 273–4 Waddington, C.H. 70
time dilation 287–8 Walden inversion 55–7
Tolman, E. 79 Walter, W.G. 81
Tomlin, C.J. 455, 457 Waskan, J. 140–1
Tooley, M. 149 water, competition for 350, 351–2
top-down experiments 356 Watson, J.B. 67
topological explanations 366–7 Weaver, D.R. 241, 242
torcetrapib 321 Weber, E. 259, 292
transcription-translation feedback loops 240, 241, Weisberg, M. 220–1, 227–8
246, 249, 251 Weiskopf, D.A. 394
transition state theory 53–5 Weiss, P.A. 70
transformational mechanisms 403–4 Wenbourne gearbox model 230
trigger-effect descriptions 452–3 Westfall, R. 31
tumbly motion 170 what–questions 390–1, 392, 398
Turchin, P. 353 wheel of fortune 174, 181
473
Index
where–questions 395–7, 398 Wnt signaling pathway 335, 336, 338, 343–4
why–questions 391–2, 398 Wood, K.L. 450–2
Williams, G.C. 109 Woodward, J. 126, 137–8, 149–50, 270
Williamson, J. 123–4, 139, 323, 353, 448 Worrall, J. 328
Willis, T. 34, 75 Wouters, A. 370
Wilson, K. 293 Wundt, W. 78
Wimsatt, W.C. 127, 189
Winship, C. 409 Yi, T.-M. 369
474
cytoplasm
nucleus
Clk
?
+ Bmal1
mCry
+
P mPer1
C B
mPer2
mPer3
−
Plate 1 A diagram of the molecular clock mechanism for circadian rhythms in mammals, exemplifying
several types of elements used in mechanism diagrams. From Reppert and Weaver 2001, Figure 1.
Reproduced with permission of Annual Review of Physiology, Volume 63  2001 by Annual
Reviews, http://www.annualreviews.org.
Output
Input
VIP
Ca2+ VP
VIP AC
2
Ca2+ CREB
V Per 1/2
BCRE EBOX
Cry 1/2
(RRE) EBOX mRNA
K +
GABA
Plate 2 A mechanism diagram showing some of the key parts of the molecular clock mechanism
and how, in SCN neurons, they interact with pathways extending into the extracellular
environment. From DeWoskin, D., Myung, J., Belle, M. D., Piggins, H. D., Takumi, T., &
Forger, D. B. (2015) “Distinct roles for GABA across multiple timescales in mammalian
circadian timekeeping,” Proc Natl Acad Sci U S A., Figure 1.
CT6
SCN firing
zz z
zz z
mRNA
CB
Anticipated drawn
Anticipated dusk
scriptio
an
na
p Tr
CT0 CB CB CT12
l fe
ed
back loo
CB
protein
SCN firing
CT18
Current Biology
Plate 3 Using a 24-hour clock face to show the timing of behavioral and neural activity (gray region)
and the underlying molecular clock (inner circle). Appropriately located iconic symbols contrast
the activity of humans (diurnal) with that of mice (nocturnal). Reprinted from Current Biology,
Vol 18, Hastings, M. H., Maywood, E. S., & O'Neill, J. S., Cellular circadian pacemaking and
the role of cytosolic rhythms, Figure 2, copyright (2008), with permission from Elsevier.
A KaiB C
PER TIM
per/tim
KaiC
KaiA
CYC CLK ?
? clk
KaiA KaiB KaiC
outputs
CRY1 CRY2
B FRQ D
wc-1 mcry1,2 PER1 PER3
PER2 ?
WC-1 WC-2
mper1,2,3
outputs ? ? frq
?
BMAL CLK ?
bmal
outputs
Plate 4 Comparison of the molecular clock mechanisms, as understood in 2001, in (A) cyanobacteria,
(B) fungi, (C) fruit flies, and (D) mice. From Harmer, S. L., Panda, S., and Kay, S. A. 2001,
Figure 2. Reproduced with permission of Annual Review of Cell and Developmental Biology,
Volume 17  2001 by Annual Reviews, http://www.annualreviews.org.

The Routledge Handbook of Mechanisms and Mechanical Philosophy 1nbsped 1138841692 978-1-138 84169 7 978 1 315 73154 4 Compress

Uploaded by

Copyright:

Available Formats

The Routledge Handbook of Mechanisms and Mechanical Philosophy 1nbsped 1138841692 978-1-138 84169 7 978 1 315 73154 4 Compress

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Routledge Handbook of Mechanisms and Mechanical Philosophy 1nbsped 1138841692 978-1-138 84169 7 978 1 315 73154 4 Compress

Uploaded by

Copyright:

Available Formats

THE ROUTLEDGE HANDBOOK OF

MECHANISMS AND MECHANICAL

• Historical perspectives on mechanisms

The Routledge Handbook of Philosophy of Pain

The Routledge Handbook of Brentano and the Brentano School

The Routledge Handbook of Metaethics

The Routledge Handbook of Philosophy of Memory

The Routledge Handbook of Evolution and Philosophy

The Routledge Handbook of Mechanisms and Mechanical Philosophy

ISBN: 978-1-138-84169-7 (hbk)

1 Introduction: mechanisms and mechanical philosophies 1

2 Mechanisms: ancient sources 13

3 From the mechanical philosophy to early modern mechanisms 26

4 The origins of the reaction mechanism 46

5 Mechanism, organicism, and vitalism 59

6 Mechanisms and the mental 74

8 Mechanisms, phenomena, and functions 104

9 The components and boundaries of mechanisms 116

10 Mechanisms and the metaphysics of causation 131

11 Mechanisms, counterfactuals, and laws 144

12 What would Hume say? Regularities, laws, and mechanisms 157

13 Probability and chance in mechanisms 169

14 Mechanistic levels, reduction, and emergence 185

15 Mechanisms and natural kinds 198

16 Mechanistic explanation and its limits 213

17 Models of mechanisms 225

18 Explaining visually using mechanism diagrams 238

19 Strategies for discovering mechanisms 255

20 Mechanisms and dynamical systems 267

21 Mechanisms in physics 283

22 Mechanisms in evolutionary biology 296

23 Mechanisms in molecular biology 308

24 Mechanisms and biomedicine 319

25 Developmental mechanisms 332

26 Mechanisms in ecology 348

27 Systems biology and mechanistic explanation 362

28 Mechanistic explanation in neuroscience 375

29 Mechanisms in cognitive science 389

30 Social mechanisms 401

31 Disaggregating historical explanation: the move to social mechanisms 413

32 Mechanisms in economics 423

33 Computational mechanisms 435

34 Mechanisms and engineering science 447

Garland E. Allen is Professor Emeritus of Biology at Washington University in St. Louis. He

Holly Andersen is Associate Professor in the Philosophy Department at Simon Fraser

Brendan Clarke is Lecturer in History and Philosophy of Medicine at the Department

Carl F. Craver is Professor of Philosophy in the Philosophy-Neuroscience-Psychology

Lindley Darden (Ph.D., conceptual foundations of science, University of Chicago) is Professor

Lane DesAutels is Assistant Professor in the Department of Philosophy and Religion at

William Goodwin is Associate Professor in the philosophy department at the University of

Sara Green is a postdoctoral fellow at the Department of Science Education, University of

Stavros Ioannidis is a philosopher of biology and currently a postdoctoral researcher at the

Meinard Kuhlmann, Professor of Philosophy at Mainz University in Germany, received dual

Daniel Little is Chancellor and Professor of Philosophy at the University of Michigan-Dearborn

Caterina Marchionni is Academy Research Fellow at the Academy of Finland Centre of

Maureen A. O’Malley is a philosopher of biology at the University of Bordeaux, France. She

Tiberiu Popa is Associate Professor of Philosophy at Butler University (Indianapolis, USA).

Mark Povich is completing his Ph.D. in the Philosophy-Neuroscience-Psychology Program at

Federica Russo is Assistant Professor at the philosophy department, the University of