Vagueness and Degrees of Truth
Vagueness and Degrees of Truth
Vagueness and Degrees of Truth
Nicholas J. J. Smith
1
1
Great Clarendon Street, Oxford
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide in
Oxford New York
Auckland Cape Town Dar es Salaam Hong Kong Karachi
Kuala Lumpur Madrid Melbourne Mexico City Nairobi
New Delhi Shanghai Taipei Toronto
With offices in
Argentina Austria Brazil Chile Czech Republic France Greece
Guatemala Hungary Italy Japan Poland Portugal Singapore
South Korea Switzerland Thailand Turkey Ukraine Vietnam
Oxford is a registered trade mark of Oxford University Press
in the UK and in certain other countries
Published in the United States
by Oxford University Press Inc., New York
Nicholas J. J. Smith 2008
The moral rights of the author have been asserted
Database right Oxford University Press (maker)
First published 2008
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
without the prior permission in writing of Oxford University Press,
or as expressly permitted by law, or under terms agreed with the appropriate
reprographics rights organization. Enquiries concerning reproduction
outside the scope of the above should be sent to the Rights Department,
Oxford University Press, at the address above
You must not circulate this book in any other binding or cover
and you must impose the same condition on any acquirer
British Library Cataloguing in Publication Data
Data available
Library of Congress Cataloging in Publication Data
Data available
Typeset by Laserwords Private Limited, Chennai, India
Printed in Great Britain
on acid-free paper by
Biddles Ltd., King’s Lynn, Norfolk
ISBN 978–0–19–923300–7
10 9 8 7 6 5 4 3 2 1
For my parents, Sybille and Vivian Smith, with love and gratitude
Acknowledgements
Introduction 1
Part I. Foundations
. Beginnings 15
1.1 Toolkit 17
1.2 The Classical Semantic Picture 24
Conclusion 317
References 321
Index 333
Introduction
Ordinarily we say that persons are vague if they regularly misplace their
car keys, stare blankly into the fridge having forgotten why they opened
it, and so on. In philosophy, however, the term ‘vague’ has a different
use. It applies primarily (although not exclusively) to predicates.¹ Amongst
all predicates, the vague ones are usually singled out in one or more of
three ways:
Borderline cases. There are some persons to whom the predicate ‘is tall’
clearly applies (e.g. most professional basketball players) and some to whom
it clearly does not apply (e.g. most professional jockeys), but then there are
other persons to whom it is unclear whether or not the predicate applies
(I’m sure you know some of them). When asked whether such a person
is tall, we tend to react with some sort of hedging response: ‘‘sort of’’, a
shrug and a certain sort of scowl or exhalation of breath, a blank look, etc.
Call these persons borderline cases for ‘tall’. Other predicates—for example,
‘is a prime number’—do not have borderline cases. Julius Caesar is not a
prime number, seven is a prime number, eight is not a prime number, and
so on: there is nothing at all of which it is unclear whether it is a prime
number.² One standard way of drawing the distinction between vague
and non-vague (or precise) predicates is to say that vague predicates have
borderline cases, while precise predicates do not.³
Blurred boundaries. Imagine a line drawn around all the things to which a
given predicate applies. One typical characteristic of vague predicates is that
¹ Predicates are items of language—e.g. ‘is tall’, ‘is happy’, ‘is running’—which go together with
names—e.g. ‘Bill’, ‘Ben’, ‘Alice’—or definite descriptions—e.g. ‘the tallest woman in the room’, ‘the
inventor of post-it notes’, ‘the man who threw the egg’—to form sentences—e.g. ‘Bill is tall’, ‘The
inventor of post-it notes is happy’, ‘Alice is running’.
² For large numbers, we may suppose that we can use a computer to help us determine the answer.
³ The borderline case characterization can be traced at least as far as Peirce 1902.
2
⁴ The blurred boundaries characterization can be traced at least as far as Frege’s statement that if we
represent concepts in extension by areas on a plane, then vague concepts do not have sharp boundaries,
but rather fade off into the background (Grundgesetze, ii, §56; trans. in Beaney 1997, 259).
⁵ More precise characterizations of Sorites series and Sorites paradoxes will be presented below in
§3.5.4.
3
many accounts, and no apparent way of deciding which (if any) of them
is correct.
Part I: Foundations
As a first step towards rectifying this situation, in Chapter 2—building
on basic preliminary material which it is the purpose of Chapter 1 to
present—I give an overview of the space of possible theories of vagueness,
and show where existing theories live in this space. For we cannot begin
to select a theory of vagueness until we fully understand how the different
theories differ from one another: not just over particular matters of detail,
but at a more fundamental and illuminating level of analysis. Accordingly,
Chapter 2 provides a conceptual map of theories of vagueness, setting
out the important axes along which different theories vary, and then
giving the coordinates of existing and future theories on these axes. At the
centre of the map is the classical view of vagueness, according to which
vague language is—from a semantic point of view—exactly the same as
precise language. This view applies the standard model theory for precise
mathematical language to vague language. I begin by isolating the key
elements of this classical model theory: these key elements give the axes of
the space. With the classical view at the origin, other theories are located
according to which of the key elements they modify, and then, at a finer
level of analysis, according to how they modify them.
It is in this chapter that a distinction which plays a central role in
this book first emerges: the distinction between worldly vagueness and
semantic indeterminacy. Here is a quick sketch of how it emerges. Model
theory involves two aspects: a part which represents a language, and a
part—a model —which represents the world. The process whereby items
of language gain meaning is represented by matching up a language with
a model—which is called giving an interpretation of the language. The
underlying thought is that our words gain meaning when we use them to
talk about the world. But language is versatile, and words could mean many
things. Consider the sentence ‘The house is on the hill’. The symbol ‘house’
could have been used to mean what the symbol ‘cat’ actually means, and the
symbol ‘hill’ could have been used to mean what the symbol ‘mat’ actually
means, so there is an interpretation of the sentence on which it means
5
that the cat is on the mat. But the interpretation on which the sentence
means that the house is on the hill is privileged: it corresponds to what
the sentence actually means. Such an interpretation is called the intended
interpretation. In isolating the key elements of the classical model theory
for precise language, I make a broad distinction between the internal nature
of classical models, and the external part of the classical story, according
to which each discourse has a unique intended interpretation. The classical
picture tells us on the one hand that the world is of such and such a sort (it
has a structure which can be represented by a classical model, and classical
models have such and such internal features), and on the other hand that
the meaning of anything we say is correctly specified by giving a single
such model. This distinction yields two broad ways in which a theory of
vagueness can differ from the classical view: it can replace classical models
with non-classical models with different internal features, or it can alter
the external story according to which only one model comes into play in
describing what some utterance actually means. This distinction between
two ways of departing from classical model theory underlies the distinction
between worldly vagueness and semantic indeterminacy. Vagueness and
indeterminacy are ruled out of the classical picture in two places. First, there
is no vagueness in the world. Classical models are entirely precise: they
represent the world as a crisp set of objects and properties such that for
each property and each object, the object either definitely possesses the
property, or definitely does not possess it. Second, there is no semantic
indeterminacy—no vagueness in the relationship between language and the
world. Each sentence has a unique intended model: so there is no vagueness
about what any sentence means. A view which differs from the classical
picture in the internal way (i.e. replaces classical models with non-classical
models with different internal features) opens up the possibility of worldly
vagueness, while a view which differs in the external way (i.e. denies that
only one model comes into play in describing what some utterance actually
means) opens up the possibility of semantic indeterminacy. For example, the
view of vagueness built on fuzzy set theory (see §2.2.1) countenances
worldly vagueness. It tells us that when we utter a vague predicate, we
pick out a unique property in the world—the extension of our predicate
on the unique intended interpretation of our utterance—but this property
is inherently vague, in the sense that some objects possess it, some do not
possess it, and others possess it to various intermediate degrees. This is a
6
matter of how things are out there in the world (on this view)—it is not
a matter of the relationship between language and the world. In contrast,
semantic indeterminacy enters when we have a view which denies that an
utterance has a unique intended interpretation. On this sort of view, when
we utter a vague predicate, there are several properties we might be picking
out—the extensions of this predicate on the various equally intended (or
equally not-unintended) interpretations—and there is no fact of the matter
as to which, in particular, we mean. Here we have indeterminacy in the
relationship between language and the world.⁶
Some surprising and significant results emerge from the survey of theories
of vagueness. One is that two quite different views have been conflated in
the literature under the heading ‘supervaluationism’. I call these two views
‘supervaluationism’ and ‘plurivaluationism’. This distinction is particularly
important, because ‘supervaluationism’ is widely regarded as the front
runner amongst existing theories of vagueness—but once we clearly
distinguish the two views that have been run together under this heading,
we will be in a much better position to discern the real disadvantages
of each.
⁶ In order to draw substantive conclusions about the location of vagueness—in the world itself,
or in the relationship between language and the world—from a system of model theory, we must
take a literal attitude towards model theory. That is, we must regard a model theory for (part of ) a
language as giving a literal (although not necessarily complete) description of the relationship between
that language and the world. I present reasons for taking such a literal attitude towards model theories
for vague language (in the context of a study such as this one) in §2.1.3.1.
7
camps. This is a very unsatisfactory situation. Given that the different types
of view present completely different pictures of the relationship between
vague language and the world, they cannot all be correct. We need some
way of deciding which type of theory is right. We need some criteria
for determining what the general form of the correct theory of vagueness
should be.
There is another loose end that also cries out to be tied up at this point.
We have three informal characterizations of vague predicates: they give
rise to borderline cases, their extensions have blurry boundaries, and they
generate Sorites paradoxes. These three characterizations seem to be quite
closely related to one another—yet not so closely related that they are
merely three ways of saying essentially the same thing. So what is the
relationship between them? Rather than three piecemeal characterizations
of vagueness, it would be desirable to have a proper definition of vagueness:
a crisp statement of what is of the essence of vagueness—of what vagueness
ultimately consists in—given which, we can see why vague predicates have
borderline cases, generate Sorites paradoxes, and draw blurred boundaries.
My strategy is to link these two tasks: the task of finding the correct
theory of vagueness and the task of finding an adequate definition of
vagueness. Chapter 3 is concerned with the latter task. In that chapter, I
discuss what we should expect from a definition of vagueness, critically
discuss existing definitions, and then present a definition of vagueness
according to which—roughly—a predicate F is vague just in case for
any objects a and b, if a and b are very similar in respects relevant to
the application of F, then the sentences Fa and Fb are very similar in
respect of truth. So, for example, ‘is tall’ will be vague just in case for any
persons a and b, if a and b are very similar in height, then the sentences
‘a is tall’ and ‘b is tall’ are very similar in respect of truth. The import of
this definition can be grasped by comparing it with the claim that vague
predicates are tolerant. A predicate F is tolerant with respect to φ if there
is some positive degree of change in respect of φ that things may undergo,
which is ‘‘insufficient ever to affect the justice with which F is applied
to a particular case’’ (Wright 1975, 334). So, for example, ‘is tall’ will be
tolerant with respect to height just in case for any persons a and b, if a and
b are very similar in height, then there is no difference in the applicability
of the predicate ‘is tall’ to a and b—that is, the sentences ‘a is tall’ and
‘b is tall’ are exactly the same in respect of truth. The great problem with
8
the claim that any predicate F is tolerant is that, when conjoined with the
claim that we can construct a Sorites series for the predicate F, it leads to
contradiction—in particular, to the claim that each object in the Sorites
series both is and is not F. My definition of vagueness is a weakening of
the claim that vague predicates are tolerant: if a and b are very similar in
height, then there need not be no difference in the applicability of the
predicate ‘is tall’ to a and b—that is, the sentences ‘a is tall’ and ‘b is tall’
need not be exactly the same in respect of truth—but there cannot be much
difference in the applicability of the predicate ‘is tall’ to a and b, and so
the sentences ‘a is tall’ and ‘b is tall’ must be very similar in respect of
truth. In the remainder of Chapter 3 I discuss the question of extending
this definition to cover vagueness of many-place predicates, of properties
and relations, and of objects, and I then explore the advantages of this
definition, some of the most important of which are that it captures the
intuitions which motivate the thought that vague predicates are tolerant,
without leading to contradiction, and that it yields a clear understanding
of the relationships between Sorites-susceptibility, blurred boundaries, and
borderline cases.
In Chapter 4, I turn to the task of determining what type of theory of
vagueness we need. Having clearly presented the different types of theory
in Chapter 2, and having presented a crisp definition of vagueness in
Chapter 3, my strategy is to ask which types of theory can accommodate
vague predicates—that is, predicates which satisfy the definition given in
Chapter 3. It is here, then, that the link is made between the task of finding
an adequate definition of vagueness and the task of finding the correct
theory of vagueness. When vagueness is characterized informally in terms
of borderline cases, blurred boundaries, and Sorites-susceptibility, all the
main existing types of theory of vagueness can be seen as accommodating
vagueness—simply because the informal characterizations are so loose. This
leads to the ‘too many theories’ problem: the existing theories cannot all
be right, as they conflict with one another; yet how to choose between
them, if they all accommodate vagueness perfectly well? The situation
changes, however, once we have to hand a sharp definition of the core
property underlying the various surface phenomena standardly used to
characterize vagueness. When we now ask whether each type of theory
allows for the existence of predicates possessing the feature isolated in
Chapter 3 as being of the essence of vagueness, it turns out that the answer
9
is No. Only one type of theory does: the type which countenances degrees
of truth.
The basic idea of degrees of truth is that while some sentences are true
and some are false, others possess intermediate truth values: they are truer
than the false sentences, but not as true as the true ones. So, for example,
if we line up a series of persons, ranging from one who is 7 feet in height
to one who is 4 feet in height, in increments of a fraction of an inch, and
then move along the series saying of each person in turn, ‘This person is
tall’, the idea is that our statements start out quite true, then gradually get
less and less true, until they end up quite false.
Thus, from the project of defining vagueness, an answer emerges to
the question of what the general form of the correct theory of vagueness
must be: it must be one which countenances degrees of truth.⁷ In order
to reach the overall conclusion of Chapter 4—that we need a theory of
vagueness that countenances degrees of truth—I need to do several things
in the course of the chapter. First, I show that of the types of theory
of vagueness distinguished in Chapter 2, only those which countenance
degrees of truth can accommodate predicates which satisfy the definition
of vagueness proposed and defended in Chapter 3. Second, I consider and
reject two different strategies which non-degree theorists might employ
to avoid my conclusion. The first strategy is to propose an error theory
of vagueness. Non-degree theorists might agree with the definition of
Chapter 3, but still maintain the correctness of their semantic theory. The
resulting position would be as follows: ‘‘For a predicate to be vague, it has
to satisfy the definition of Chapter 3; my semantic theory does not allow
for the existence of predicates that satisfy the definition of Chapter 3; but
my semantic theory is correct; therefore there are no vague predicates.’’ I
offer reasons for rejecting such an error theory of vagueness. The second
strategy is to reject the definition of Chapter 3, to propose in its place an
alternative definition which is compatible with a given non-degree theory,
and to argue that this alternative definition has all the advantages which
I claim for my definition in Chapter 3. I consider, and give reasons for
rejecting, several proposals along these lines, including the proposal that a
⁷ Clearly, in order for the definition of vagueness to play the role of deciding between theories of
vagueness, we must reject the widespread idea that a definition should be theory-neutral—that is, that
the definition of P must be something on which all the candidate theories of P can agree. I discuss this
issue in Ch. 3.
10
predicate F is vague just in case for any objects a and b, if a and b are very
similar in respects relevant to the application of F, then the sentences Fa
and Fb are very similar in respect of assertibility (rather than truth).
understand clearly how the different theories of vagueness differ from one
another, not just over particular matters of detail, but at a fundamental
level—and conceptual clarification and illumination of this sort is an end
in itself in philosophy. Second, just as Williamson’s and Keefe’s dialectics
require them to survey objections and replies comprehensively—for they
each argue for their favoured theory by way of some form of cost/benefit
analysis of all the available theories (see pp. 129–30 below)—my dialectic
requires me to explore carefully the inner workings of each theory. As
discussed in the Introduction, my strategy for selecting amongst theories of
vagueness is to ask of each theory whether it can allow for the existence
of vague predicates—that is, first to present a fundamental definition of
vagueness, and then to ask of each theory whether it can allow that there
are any predicates which satisfy this definition. In order to pursue this
strategy, we need not only a very clear definition of vagueness, but a very
clear understanding of the workings of each theory. Hence the need to
focus on formal details, rather than surveying all known objections and
replies to each theory. That said, I do note what I regard as the major
objections and replies to each theory, in order to give a feel for these
views to newcomers to the vagueness literature—and of course I explore
objections and replies to degree-theoretic treatments of vagueness in great
detail, in Part III.
My emphasis on formal details in the foregoing should not trigger alarm
bells in readers with little or no background in logic. The kind of reader that
I had in mind when writing the book was one with no formal background
other than a first introductory logic course of one of the standard sorts.
Indeed, the book is self-contained, and someone without even this much
logic will—with a little more effort—be able to follow the discussion.¹
(Readers with more formal background might, accordingly, find that I
am labouring certain points, and should of course skim forwards at such
places.)
Chapter 2 maps the space of possible theories of vagueness. The present
chapter deals with preliminaries. §1.2 presents the classical semantic picture
which will be located at the origin of the map. Prior to that, §1.1 introduces
some basic formal tools.
¹ I have taught several honours/postgraduate seminars based on drafts of this book: most of the
students had taken no more than a first course in logic, and some had no formal background
at all.
17
1.1 Toolkit
This section can be read through as a first introduction to the notions
discussed, suitable for readers with little or no formal background. Altern-
atively, it can be treated as a glossary, to be skipped over entirely, and then
consulted only in the event that one encounters a notion later on with
which one is unfamiliar.
1.1.1 Sets
A set is a bunch of objects; these objects are said to be members or elements
of the set. We denote the set containing a, b, and c as members as follows:
{a, b, c}. We denote the set of all things such that some condition C
holds as follows: {x : C}. For example, the set of all red things is denoted
{x : x is red}. We use the symbol ∈ (epsilon) to denote membership, as
in a ∈ {a, b, c}. To say that something is not a member of a set we use
the symbol ∈, / as in d ∈
/ {a, b, c}. Sets are individuated by their members.
That is, if ‘two’ sets have exactly the same members, they are identical;
i.e. they are in fact one and the same set. So, for example, {a, b} = {b, a}
(where = means is identical to, i.e. is one and the same object as).
Note that sets are objects. The set containing my pen, my retractable
pencil, my packet of polo mints, and my gonk is another object in its
own right, distinct from the four objects in it. That is why sets can
be members of other sets: the set just mentioned is, for example, a
member of the set of all four-membered sets. The set just mentioned is
not a visible, kickable object—unlike my retractable pencil or the other
objects in the set. For this reason, sets are often referred to as ‘abstract
objects’. But this does not make them any less objects: they are just objects
which are not visible or kickable or located in space or time. This is
the guiding idea of set theory: to treat a bunch of objects as an object in its
own right, which we can then do things with—for example, make it the
argument or value of a function (see below), put it in another set, and
so on.
1.1.2 Subsets
We say a set S is a subset of a set T (in symbols, S ⊂ T or S ⊆ T) iff every
member of S is a member of T. Note that this leaves open whether or not
18
contains everything in the background set that is not in A, not everything at all
that is not in A.) In symbols:
A = {x : x ∈
/ A}
The union of two sets A and B, denoted A ∪ B, contains everything which
is in either A or B (or both):
A ∪ B = {x : x ∈ A or x ∈ B}
The intersection of two sets A and B, denoted A ∩ B, contains everything
which is in both A and B:
A ∩ B = {x : x ∈ A and x ∈ B}
Two sets A and B are said to be disjoint if they have no members in
common, i.e. if A ∩ B = ∅. The set A\B is the set of things which are in
A but not in B:
A\B = {x : x ∈ A and x ∈
/ B} = A ∩ B
If B ⊂ A, it is common to write A − B instead of A\B.
1.1.5 Relations
A relation from a set X to a set Y is a subset of X × Y , that is, a set of
ordered pairs whose first elements are in X and whose second elements are
in Y . For example, if X is the set of females {Alice, Ethel} and Y is the set
of males {Bill, Charles, Dennis}, then if Alice, Bill, and Charles are siblings,
and Dennis and Ethel are siblings, the ‘brother of’ relation from X to Y is
{(Alice, Bill), (Alice, Charles), (Ethel, Dennis)} and the ‘sister of’ relation
from Y to X is {(Bill, Alice), (Charles, Alice), (Dennis, Ethel)}. Where X and
Y are the same set, what I have called a relation from X to Y is usually called
a binary relation on X, i.e. a subset of X 2 . If R is a binary relation on X, and
x and y are elements of X, then it is common to express the fact that x stands
in the relation R to y in any of the following ways: (x, y) ∈ R, R(x, y), Rxy,
xRy. A ternary relation on X is a subset of X 3 (i.e. a set of ordered triples
of elements of X), and in general an n-place relation on X is a subset of X n .
A binary relation R on X is said to be reflexive if it satisfies the condition
that ∀x ∈ X(xRx); transitive if ∀x, y, z ∈ X(xRy ∧ yRz → xRz); symmetric if
∀x, y ∈ X(xRy → yRx); antisymmetric if ∀x, y ∈ X(xRy ∧ yRx → x = y);
and connected if ∀x, y ∈ X(xRy ∨ yRx). If R is reflexive, symmetric, and
transitive, it is said to be an equivalence relation.
1.1.6 Functions
A function (aka map, mapping, operation) f from a set A to a set B, written
f :A→B
assigns particular objects in B to objects in A. A is called the domain of the
function, and B the codomain. The essential feature of a function is that it
never assigns more than one object in B to any given object in A. If x is a
member of A, f (x) is the object in B which the function f assigns to x. We
say that f (x) is the value of the function f for the argument x, or is the value
at x; we also say x is sent to f (x) and that f (x) is hit by x. Note that A and
B may be the same set. For example, the ‘mother of’ function assigns to
each person the person who is his or her mother; the ‘successor’ function
assigns to each natural number the number which comes directly after it in
the sequence of natural numbers.
A function f : A → B is commonly identified with the set of all ordered
pairs (x, f (x)), where x is an object in A which is sent to some object in
21
A f B
Figure 1.1. Picturing a function as a bunch of arrows.
² This usage of the terms ‘structure’ and ‘algebra’ is neither non-standard nor universal.
23
with a unary operation and two binary operations ∨ and ∧, satisfying the
following identities:
1. x ∨ y = y ∨ x and x ∧ y = y ∧ x
2. x ∨ (y ∨ z) = (x ∨ y) ∨ z and x ∧ (y ∧ z) = (x ∧ y) ∧ z
3. x ∨ (x ∧ y) = x and x ∧ (x ∨ y) = x
4. x ∨ (y ∧ z) = (x ∨ y) ∧ (x ∨ z) and x ∧ (y ∨ z) = (x ∧ y) ∨ (x ∧ z)
5. x ∨ (y ∧ y ) = x and x ∧ (y ∨ y ) = x
A unary operation on a bounded lattice which satisfies the following
conditions is an involution (or duality):
• (x ) = x
• if x ≤ y then y ≤ x
If is an involution, then the following identities—the De Morgan laws—
may or may not hold:
• (x ∨ y) = x ∧ y
• (x ∧ y) = x ∨ y
A bounded distributive lattice with an involution that satisfies the De
Morgan laws is a De Morgan algebra; if the underlying lattice is complete,
it is a complete De Morgan algebra. If a De Morgan algebra satisfies the
following condition:
• x ∧ x ≤ y ∨ y
then it is a Kleene algebra; if the underlying lattice is complete, it is a complete
Kleene algebra.
³ This is exactly the ordering we get on {0, 1} if we restrict the standard ordering of the real numbers
to the subset {0, 1}—and this is in fact one reason why we use 0 to represent falsity and 1 to represent
truth.
26
• [A ∨ B ] = [A ] [B ]
• [A ∧ B ] = [A ] [B ]
• [¬A ] = [A ]
Thus we are saying that the truth value of the wf A ∨ B —the object
assigned to this sequence of symbols on the interpretation in question—is
found by locating the object assigned to A and the object assigned to B ,
and then performing the operation on these two objects. The result of
this operation is an object in our algebra of truth values—and this object
is the truth value of A ∨ B .
Note that the foregoing information is often presented by means of
truth tables (see Table 1.1). These tables combine our earlier definitions
of the operations , , and on the set {0, 1} of truth values (these now
occur in the body of the table—for example, the definition of is in the
final column) with our recursive definition, in the previous paragraph, of
the truth values of compound propositions. I think it is more perspicuous
to keep the two definitions apart. First, it makes us less likely to confuse
the symbols of our formal language (∨, ∧, and ¬) with the operations on
our set of truth values (, , and ).⁵ Second, it will allow us to see more
clearly the differences between certain theories of vagueness which we shall
examine below—for example, between three-valued views (§2.2) and the
supervaluationist view (§2.4).
Two wfs A and B are logically equivalent (written A ≡ B ) if they have
the same truth value on every interpretation. ≡ is an equivalence relation
on F, and F/≡ is the set of equivalence classes of F under ≡. Letting |A |
be the equivalence class containing the wf A , we may set:
• |A | ∨ |B | = |A ∨ B |
• |A | ∧ |B | = |A ∧ B |
• ¬|A | = |¬A |
Note that the occurrences of ∨, ∧, and ¬ on the right of these identities
represent the sentential connectives, while the occurrences on the left
denote newly defined algebraic operations on F/≡. It is a straightforward
matter to check that these operations are well defined, and that F/≡
together with these operations is a Boolean algebra.⁶ It is not uncommon to
refer to these equivalence classes of wfs as propositions, and to this Boolean
algebra of equivalence classes as the classical propositional calculus.
What about the conditional? That is, what are we to say of sentences
of the form A → B ? There are two strategies: both yield the result that
A → B is equivalent to ¬A ∨ B and to ¬(A ∧ ¬B ), but via different
⁵ Such confusion is easier if—as is standard practice—we use the same symbols (∨, ∧, and ¬) to
represent both symbols of the language and operations on the truth values. To avoid this confusion,
I have used different symbols: ∨, ∧, and ¬ for the symbols of the language, and , , and for the
operations on the truth values. Later on—once we have the distinction between symbols of the formal
language and operations on the set of truth values clearly in mind—I will revert to the standard practice
of using the same symbols for both and letting context disambiguate.
⁶ For more details see Halmos and Givant 1998. This algebra of equivalence classes is often called
the Lindenbaum algebra.
28
⁷ The characteristic function fS of a subset S of X (also known as the indicator function) is a function
from X to the set {0, 1} of truth values, which assigns 1 to every member of X which is in S, and 0 to
every member of X which is not in S.
29
intersection, and complementation. With , , and interpreting ‘or’,
‘and’, and ‘not’, we have:
• x is in the union of f and g iff x is in f or x is in g
• x is in the intersection of f and g iff x is in f and x is in g
• x is in the complement of f iff x is not in f .
As well as unions and intersections of pairs of sets f and g, we can
also define unions and intersections of arbitrary families of sets {fi } (the
above definitions for pairs of sets are just special cases of these more general
definitions):
• {f }(x) = {fi (x)}
i
• {fi }(x) = {fi (x)}
Finally, let us consider classical predicate logic. In addition to the
propositional constants, the connectives, and the punctuation marks, we
now have individual variables x, y, z, . . . , individual constants a, b, c, . . . ,
n-ary predicate letters P n , Qn , Rn , . . . for each n ≥ 1,⁸ and quantifiers ∃ and
∀. Terms and well-formed formulae are defined as follows:
• Variables and individual constants are terms; nothing else is a term.
• Propositional constants are wfs.
• If t1 , . . . , tn are terms and P is an n-ary predicate letter, then P(t1 , . . . , tn )
is a wf.
• If A and B are wfs and x is a variable, then (¬A ), (A ∨ B ), (A ∧ B ),
((∃x)A ), and ((∀x)A ) are wfs.
• Nothing else is a wf.
Thus, as before, a wf is a sequence of symbols. Such a sequence gets a mean-
ing through being given an interpretation. An interpretation M = (M, I)
of the language consists in a nonempty set M (the domain), together with
a function I which assigns:
• a truth value [p] to every propositional constant p;
• an object I(a) in M to every individual constant a;
• an n-ary relation I(P) on M to every n-ary predicate P.
⁸ The superscript indicating the arity (number of places) of the predicate will be omitted when it is
obvious from the context.
30
⁹ Looking at the definition of a wf, we see that the only way a quantifier can get into a wf is by
being stuck on the front of a wf A , as in ((∃x)A ) or ((∀x)A ). This wf A is called the scope of the
quantifier. A variable x occurring within the scope of a quantifier (∀x) or (∃x) is said to be bound. ( The
variable x in (∀x) or (∃x) itself is also said to be bound.) A variable x which is not within the scope of
any quantifier (∀x) or (∃x) is said to be free. A closed wf is one containing no free variables.
31
The idea here is that to find out the truth value of, say, ∀xPx, we consider
the sentence Pa, and we ask what truth value it would have if everything
about our interpretation were the same, except that a denoted this object
in the domain; we note this truth value. Then we ask what truth value Pa
would have if everything about our interpretation were the same, except
that a denoted this other object in the domain; we note this truth value.
And so on, for all objects in the domain. We now have a set of truth
values—the ones we noted down along the way. To get the truth value
of ∀xPx, we apply our infimum operation to this set of truth values,
yielding another truth value. The analogy with conjunction enters in the
fact that is the generalization to arbitrary sets of truth values of our
operation on pairs of truth values.
We have now finished our tour through classical logic (propositional
and predicate) and set theory. We have seen that the Boolean algebra of
classical truth values is at the heart of the story. The set-theoretic relations
between, and operations on, sets, and the logical relations between, and
operations on, sentences, are just the images projected into the worlds
of sets and of sentences respectively by the algebraic operations on our
set of truth values. If we unplugged the classical algebra of truth values,
and plugged in a different one, and then told the rest of the ensuing
story of set theory, propositional logic, and predicate logic in the same
way as above, we would get different images appearing in the worlds
of sets and sentences—i.e. different set-theoretic relations between, and
operations on, sets, and different logical relations between, and operations
on, sentences—but once we have the recipe for telling the story of logic
and set theory given an initial algebra of truth values, how the story ends is
completely determined by our initial choice of algebra. Later on we shall
indeed vary the choice of algebra, and look at the results for set theory
and logic. The point for the moment is simply that once we have the
underlying algebra, the rest of the story comes in a natural progression.
One thing has been left out of our story so far: something that is essential
if we wish to apply the classical semantic picture to ordinary language.
Typically we want to know whether a sentence is true (simpliciter), not just
that it is true on such-and-such interpretations and false on others. What is
going on here is that when we utter a sentence, we mean something —some
particular thing—by the sentence we utter. That is, we utter it relative to
a particular interpretation. The particular interpretation that is relevant in a
32
given case has been called many things, including intended (Putnam 1983
[1977]; Merrill 1980; Lewis 1999 [1984], 1993; Weiner 2004), correct (Lepore
1983; Davidson 2005 [1986]; van Cleve 1992; Islam 1996; Field 2000), actual
(Lepore 1983), real (Hájek 1999), standard (Abbott 1997), proper (Przełȩcki
1976), and the interpretation which accords with the semantics of the language
being spoken (Field 1974, 210–11). Whatever we call the interpretation in
question,¹⁰ the final part of the classical semantic picture is the supposition
that every time you utter a sentence, you invoke a particular interpretation
of the language, and utter the sentence relative to that interpretation; that
is, it is this special or designated interpretation that gives the actual meaning
of what you say on that occasion. An utterance of a sentence is then true
simpliciter if the sentence uttered is true on the intended interpretation (as
invoked by that utterance of it), or in other words, if it is an utterance of
a sentence relative to an interpretation on which that sentence is true. In a
slogan: truth simpliciter is truth on the intended interpretation.¹¹
¹⁰ In this book, I shall regard all these terms—and some others besides, such as ‘designated
interpretation’ and ‘special interpretation’—as interchangeable. The term ‘intended interpretation’
might be felt to carry the connotation that it is the speaker’s intentions that make a certain interpretation
the one relative to which she is speaking. I do not intend any such connotation to go with my use
of this term. The question of what does determine the semantic facts will be discussed below: see in
particular §2.1.1 and §6.1.1.
¹¹ Cf. e.g. Field 1974, 212; Przełȩcki 1976, 376; Lepore 1983, 182; Menzel 1990, 358; Islam 1996;
and Weiner 2004, 165. Thanks to Amitavo Islam for helpful discussions here.
2
The Space of Possible Theories
of Vagueness
¹ One thing we assume throughout is that our language never changes: it is always our standard first-
order language. In other words our syntax never changes—we only consider changes to our semantics.
² Advocates of epistemic theories of vagueness include Cargile 1997 [1969], Campbell 1974, Sorensen
1988, ch. 6; 2001, Williamson 1997 [1992]; 1994, chs. 7–8, and Horwich 1998.
³ Note that the classical semantic picture, all by itself, does not require that the set of heaps be so
‘nicely’ organized; i.e. as far as classical semantics alone is concerned, it would be fine if the set of heaps
consisted of piles with an even number of grains—thus as we remove grains from our heap, we would
go ‘heap’, ‘non-heap’, ‘heap’, ‘non-heap’, etc. All that classical semantics, by itself, requires, is that ‘is a
heap’ have a crisp set of objects as its extension—i.e. each object is either in the set, or is out, with no
fuzziness or blurriness at the edges of the set.
35
⁴ Cf. Williamson 1994, 202: ‘‘ignorance is the real essence of the phenomenon ostensively identified
as vagueness’’, and Williamson 1996b, 327: ‘‘This is not to deny that vagueness exists; it is to assert that
its underlying nature is epistemic.’’
36
just show that the starting points are not acceptable: we already know they
cannot all be acceptable, because they lead to a conclusion which we cannot
accept. We have to explain why we thought they were acceptable: we
have to explain why we were taken in. Only then do we have a satisfying
solution to the paradox.⁵ The second part of the epistemicist’s story is that
we cannot know where the cut-off is. This is turned into an explanation of
why we were taken in by the Sorites, as follows. We cannot know where
the cut-off is, so we mistakenly think it isn’t anywhere. Our ignorance of
where the cut-off is makes us think there is no cut-off at all. That is why
we are inclined to accept the inductive premiss, even though, according to
the epistemicist, the inductive premiss is actually false.
⁵ Thanks to Delia Graff Fara for first helping me to appreciate this point. Cf. Stalnaker 1999, 74;
Fara 2000, 50; and Schiffer 2000 233 ff.
⁶ Cf. e.g. Wright 1995, 2003 and Schiffer 1999, to name just two of the critics of epistemicism who
have focused on this issue.
⁷ Cf. e.g. the opening sentence of Lewis 1992: ‘‘Surely it is our use of language that somehow
determines meaning’’, or Williamson 1994, 205: ‘‘Words mean what they do because we use them as
we do’’, or Shapiro 2006, 5: ‘‘I take it to be a truism that the competent users of a language determine
the meaning of its words and phrases.’’
37
(MU) The claim Pa is true if and only if most competent speakers would
confidently assent if presented with a in normal conditions and asked
whether it was P, and is false if and only if most competent speakers
would confidently dissent if presented with a in normal conditions and
asked whether it was P.
The biconditionals in (MU) connect meaning (i.e. extension on the
intended interpretation, which determines truth and falsity) with use
(actual and counterfactual) only in the weak sense of saying that they
co-vary. (MU) is compatible with, but does not require, the view that
usage directly determines meaning—that is, the view that a certain object
is in the extension of P because we do or would apply P to this thing.
(MU) is also compatible with the view that we apply words in certain ways
because doing so allows us to speak truthfully; in conjunction with this
view, MU merely requires that competent speakers in general have access
to the meanings of their terms—in keeping with the guiding idea that
language is a (useful) human tool.
The epistemicist must reject (MU), for the following reason. Take a
clear case of ‘is bald’ (say Yul Brynner). ‘Yul Brynner is bald’ is true, and
also, everyone would assent to it, so here meaning and use match up. Take
a clear non-case of ‘is bald’ (say Fabio). ‘Fabio is bald’ is false, and also,
everyone would dissent from it, so here again meaning and use match up.
So far so good. But what about the borderline cases? According to the
epistemicist, for any borderline case x of baldness, ‘x is bald’ is either true,
or false. Yet most competent speakers would neither assent to nor dissent
from ‘x is bald’: we shrug our shoulders, we say ‘he is and he isn’t’ or ‘he’s
sort of bald’, we say nothing either way, etc. So within the borderline cases,
there is—given the epistemicist view—a failure of match-up between
meaning and use. We can picture the situation—say for the predicate ‘is
tall’—as in Figure 2.1. What we see is that epistemicism entails a mismatch
between use (on the left) and meaning (on the right).
There are standard counterexamples to (MU). For example, consider the
predicate ‘is a lump of gold’. Most ordinary speakers would classify a lump
of fools’ gold as a lump of gold, and yet we do not think that ‘is a lump of
gold’ truly applies to a lump of fools’ gold. In other words, fools’ gold is
not gold, even though most of us would classify it as such: most of us are
just wrong. The mechanism which makes ‘is a lump of gold’ fail to have
38
use meaning
8′
assert
true
6′
hedge
4′ false
deny
2′
persons
Figure 2.1. Use and meaning.
any lumps of fools’ gold in its extension, even though we would assert of
many lumps of fools’ gold that they are lumps of gold, is as follows. We
originally applied our term ‘gold’ to some instances of a certain natural kind.
From there, the natural kind did the work of determining the extension of
our term. The term truly applies to all instances of the natural kind, even
if we would not co-classify them with the original samples, and it fails to
apply to any non-instance of the natural kind, even if we would co-classify
it with the original samples. By latching our term onto a natural kind, we
take away the rest of the job of determining its extension (i.e. beyond the
initial instances to which we applied the term) from usage, and hand it over
to the world.⁸
This sort of counterexample to (MU) will not help the epistemicist,
however—precisely because of the essential role it gives to natural kinds.
For with vague predicates, there are no natural kinds in the vicinity to
bridge the gap between use and meaning. Consider the predicate ‘is tall’.
¹¹ On a related note, Burgess 2001, 513 argues that ‘‘the critic [of epistemicism] is misrepresented
if he or she is taken to be looking primarily for a supervenience thesis. The critic . . . wants to know,
primarily, of particular borderline examples, what makes them heaps, bald men, red objects, or whatever,
when other objects, perhaps indiscriminable from them by the standard methods of testing performed
under suitable conditions, are not correctly so describable.’’ See also Keefe 2000, 80–3.
41
Burgess (2001, 519) sees here a ‘parasitic’ strategy for saying how use
determines meaning; as Weatherson (2003a, 277) puts it: ‘‘wait for the
indeterminist to offer a theory of when sentences are true, accept that
part of the indeterminist theory, and say all other sentences that express
propositions are false’’.¹² In our terms, we can say that the idea is to accept
the first biconditional in (MU) (the one connecting assent and truth), while
providing a principled reason (viz. the asymmetry of truth and falsity)
for rejecting the second biconditional (the one connecting dissent and
falsity).¹³
We can see, however, that the parasitic strategy will not work as part of
the overall epistemicist view. The epistemicist wants to explain our hedging
over Fa in borderline cases as a manifestation of our ignorance as to whether
a is or is not F, rather than in terms of its being neither true nor false that
a is F.¹⁴ As we shall see in the next section, the best account of why we
would be ignorant in borderline cases—when in fact, in such cases, either
a is F, or a is not F, according to the epistemicist—employs the idea of
knowledge requiring a margin for error, which ensures that we cannot know
of objects in the vicinity of the non-F/F boundary whether or not they are
F. As Williamson (1997 [1992], 278) puts it: ‘‘an utterance of ‘[Williamson]
is thin’ is an expression of knowledge only if I am some way from the
boundary of ‘thin’ ’’. Thus, this part of the epistemicist picture requires
that the hedging region cover the false/true boundary. But now consider
the parasitic strategy. In accordance with (MU), the indeterminist locates
false/neither-true-nor-false and neither-true-nor-false/true boundaries in
the same places as the deny/hedge and hedge/assert boundaries provided
by usage. The epistemicist who follows the parasitic strategy will then
end up with a false/true boundary which coincides with the hedge/assert
boundary—thus conflicting with the idea that the hedging region must
cover the false/true boundary.¹⁵
¹² The ‘indeterminist’ here is the theorist who thinks, in opposition to the classical semantic picture
which the epistemicist accepts, that statements about borderline cases are neither true nor false.
¹³ Thanks to Robbie Williams for helpful comments here.
¹⁴ See e.g. Williamson 1994, 3.
¹⁵ Apart from this problem with the parasitic strategy, I also have deep misgivings about the
underlying idea that there is a fundamental asymmetry between truth and falsity—but discussing this
issue would take us far afield. In addition to these worries, Burgess 2001 argues that the parasitic strategy
yields the wrong truth values in some cases; Weatherson 2003a correctly points out that Burgess is
wrong about these cases, but substitutes other cases in their place.
42
¹⁶ The view that knowledge requires a margin for error has obvious affinities with both reliabilist and
tracking theories of knowledge. For more on the surrounding epistemological issues, see Williamson
2000.
44
which in the actual situation, Bob is 1.6 m in height, and we claim he is not
tall. Community usage would have to be very different for the cut-off for
tall to be shifted below 1.6 m: that is a big change from the actual cut-off of
1.8 m. Assuming I would recognize such a difference in usage and not say
in that situation that Bob is not tall, I do know in the actual situation that
Bob is not tall: I am safely within my margin for error. But consider the
case in which in the actual situation Bob is 1.8 m tall. We might suppose
that for the cut-off to be shifted just above 1.8 m, usage would not have
to be very different: just a few people here or there would have to have
different dispositions regarding the use of ‘is tall’ to make such a tiny shift
in its extension.¹⁷ Plausibly, were such a change in usage to take place, I
would be none the wiser: so I do not know in the actual situation that Bob
is tall, for I have no margin for error.
If this is correct, then we can see why we cannot know where the cut-off
for tall is. In order to know that the cut-off between tall and not-tall is at
1.8 m, I would have to know both that a person 1.8 m in height is tall, and
that a person arbitrarily less than 1.8 m in height is not tall. But we have
just seen that I cannot know such things, because when we are so close to
the cut-off that the amount of change in usage required to push the cut-off
over the point we are considering is small enough for us not to be able to
notice it, there is no margin for error. Thus I cannot know where the cut-off
between tall and not-tall is—and similarly for other vague predicates.
¹⁷ In fact, it is hard to see how Williamson could be entitled to claims of this sort, given his view
(discussed in the previous section) that ‘‘Meaning may supervene on use in an unsurveyably chaotic
way’’ (1994, 209).
45
Classical models are entirely precise: they represent the world as a crisp
set of objects and properties such that for each property and each object,
the object either definitely possesses the property, or definitely does not
possess it. Second, there is no semantic indeterminacy—no vagueness in the
relationship between language and the world. Each discourse has a unique
intended model: so there is no vagueness or indeterminacy concerning
what any name, predicate, or sentence means. For the epistemicist, then,
vagueness is neither a metaphysical nor a semantic phenomenon. It is an
epistemic phenomenon. Although the world itself is completely precise,
and all items of our language have a unique, precise meaning, some of
these meanings are unknowable. For example, ‘is tall’ picks out a particular
crisp set of objects in the world, but we cannot know the full details of
the membership of this set—and it is in this necessary ignorance that the
vagueness of ‘is tall’ is located.
possible worlds and merely possible objects, but not take this model theory
literally in my sense: when his model theory tells him that ‘Possibly P’ is true
iff there is a possible world in which P is true, he does not regard this as a lit-
eral description of what goes on when I say ‘Possibly P’; that is, he does not
think that I stand to a bunch of possible worlds, which between them do or
do not make my statement true, in the same way that the teacher who says
‘Someone threw that apple core’ stands to the collection of students in the
room, who between them do or do not make that statement true. Rather,
he might treat his model theory in an instrumental fashion, regarding it as
a calculus for keeping track of our modal commitments.¹⁸ This is all quite
right, but in the present context—i.e. in the various sections of this chapter
entitled ‘Worldly Vagueness?’—a literal attitude towards model theory is
appropriate. This is because a presupposition of the question being asked in
these sections is semantic realism, and where one is a semantic realist about an
area of discourse, one should be a model-theoretic literalist. Let me explain.
Semantic realism, in the sense I mean it here, is the view that there
are genuine semantic properties and relations: in particular, relations of
reference holding between sub-sentential or sub-propositional expressions
(names, predicates, relation symbols) and parts of the world (objects, prop-
erties or sets of objects, relations or sets of n-tuples of objects), and properties
of truth possessed by (some) sentences or propositions.¹⁹ Note that while
semantic realism implies that there are real relations of reference, it does
¹⁸ This non-literal attitude towards modal model theory is to be contrasted with the ersatzist
perspective, which involves taking some form of Kripke-style modal model theory employing possible
worlds literally, but denying that the possible worlds and mere possibilia which figure in the model
theory are to be regarded as concrete things on a par with the actual world and its inhabitants. In relation
to the distinction between literal and non-literal attitudes towards model theory, cf. the discussions in
the literature on modal logic of the distinction between pure and applied (aka depraved) semantics.
The original distinction is due to Plantinga 1974, 126–8. For some recent discussion see Gregory 2005;
Zimmerman 2005, 414 ff.; and Divers 2006. My first introduction to the idea of seeing (classical) model
theory as providing an account of the relationship between natural language and the world was in
helpful discussions with Amitavo Islam.
¹⁹ Within the class of semantic realists—characterized by adherence to the view that both reference
and truth are real—we might furthermore distinguish those who hold that reference is primary, and
those who hold that truth is primary. On the former view (which I call ‘reference-first semantic
realism’), the truth status (e.g. true, false, no truth value, an intermediate truth value) of sentences is
explicable in terms of the referents of their parts. For example, in the sentence ‘Maisy is hungry’, the
name ‘Maisy’ refers to a certain dog, and the predicate ‘is hungry’ picks out a certain property (or set
of things), and it is because Maisy in fact has this property (is a member of this set of things) that the
sentence is true. On the latter view (which I call ‘truth-first semantic realism’), what fixes or determines
reference is the truth conditions of certain sentences. For example, one might hold that the reference of
some term is whatever it has to be in order to make all the sentences in some overarching theory true.
47
²⁰ Indeed, this follows from the fact (n. 19) that semantic realism is compatible with the view that
the truth conditions of sentences determine all other semantic facts, together with the now well-known
fact (emphasized by Quine, Wallace, and Putnam, amongst others—see e.g. Wallace 1977 and Putnam
1981, 33), that in general the truth conditions of whole sentences do not determine unique references
for their parts.
²¹ Abbott 1997, 128–30 has a useful categorization of views of reference, which distinguishes
two other views besides semantic realism and antirepresentationalism (which Abbott calls ‘bold
antirepresentationalism’).
²² Referential antirealism of this sort must be distinguished from truth-first semantic realism. The
latter view agrees with Davidson on the primacy in semantics of the truth conditions of whole sentences,
but disagrees that there are no real relations of reference between sub-sentential expressions and (parts
of ) the world.
48
²³ Of course, as I have already noted, there may well be other contexts in which we wish to employ
some system of model theory for different reasons and in which a literal attitude to that model theory
would not be appropriate. This also means that when I say, for example, that many-valued theories of
vagueness locate vagueness in the world (§2.2.2), this cannot be read as saying that previous advocates
of many-valued semantics think that vagueness is a worldly, not a semantic, phenomenon: for while an
author may have advocated a many-valued model theory for vague language, she may not have taken
her model theory literally in my sense.
50
²⁴ In §1.2, I used different symbols for the algebraic operations on the set of truth values and for
the connectives in our language, in order to make this distinction very clear. As foreshadowed in ch. 1
n. 5, I now revert to the standard practice of using the same symbols for both, and letting context
disambiguate.
51
are neither true nor false, but have a third truth value.²⁵ One particularly
nice aspect of this approach is that it allows us to maintain an attractive
parallelism between meaning and use, of the sort that the epistemicist
was forced to reject. We can now say that a vague sentence Pa is true if
most competent speakers would confidently classify a as being P, is false if
most competent speakers would confidently deny that a is P, and is neither
true nor false if most competent speakers would hedge over whether a is
P. The epistemicist had only two semantic statuses for sentences—true
and false—whereas ordinary speakers have at least three pragmatic stances
which they can adopt towards a sentence—assertion, denial, or hedging.
The three-valued approach posits a new semantic status, thereby allowing
a parallelism between meaning and use to be maintained.²⁶
In developing any three-valued view, our first step is to specify an algebra
({0, 1, ∗}, ∨, ∧, ) which has three truth values, 1, 0, and the new truth value
∗, and three operations, ∨, ∧, and . Consider Table 2.1. Every different
way of filling in the blank cells with a 1, 0, or ∗ yields a different algebra
of truth values. (A cell marked with the ‘ditto’ symbol must be filled
with whatever is in the cell above it.) There are twenty-one empty cells,
and three possible fillings for each; thus we have 321 = 10, 460, 353, 203
possibilities. Thus, our family of three-valued logics is large!
Some members of the family are more interesting than others. For a start,
it is natural to be more interested in those operations ∨, ∧, and which,
when fed only 1’s and 0’s, output what the classical operations would
output. This fixes every row which does not have a ∗ in it, reducing the
number of empty cells to eleven (and thus leaving 311 = 177, 147 possible
algebras remaining). It is also natural to want the truth values of A ∨ B and
A ∧ B to be the same as those of B ∨ A and B ∧ A respectively—thus
we want our binary operations ∨ and ∧ to be symmetric. Introducing
²⁵ Other common motivations for introducing three-valued logics include: Future contingents.
Consider a statement about the future, such as ‘Bob will eat pizza for dinner next Tuesday’. You might
think that as of right now, this statement is neither true nor false: it’s up to Bob, and he has not decided
yet. Semantic paradoxes. Consider the sentence: ‘This sentence is false.’ Suppose it is true; then what it
says is indeed the case; but what it says is that it is false. So if it’s true, it’s false, so it’s not true. So what
if it is false? Then what it says is not the case. What it says is that it is false; so it is not false, i.e. it is true.
So if it is false, it is true, so it is not false. So it seems that the sentence can be neither true nor false. It
might reasonably be thought that these motivations lead most naturally not to the idea of a third truth
value, but to the idea that some sentences lack a truth value altogether. We shall come to this idea in
§2.3.
²⁶ For further discussion of this issue see p. 58 below.
52
1 1
1 ∗
1 0
∗ 1
∗ ∗
∗ 0
0 1
0 ∗
0 0
the convention that a cell marked with the symbol 2 must be filled with
whatever is in the cell above the cell above it, this leaves us with seven
empty cells (see Table 2.2). Finally, we are naturally more interested in
those algebras where ∨ and ∧ are duals (with respect to ).²⁷ Thus, once we
1 1 1 1 0
1 ∗
1 0 1 0
2 2
∗ 1
∗ ∗
∗ 0
0 1 1 0 1
2 2
0 ∗
0 0 0 0
have fixed one of these operations, the other is fixed too. That removes
three more cells. The number of algebras left is still 34 = 81. Of these,
I shall mention only two!
The first is arrived at via the idea that any compound proposition with
a component that has the value ∗ should itself have the value ∗. Given
that, once we have fixed the algebra of truth values, we are simply going
to run the rest of the classical story without further changes—thus, for
example, the truth value of A ∨ B will be that truth value to which the
operation ∨ in our algebra of truth values maps the pair (the truth value
of A , the truth value of B )—this means that our operations must be as in
Table 2.3 (where every blank cell in Table 2.2 has been filled with a ∗).²⁸
A second rationale for filling in the blank cells in Table 2.2 is as follows.
Suppose the ∗ were a 1, and calculate the value of the corresponding
classical operation, and suppose the ∗ were a 0, and calculate the value of
the corresponding classical operation; if you get 1 both times, the value of
the new operation that you are defining is 1; if you get 0 both times, the
value of the new operation is 0; if you get 1 once and 0 once, then the value
of the new operation is ∗. (Where you are trying to determine what the
1 0 1 0
∗ 1 ∗ ∗ ∗
∗ ∗ ∗ ∗
∗ 0 ∗ ∗
0 1 1 0 1
0 ∗ ∗ ∗
0 0 0 0
²⁸ This gives us something analogous to Kleene’s 1952 weak truth tables. It does not give us Kleene’s
weak truth tables themselves, because what I have just given is not a truth table (recall the discussion
on p. 26).
54
new operation should assign to a pair of ∗’s, you need to calculate all four
possibilities where each ∗ is replaced by a 1 or a 0.) Following this rationale
through, we get Table 2.4.²⁹ The rationale used to define the latest set
of operations may have sounded like supervaluationism, but the outcome
is in fact very different. The three-valued view just presented—along
with every other view constructed from the classical view by replacing
the algebra of classical truth values with an alternative algebra—is truth-
functional, which is to say that the truth value of a compound sentence is
determined solely by the truth values of its components. Truth-functionality
is ensured by the part of the picture which says that [A ∨ B ] = [A ] ∨ [B ],
[A ∧ B ] = [A ] ∧ [B ] and [¬A ] = [A ] (see p. 26 above). On the other
hand, as we shall see when we discuss it in §2.4, the supervaluationist view
is not truth-functional.
A third rationale for filling in the blank cells in Table 2.2 is as follows.
Recall that at the beginning of our discussion of the classical picture, we
considered the appropriate ordering of our two classical truth values 1 and 0.
Let us do this again, with our three truth values 1, 0, and ∗. We may ask,
under what assignments of truth values to sentences S and T is S at least as
true as T? Obviously, if S and T have the same truth value, then S ≥ T
1 0 1 0
∗ 1 1 ∗ ∗
∗ ∗ ∗ ∗
∗ 0 ∗ 0
0 1 1 0 1
0 ∗ ∗ 0
0 0 0 0
²⁹ This is Łukasiewicz’s three-valued logic (minus his treatment of the conditional, which is the
three-valued analogue of the fuzzy Łukasiewicz conditional, to be introduced in §2.2.1.1). It is also
analogous to Kleene’s 1952 strong truth tables (cf. n. 28).
55
(letting ≥ mean ‘at least as true as’), and if S is true and T is false, then S is
truer than T (i.e. S ≥ T and T S). But what if S has the value ∗ and T
does not? One natural thought is that if T is false, then S is truer than T,
and that if T is true, then T is truer than S. This yields the following lattice
of truth values:
1
|
∗
|
0
Setting x ∨ y equal to the supremum of x and y relative to this ordering,
x ∧ y equal to the infimum of x and y relative to this ordering, and setting
0 = 1, ∗ = ∗, and 1 = 0, yields an algebra of truth values. As is easily
verified, it is in fact the same as our previous algebra.
We wanted to replace the algebra of classical truth values with a different
structure. We have seen that we have many choices here. Making this
choice is our only substantive step. After making it, we then run the rest of
the classical story just as before. Thus, when it comes to propositional logic,
the truth value [A ∧ B ] of the proposition A ∧ B will be [A ] ∧ [B ],
where [A ] is the truth value of A , [B ] is the truth value of B , and ∧ is
the operation in our new algebra of truth values that replaces the classical
operation, symbolized by the same symbol, in our original Boolean algebra
of truth values.³⁰ When it comes to set theory, we have a ready-made
³⁰ Of course, not all the details of the resulting story of propositional logic will be the same as in
the classical story: what we are holding fixed is the method of generating the story, given an algebra
of truth values as input; when we vary that algebra, the details will, in general, change. The changes
are, however—and this is the point—perfectly predictable. Thus, e.g., the classical algebra of truth
values is a Boolean algebra, and so all the laws of Boolean algebras reappear, in different guise, as
laws of classical propositional logic. For example, mirroring law 5 of Boolean algebras on p. 24, two
formulas A and A ∨ (B ∧ ¬B ) of classical propositional logic always have the same truth value. Our
second algebra of three truth values is not a Boolean algebra—for example, in relation to law 5 just
considered, 0 ∨ (∗ ∧ ∗ ) = ∗ = 0—and so, correspondingly, two formulas A and A ∨ (B ∧ ¬B ) will
have different truth values when A has the value 0 and B has the value ∗. Our first algebra of truth
values (unlike our second) is not even a lattice—for example, in relation to law 4 of lattices on p. 23,
1 ∨ (1 ∧ ∗) = ∗ = 1—and so, correspondingly, two formulas A and A ∨ (A ∧ B ) will have different
truth values when A has the value 1 and B has the value ∗—unlike in classical propositional logic
(and the logic resulting from our second algebra of truth values), where such formulae always have the
same truth value. However, the point is that the laws of each logic do correspond perfectly to the laws
of its own underlying algebra of truth values—thanks to the uniform method of generating the logic
from the algebra.
56
paradox is that these things that we want to say cannot all be true (at least in
the classical picture). One obvious strategy for a three-valued response to
this situation—illustrated with respect to the second of the two particular
three-valued views introduced above—is as follows. Where Bob is a clear
case of tallness, ‘Bob is tall’ has the value 1; where Bob is a borderline case
of tallness, ‘Bob is tall’ has the value ∗; and where Bob is a clear countercase
of tallness, ‘Bob is tall’ has the value 0. Going from the 2000 end to the 0
end, our Sorites series begins with people who are clear cases of tallness and
ends with people who are clear countercases of tallness, and in between
contains people who are borderline cases of tallness. So now consider our
Sorites conditionals, after recalling the part of the picture which says that
A → B is an abbreviation of ¬A ∨ B , and then noting that 1 ∨ 1 = 1,
1 ∨ ∗ = ∗, ∗ ∨ ∗ = ∗, ∗ ∨ 0 = ∗, and 0 ∨ 0 = 1.³¹ The upshot is that
some of the Sorites conditionals are not true. So that’s the error in the
paradoxical reasoning. Why, then, do we get taken in by the paradox?
Because none of the conditionals is false, either. The only way for ‘if P
then Q’ to have the value 0 is for P to have the value 1 and Q the value 0.
This does not happen in our series: between the persons who are tall (i.e.
are mapped to 1 by the set of tall things) and those who are not tall (i.e. are
mapped to 0 by the set of tall things) are the borderline cases—that is, the
persons who are neither tall nor not tall (i.e. are mapped to ∗ by the set of
tall things).³²
There are two main objections to three-valued approaches to vagueness.
One of them focuses on the fact that the three-valued approach is truth-
functional. I shall postpone discussion of it until p. 85.³³ The other
objection is that the three-valued approach falls foul of ‘the problem of
higher-order vagueness’, by imposing sharp cut-offs between the objects to
which a predicate does not apply and its borderline cases, and between the
borderline cases and the objects to which the predicate applies. In fact we
³⁴ Note that I do not think that there is no problem for the three-valued view stemming from its
positing of perfectly sharp divisions between the men mapped to 0 by the characteristic function of
the extension of ‘is tall’ and the men mapped to ∗, and between the men mapped to ∗ and the men
mapped to 1. In fact I think there is a devastating problem here: the jolt problem, to be discussed in
Ch. 4. My point at present is that there is not a big problem concerning how our usage—as a speech
community—could determine the position of these sharp divisions.
³⁵ The finitely many-valued logics of Łukasiewicz take this form; see Łukasiewicz and Tarski 1970
[1930], 141 or Malinowski 1993, 36.
60
³⁶ The study of infinite-valued logics begins with Łukasiewicz; see e.g. Łukasiewicz and Tarski 1970
[1930]. Fuzzy set theory was born in Zadeh 1965; a related earlier idea is Post’s notion of an n-valued
set (Malinowski 1993, 47, 98). The history of both branches of the fuzzy view (i.e. logic and set theory)
is rich; see Hájek 1998, ch. 10, and Novák et al. 1999, ch. 8, for overviews. For technical introductions
to fuzzy logic and set theory, see e.g. Klir and Yuan 1995; Hájek 1998; Novák et al. 1999; and Nguyen
and Walker 2000. The chief sources for the fuzzy view of vagueness are Goguen 1968–9; Lakoff 1973;
and Machina 1976; Black 1997 [1937] is a relevant earlier work.
61
⁴⁰ So are the lattices of finitely many truth values introduced at the beginning of §2.2.1. However,
the set of all rational numbers between 0 and 1 inclusive, under the standard ordering, yields a
lattice which is not complete (e.g., the set of all such rationals which are less/greater than √12 has no
supremum/infimum). That is why we do not consider taking the set of all rational numbers between 0
and 1 inclusive as our set of truth values, but instead move straight from finitely many truth values to
continuum-many.
65
⁴¹ One might think that as in the case of the definitions of union and intersection for arbitrary
families
offuzzy sets, in order for these truth definitions for quantified sentences to work, we need
S and S need to be well-defined for an arbitrary set S of truth values—in other words, the lattice
of truth values must be complete. This time, however, completeness is a sufficient
but not necessary
condition for our definitions to work. In order for the definitions to work, S and S need not
be defined for every set S of truth values: they need only be defined for every set S of truth values
which we might arrive at by taking a closed wf ∀xA or ∃xA , stripping off the quantifier, replacing
any free occurrences of x by a new constant a, and then putting one truth value, viz. [Ax a]Mao , into
S for eachobject o in the domain. Thus the number of sets S of truth values for which we need
S and S to be defined will not be greater than the number of wfs of our language. (NB: We are
talking about the number of sets of truth values, not the number of truth values in any of these sets.)
If (as is standardly the case, and is the case in my presentation here) there are countably many wfs,
then (given that we have uncountably
many fuzzy truth values) the number of sets S of truth values
for which we need S and S to be definedwill be less than the cardinality of the power set of
the set of truth values, and thus we do not need S and S to be well-defined for arbitrary sets S of
truth values. Thus, lattice completeness gives us more than we actually need. Getting by on just what
we do need, without the completeness requirement, requires some effort (see Restall 1994, §4.1, who
draws on Brady 1988). I shall not go into the details, because we need our truth values not just for
logic, but also for set theory—and there, as we have seen, we do need the lattice of truth values to be
complete.
67
Earlier (p. 61) I mentioned that properties 2 and 4 are desirable; as a matter
of fact, if a function c : [0, 1] → [0, 1] possesses these properties, then it
possesses the other two as well (Klir and Yuan 1995, 52). In the fuzzy
literature, conditions 1 and 2 are taken as minimal, and negations which
satisfy only these two conditions are studied; notable examples are negations
of the threshold type, of which there is one for each value of the parameter
t ∈ [0, 1):
1 for x ≤ t
c(x) =
0 for x > t
As for negations which satisfy all four conditions, the standard fuzzy negation
is not the only example. Other notable examples are Sugeno negations, of
which there is one for each value of the parameter λ ∈ (−1, ∞):
1−x
cλ (x) =
1 + λx
and Yager negations, of which there is one for each value of the parameter
w ∈ (0, ∞):
cw (x) = (1 − xw )1/w
⁴² See e.g. Malinowski 1993, 89, and Hájek 1998, 31. For some more details on fuzzy negations, see
Klir and Yuan 1995, 51–61, and Nguyen and Walker 2000, 100–8.
68
Fuzzy conjunctions and disjunctions are functions i and u from [0, 1] × [0, 1]
to [0, 1]. Some basic properties which we would like them to have are:
1. i(x, 1) = x and u(x, 0) = x
2. i(x, y) = i(y, x) and u(x, y) = u(y, x)
3. i(x, i(y, z)) = i(i(x, y), z) and u(x, u(y, z)) = u(u(x, y), z)
4. If y ≤ z, then i(x, y) ≤ i(x, z) and u(x, y) ≤ u(x, z)
It turns out that functions with these properties had already been studied
in connection with the theory of statistical metric spaces, where they were
referred to as t-norms (or triangular norms) and t-conorms respectively.⁴³ There
are further conditions which t-norms and t-conorms may or may not satisfy.
One condition which was mentioned earlier is idempotence: i(x, x) = x
and u(x, x) = x. It turns out that the standard fuzzy conjunction is the only
idempotent t-norm, and the standard fuzzy disjunction is the only idem-
potent t-conorm (Klir and Yuan 1995, 63, 77). Nevertheless, other t-norms
and t-conorms are studied in the literature; notable examples are Yager
t-norms, of which there is one for each value of the parameter w ∈ (0, ∞):
iw (x, y) = 1 − min(1, [(1 − x)w + (1 − y)w ]1/w )
and Yager t-conorms, of which there is one for each value of the parameter
w ∈ (0, ∞):
uw (x, y) = min(1, (xw + yw )1/w )
Other notable conjunctions are the Łukasiewicz conjunction:⁴⁴
i(x, y) = max(0, x + y − 1)
and the product conjunction:⁴⁵
i(x, y) = xy
Fuzzy conditionals are functions m from [0, 1] × [0, 1] to [0, 1]. Notable
examples include the Łukasiewicz conditional:
1 if x ≤ y
m(x, y) =
1 − x + y otherwise
⁴⁶ See e.g. Rescher 1969, 44, and Malinowski 1993, 89, 100.
⁴⁷ This trio of operations is not unique in this respect; see e.g. Klir and Yuan 1995, 83–7.
⁴⁸ See Hájek 1998, 27–32.
70
Given a set of two truth values, there are only 22 = 4 unary operations
2
on that set, and only 2(2 ) = 16 binary operations. On the set of fuzzy truth
values [0, 1], by contrast, there are uncountably many n-place operations
for each n ≥ 0.⁴⁹ This structural richness makes fuzzy logic an interesting
object of study from a technical point of view. From the point of view
of studying vagueness, on the other hand, while I would not want to
rule out the possibility that further down the track, distinctions amongst
different connectives might shed light on important vagueness-related
issues, nevertheless I cannot see, at the present stage of enquiry, that there
is anything to be gained by moving beyond the basic fuzzy operations
∨, ∧, and . Therefore, with the exception of the discussion of the
conditional in §5.5.1, I shall not discuss other operations further in this
book.
vague is a matter to which we shall come. The point for now is that the
divergence from the classical picture comes on the world side, not in the
relationship between language and the world.
Forbes (1983, 245) writes: ‘‘Note that on [the fuzzy] approach, vagueness
resides entirely in concepts. The objects in [the domain of the fuzzy mem-
bership function] are perfectly determinate and the fuzzy sets themselves
also have exact identity conditions: two such sets are the same iff the
same things are members of each to the same degree.’’ Forbes is clearly
assuming that worldly vagueness requires vague objects and/or vague iden-
tity. I disagree. If the fuzzy view is correct, there exist vague properties
and relations alongside the precise ones (indeed, the precise ones are those
special cases of the vague ones that map all n-tuples to 0 or 1)—and this is
one perfectly genuine sort of worldly or metaphysical vagueness. In order
to describe the world, it is not enough just to list the objects it contains: one
must also describe their properties and the relations they bear to one another.
Metaphysical vagueness is often taken to mean vagueness in objects—but
if there is vagueness in properties or relations, this will also be a kind of
metaphysical or worldly vagueness. On the fuzzy semantic picture, there
are properties and relations in the world that are vague. Contra Forbes, this
is not vagueness in concepts.
assigns x the value 0, this means x is not in S. We now allow that S might
assign x no value at all—which we interpret as meaning that x is neither in
S nor not in S. Thus, we now have truth value gaps for simple sentences,
and we have gappy sets.⁵⁰ Given these changes, we now want to run the
rest of the classical story with no further changes—or, where this is not
possible, with only the minimal changes required to accommodate the two
basic changes which we have just made.
One immediate upshot is that we get truth gaps for atomic sentences,
as well as for simple sentences. For recall the classical story, in which the
truth definition for an atomic wf P(a1 , . . . , an ), which consists of an n-ary
predicate P followed by n individual constants a1 , . . . , an , is as follows:
[P(a1 , . . . , an )] = I(P)(I(a1 ), . . . , I(an ))
Recall the idea here: each individual constant a is assigned an object I(a)
in the domain, and P is assigned an n-ary relation on the domain, that is
a function I(P) from the set of all n-tuples of members of M to our set
of truth values {0, 1}. Whatever value this function assigns to the n-tuple
(I(a1 ), . . . , I(an )), this value is the truth value of the sentence P(a1 , . . . , an ).
This function may now be partial. So suppose it assigns nothing to the n-tuple
(I(a1 ), . . . , I(an )); then the atomic sentence P(a1 , . . . , an ) will have no truth
value. More concretely, ‘Bob is bald’ is true if Bob is in the extension
of ‘is bald’ (i.e. the characteristic function of the extension of ‘is bald’
assigns 1 to the referent of ‘Bob’), is false if Bob is not in the extension
of ‘is bald’ (i.e. the characteristic function of the extension of ‘is bald’
assigns 0 to the referent of ‘Bob’), and is neither true nor false if Bob
is neither in nor out of the extension of ‘is bald’ (i.e. the characteristic
function of the extension of ‘is bald’ assigns nothing to the referent of
‘Bob’). This idea has a natural application in the case of vagueness—for
it is a fairly natural thought about vague sentences such as ‘Bob is bald’,
where Bob is a borderline case for baldness, that they are neither true nor
false. As in the case of the three-valued logicians discussed in §2.2, this
allows us to maintain an attractive parallelism between meaning and use,
of the sort that the epistemicist was forced to reject. The difference this
⁵⁰ In principle, we could separate these aspects: we could, for example, consider a view according to
which all simple sentences have a truth value, but some sets are partial. In practice, however, the two
aspects go together very naturally, and we shall not explore the option of separating them.
73
⁵¹ In the literature, many writers who take this sort of approach implement it by saying that a
predicate, instead of being assigned an extension (as on the classical view), is assigned a pair of an
extension and an antiextension, which need not exhaust the domain (Kripke 1975; Soames 1999). I say
instead that a predicate is assigned a set as its extension, both on the new view and on the classical view,
but the difference is that sets may now be partial, that is, their characteristic functions may be partial,
rather than total (as on the classical view). Given the way that I presented the classical picture, my
formulation is smoother. The two formulations are, however, equivalent. The set of things assigned
1 by the characteristic function of my extension is their extension; the set of things assigned 0 by the
characteristic function of my extension is their antiextension; and the set of things assigned nothing by
the characteristic function of my extension is the set of things which are in neither their extension nor
their antiextension.
74
1
0
0 1 1 0 1
0
0 0 0 0
75
1 1
0 0
0 1 1 0 1
0 0
0 0 0 0
76
⁵³ Cf. n. 3 above.
81
• Suppose that Bill is 16, Ben is 18, and Bob is 20. If we precisify
‘juvenile’ and ‘adult’, it must turn out that each of Bill, Ben, and Bob
falls in exactly one of these categories, and it must not be the case that
Bill and Bob fall in one category, while Ben falls in the other.
The first constraint concerns how one object should be classified relative to
several predicates; the second constraint concerns how several objects should
be classified relative to one predicate; and the third constraint concerns how
several objects should be classified relative to several predicates. This is just a
tiny sample of the constraints on admissible precisification, not a full list. But
it should be enough to give the idea. Now suppose we have a partial inter-
pretation which corresponds to our actual use of a vague language. A classical
extension of this partial interpretation will (to repeat) be admissible just in
case it corresponds to a legitimate precisification of that vague language.
Note that we have defined the notion of an admissible classical exten-
sion of a given partial interpretation. Do not call the admissible extension
an admissible interpretation of the given vague language—this will lead to
confusion later. Our delineation of a certain class of classical interpretations
as admissible extensions is always relative to a given partial interpretation,
which is assumed to correspond to a body of usage of vague language. That
is, there is a unique intended partial interpretation of some vague language,
and then it has many admissible classical extensions. These correspond to
legitimate precisifications of the vague language whose intended interpreta-
tion was given by the original partial interpretation—they are not themselves
intended interpretations of the vague language. (They could not be, because
they remove all its vagueness! But more on this below—see especially §4.4.)
Returning to supervaluationism as applied to vagueness, the strategy for
assigning truth values to compound sentences in our partial interpretation
Mp is now as follows. A compound sentence gets assigned the value 1
on Mp if it comes out as having the value 1 on every admissible classical
extension of Mp ; it gets assigned the value 0 on Mp if it comes out as
having the value 0 on every admissible classical extension of Mp ; and it
gets assigned no truth value on Mp if it comes out as having the value
1 on some admissible classical extensions of Mp and 0 on others. The
intuitive gloss on this is that a vague sentence is true simpliciter if it would
be true no matter how its vagueness were removed; it is false simpliciter
if it would be false no matter how its vagueness were removed; and it is
82
neither true simpliciter nor false simpliciter if there are legitimate ways of
removing its vagueness that would render it true, and others that would
render it false.
One point of interest about the supervaluationist approach is that,
although its semantics is non-classical, it yields classical logic. That is, any
classical tautology is a supervaluationist tautology, and vice versa, and
any inference which is valid in classical logic is valid according to the
supervaluationist, and vice versa. Let’s be more precise about this. I am
assuming here—as throughout this chapter—that we are dealing with our
standard first-order language; if we enrich the language in certain ways,
then the following result does not hold.⁵⁴ Where W is the set of wfs of
our language, a consequence relation Cx is a subset of PW × W . Where is
a set of wfs and A is a wf, we say |=x A iff (, A ) ∈ Cx . Any A for
which (∅, A ) ∈ Cx is called an x tautology. Now, we define the classical
consequence relation Cclass thus: (, A ) ∈ Cclass iff there is no classical
interpretation on which all the members of are true and A is false. We
define the supervaluationist consequence relation Csval thus: (, A ) ∈ Csval
iff there is no partial interpretation on which all the members of are true
and A is not true (i.e. is false, or is neither true nor false).⁵⁵ It then turns
out that Cclass = Csval .⁵⁶
⁵⁴ As Hyde 1997, 652 emphasizes, the result also does not hold—even for our standard first-
order language—if we consider the multiple-conclusion consequence relation, in place of the usual
multiple-premiss-single-conclusion consequence relation considered in what follows. See §2.4.2 below.
⁵⁵ It has been said that supervaluationists have a choice as to how define validity (cf. Dummett 1997
[1975], 108, and Fine 1997 [1975], 137, and for discussion see e.g. Williamson 1994, 147–8): they can
say what we have just said (‘global validity’) or they can say that an argument is valid just in case there
is no classical extension of any partial interpretation on which all the members of are true and A
is not true (‘local validity’). It should be obvious that the latter definition in fact has no plausibility at
all—i.e. our definition is obviously the correct one. However, we shall see in §2.5 that there is a view
quite different from supervaluationism, which I call plurivaluationism—and for this view, the obvious
definition of validity corresponds to the latter idea. Plurivaluationism and supervaluationism have been
conflated in the literature, leading to the illusion that there is one view of vagueness which has two
natural options concerning how to define validity.
⁵⁶ Proof. (i) Suppose (, A ) ∈ Cclass . Then (A) there is no classical interpretation on which all the
members of are true and A is false. Now suppose we have an arbitrary partial interpretation on
which every member of is true. This means that on every extension of our partial interpretation,
every member of is true. Then by (A), on every extension of our partial interpretation, A is
true. Hence A is true on our partial interpretation. Thus on any partial interpretation on which all
the members of are true, A is true; i.e. (, A ) ∈ Csval . (ii) Suppose (, A ) ∈ / Cclass . Then there
is a classical interpretation on which all the members of are true and A is false. But note that a
classical interpretation is a partial interpretation: it is a special case of the latter, where the only classical
interpretation which extends it is itself. So, there is a partial interpretation on which all the members of
are true and A is false; hence (, A ) ∈ / Csval .
83
How will the supervaluationist view handle the Sorites paradox? Suppose
we remove one grain at a time from a 10,000-grain pile of sand, until we
have one grain left. Call the pile with n grains ‘pile n’. Our ordinary, vague
usage of the term ‘heap’ corresponds to a partial interpretation on which, let
us say, piles 100 through 10,000 are assigned the value 1 by the characteristic
function of the extension of ‘is a heap’, piles 1 through 10 are assigned 0, and
piles 11 through 99 are assigned no value. (I am talking here about the usage
of the whole community of speakers—recall the discussion on pp. 58–9.)
Now, for each 11 ≤ n ≤ 100, there is an admissible extension—which we
shall call extension N —of this interpretation on which piles 1 through
n − 1 are assigned 0, and piles n through 10,000 are assigned 1. (There
are also extensions on which, for example, piles 1 through 10, 35, 39,
and 42 are assigned 0, and the rest are assigned 1, but these are not
admissible.) Let the sentence ‘the cut-off is at n’ mean that piles 1 through
n − 1 are non-heaps, and piles n through 10,000 are heaps. Consider the
inductive premiss of the Sorites argument associated with this setup: ‘For
every 2 ≤ n ≤ 10,000, if n is a heap, then so is n − 1.’ This sentence is
false on each of our admissible extensions: on interpretation N, where
the cut-off is at n, the corresponding instance of this universal claim is
false, and hence the universal claim itself is false. Thus, according to the
supervaluationist approach, the inductive premiss gets the value 0 in our
partial interpretation. This, then, is the mistake in the paradoxical reasoning:
the inductive premiss is false. So why is the paradox compelling? Consider
the claim ‘The cut-off is at n’. For each 11 ≤ n ≤ 100, this claim is true on
exactly one of our admissible extensions. Thus, in our partial interpretation,
it is neither true nor false. There is, then, no point n such that we can truly
say ‘The cut-off is at n’. So far, so good. But here, says the supervaluationist,
is where we make a natural mistake, and so get drawn into the paradox.
We conclude from the foregoing that ‘There is no cut-off’ is true—or in
other words, we conclude that the inductive premiss is true. But here we
are mistaken. ‘There is no cut-off’ is false on every admissible extension
(for each extension puts the cut-off somewhere—just at a different point on
different extensions), and hence—according to the supervaluationist—false
simpliciter. Thus, the paradox is compelling because we tend to assume that
because the cut-off is not here, or here, or here, . . . through all the possible
positions where the cut-off might be, it follows that ‘There is no cut-off’
is true—that is, that the inductive premiss is true. On the supervaluationist
84
account, this does not follow. There is no point n such that we can truly
say ‘The cut-off is at n’, and yet ‘There is a cut-off’ is true simpliciter —and
so the inductive premiss, which denies this, is false simpliciter.
The major objection unique to supervaluationism arises directly from the
solution to the Sorites just presented. I call it the problem of missing witnesses
and counterexamples, or just the ‘problem of missing witnesses’ for short.
Intuitively, if a universal claim is false, then there must be a counterexample
which makes it false, and if an existential claim is true, then there must be
a witness which makes it true. For example, if ‘Everyone is female’ is false,
there must be some particular person who is not female, and if ‘Someone is
female’ is true, there must be some particular person who is female. But on
the supervaluationist view, these relationships between quantified claims
and instances do not hold. Consider the inductive premiss of a typical
Sorites argument:
For every x in the series, if Fx then Fx .
As we have seen, the supervaluationist account has this come out false. This
would naturally lead us to suppose that one of the following sentences is
true, to give us a counterexample to the universal claim:
Fx1 but not Fx2
Fx2 but not Fx3
..
.
Fxn−1 but not Fxn
However, according to the supervaluationist, none of these is true. Consider
the sentence ‘Fx but not Fx ’. There are several possible cases:
• If x and x are both clear cases of F, or if x and x are both clear
countercases of F, then the sentence is false on every admissible
extension and hence false in the partial interpretation.
• If x is a clear case of F and x is a borderline case, or if x is a borderline
case of F and x is a clear countercase, or if x and x are both border-
line cases of F, then the sentence is true on some admissible extensions
and false on others, and hence neither true nor false in the partial
interpretation.
Thus the sentence will be either false, or neither true nor false, but will
never be true. So we have a false universal claim without a counterexample
85
⁵⁷ Cf. e.g. Sanford 1976; Rolf 1984, 232; and Tappenden 1993, 564. An analogous problem arises
at the level of sentential connectives (which is not surprising, given the analogies between disjunction
and existential quantification, and conjunction and universal quantification): a disjunction can be true
without either disjunct being true, and a conjunction can be false without either conjunct being false.
86
These reactions fit with the recursive assignments of truth values, not the
supervaluationist assignments. Ordinary speakers hedge over ‘x is red’ and
‘x is not red’ when x is a borderline case of redness, and they hedge in just
the same way over ‘x is or isn’t red’ and ‘x is and isn’t red’. I have found
these reactions to be robust, over a range of examples. Where Bob is a
borderline case of baldness, ordinary speakers do not think that ‘Bob is bald
or he is not bald’ is clearly true, and they do not think that ‘Bob is bald and
he is not bald’ is clearly false: they react to these two sentences with just the
sort of hesitancy with which they react to the sentence ‘Bob is bald’; they
regard all three sentences as equally dubious. Similarly for ‘heap’, ‘tall’, and
so on.⁵⁸
We thus have two sets of intuitions on the table, one of which fits
naturally with the supervaluationist assignments of truth values, and one of
which fits naturally with the recursive assignments. But at least one other
set of intuitions has also been reported in the literature.⁵⁹ According to
this third set, when a is a borderline case of F, ‘a is F or a is not F’ is
not assertible (which is, prima facie, in conflict with the supervaluationist
view), but ‘It is not the case that a is F and that a is not F’ is assertible
(which is, prima facie, in conflict with the recursive view).
What are we to make of all this? There is much to say, but I will save
this discussion for §5.5. For now, I think that any fair-minded person who
steps back from her own favoured set of intuitions and even-handedly
surveys the literature would have to agree—given first that there are
conflicting (reports of the) data, and second that some of the data do not
fit (easily) with the supervaluationist view, some do not fit (easily) with the
recursive view, and some do not fit (easily) with either view—that there
is no simple, knockdown intuitive argument here either way (i.e. in favour
of supervaluationism over the recursive approach, or vice versa). This is
still news: for in some quarters it is thought that the supervaluationist’s
penumbral connection argument against recursive approaches is a simple,
knockdown argument from intuition. In light of the foregoing, I consider
that position untenable—and even more so once we notice that the
feature of the supervaluationist semantics which leads to the assignment
of True to ‘Bob is or isn’t bald’ even when ‘Bob is bald’ and ‘Bob isn’t
bald’ are assigned no truth value is the very same feature which leads
to the missing witness problem. It is hard to claim—in the context of
discussing truth-functionality—that this feature of supervaluationism gives
it a clear intuitive advantage over recursive approaches, when it is the very
feature that leads to the missing witness problem—the latter being a great
intuitive disadvantage of supervaluationism, and one which the recursive
approach does not share. The upshot, then, is that the debate between
supervaluationist and recursive approaches will need to be settled either
by a much more detailed argument from data about the assertibility of
sentences about borderline cases, based on proper empirical foundations
(see §5.5), or else by other means altogether.⁶⁰
⁶⁰ As far as the dialectic of this book is concerned, I present an argument against all non-degree
theories of vagueness (Part II) and a number of arguments against supervaluationist degree theories
(§2.4.1)—thereby leading us to recursive degree theories. I then return to a detailed discussion of the
truth-functionality objection to recursive degree theories in §5.5.
⁶¹ We start with a three-valued interpretation, which consists in a nonempty domain M together
with a (total) interpretation function I which assigns a truth value (1, 0, or ∗) to each propositional
88
have more than three truth values, however, then things can get more
interesting on supervaluationist versions of the many-valued approach than
they can on supervaluationist versions of the partial two-valued approach.
In particular, suppose we have a fuzzy interpretation—i.e. a many-valued
interpretation where the underlying algebra of truth values is the Kleene
algebra [0, 1] discussed in §2.2.1. Instead of extending it to a fuzzy model
of the entire language in the recursive way examined in §2.2.1, how might
we extend it to a model which assigns an element of [0, 1] to each sentence
as its degree of truth, but where the assignments to compound sentences
are not a function of the assignments to simple sentences? One way would
be to say that a compound sentence gets assigned the value 1 if it comes out
as having the value 1 on every classical extension; it gets assigned the value 0
if it comes out as having the value 0 on every classical extension; and it gets
assigned the value 0.5 in all other cases. This is not very attractive, however:
the assignment of 0.5 in the third case is arbitrary, and the proposal allows
simple and atomic sentences to have any degree of truth, while restricting
other types of sentence to the values 1, 0, and 0.5, which seems odd.⁶²
How, then, might we allow compound sentences to have the full range
of intermediate degrees of truth, but in a non-recursive way? One idea is
to introduce a further piece of machinery: a measure defined over the set
of classical extensions of our fuzzy interpretation.⁶³ We stipulate that the
measure is normalized, and we say that the degree of truth of a compound
sentence is equal to the measure of the set of classical extensions on which
it is true. A classical contradiction is true on no classical interpretation
constant, an object in M to each individual constant, and an n-ary relation on M (the characteristic
function of which is a total function from M n to our set of three truth values) to each n-ary predicate.
Instead of taking the recursive approach from this point on, we say that a classical interpretation
extends a many-valued interpretation if it agrees with all the assignments of classical truth values (i.e.
0 and 1) that the many-valued interpretation makes (this can be made precise by making the obvious
changes to our definition of a classical interpretation extending a partial interpretation); we introduce
the notion of an admissible extension of a many-valued interpretation (in a way perfectly analogous
to our introduction of the notion of an admissible extension of a partial interpretation); and we then
follow the supervaluationist route: a compound sentence gets assigned the value 1 on our many-valued
interpretation if it comes out as having the value 1 on every admissible extension, etc.
⁶² I am assuming here that atomic sentences get their truth values directly from the base interpretation,
rather than from the supervaluation—recall the two options regarding the assignment of truth values
to atomic sentences discussed on p. 78 above, and see below for further discussion of this point.
⁶³ A measure over a set S is a function μ from the power set of S to the non-negative real numbers,
such that μ(∅) = 0, and for any disjoint subsets A and B of S, μ(A ∪ B) = μ(A) + μ(B). Intuitively, the
measure μ assigns to each subset of S a size (a real number). The measure is said to be normalized if
μ(S) = 1.
89
⁶⁴ For views in the ballpark of degree-theoretic supervaluationism as just presented, see Kamp 1975;
1981, 234–5; and Lewis 1983 [1970], 228–9; 1983 [1976], 69–70. (Sanford 1993, 225 presents a different
sort of view, in which the admissible valuations are fuzzy rather than classical, and the supervaluation
assigns a sentence not a single degree of truth, but a range of values.) Note that, instead of our fuzzy
base interpretation, we could take the degree supervaluationist’s treatment of compound wfs and apply
it over a three-valued, or partial two-valued, base interpretation. But then we would face the opposite
problem to that faced by the idea of applying the original supervaluationist’s treatment of compound
wfs over a fuzzy base interpretation: we would allow compound sentences to have any degree of truth,
while simple and atomic sentences could only have the values 1, 0, or ∗.
⁶⁵ |X| is the cardinality of the set X.
⁶⁶ Thus Akiba 2004, 422 moves too fast when he says ‘‘suppose that we assign real numbers from 0
to 1 as degrees of Bruce’s baldness. If you accept . . . supervaluationism, such an assignment is very easy
to obtain: you may just assign the ratio that the number of . . . interpretations in which ‘Bruce is bald’
is true bears to the number of all [interpretations].’’ This will work only where there are finitely many
interpretations to consider.
⁶⁷ Our set of admissible extensions is not like a set of points in physical space, or in a space such as
Rn , where we have a natural notion of the size of a subset.
90
different reason: because one quarter of the classical extensions of our fuzzy
interpretation assign it the value 1. That is, on one-quarter of the admissible
ways of precisifying our language, this sentence comes out (completely,
utterly) true. I think it is misleading to have these two completely different
routes to truth disguised under the same description (i.e. an attribution
of a degree of truth between 0 and 1 inclusive). But this is not the real
problem—for we could gain facility with this framework, and never be
confused as to the origin or meaning of a particular sentence’s degree
of truth. The real problem is that the rules for compound sentences are
different from those for atomic and simple sentences, in a way that seems
arbitrary, unmotivated, and downright odd. If you say ‘Bob is bald’ and I
say ‘John is tall’, each of our statements is judged by looking at the world,
finding the extensions of our predicates and the referents of our names, and
seeing how they lie with respect to one another. However if, rather than
each of us saying one of these things, you (or I) say both of these things,
in the form ‘Bob is bald and John is tall’, then this claim is assessed in a
completely different way. Now we look at all the possible precisifications
of our language, and determine which of them make this sentence true; its
degree of truth is then the proportion of precisifications on which it comes
out true. But why should the assessment of two individual sentences be
so totally different from the assessment of their conjunction? It’s as if we
were to assess individual runners by their time around the track, and assess
relay teams not by their cumulative time around the track, but by how
colour-coordinated their clothes are.
I mentioned earlier (p. 78) that there are two options for the super-
valuationist as regards the assignment of truth values to atomic sentences.
First, he may regard atomic sentences as getting their truth values directly
from the base interpretation, without any role being played by the classical
extensions: the truth value of P(a1 , . . . , an ) on the base interpretation
M = (M, I) is the value assigned by I(P) to the n-tuple (I(a1 ), . . . , I(an )).
Second, he may regard atomic sentences as getting their truth values from
the supervaluation, just as compound sentences do. In the present discussion
of supervaluationism as applied to many-valued base models, I have so far
been assuming the first option (cf. n. 62 above). We can avoid the problem
just encountered by moving to the second option, according to which
atomic sentences—like compound sentences—are judged according to
the proportion of admissible extensions which make them true. Thus the
92
base interpretation assigns an atomic sentence the truth value 0.3 just in case
30 per cent of the admissible extensions of the base interpretation assign
this sentence the value 1. So now we avoid the strange situation in which
atomic sentences and compound sentences get their degrees of truth by
completely different means. Nevertheless, this second option faces severe
problems of its own. The view under consideration is that even though the
intended base interpretation assigns ‘is bald’ a function f as its extension,
and ‘Bob’ an object x as its referent, and f (x) = 0.3, still it does not follow
from this that ‘Bob is bald’ is 0.3 true. Rather, to determine the truth of ‘Bob
is bald’ we have to see what proportion of admissible classical extensions of
the base interpretation make this sentence true. But when we think about
it, this is a bizarre view. It flies in the face of the most basic intuitions about
truth: in particular, about the way truth is determined by the meanings of
our words together with the way the world is. Let’s think it through: the
view is that ‘Bob’ means this man and ‘is bald’ means that property—and this
man possesses that property to degree 0.3—and yet this does not make it
the case that ‘Bob is bald’ is 0.3 true. Rather, what makes the sentence 0.3
true is the fact that 30 per cent of the admissible classical extensions of the
base interpretation make this sentence true. One wants to say at this point
that what this view is calling the ‘truth values’ of atomic sentences have
very little to do with truth as we know it—i.e. with truth as something
that is jointly determined by meanings and the way the world is. We have
thus arrived at a situation whereby atomic and compound sentences get
their ‘truth values’ in the same way only by calling something the ‘truth
value’ of the atomic sentence which in fact has very little to do with truth.
Here is a way to bring out the worry. I noted in the original case of
supervaluationism built over a base of partial two-valued models that our
two methods of assigning truth values to atomic sentences always yield the
same results. In the case currently under discussion—i.e. supervaluationism
built over a base of fuzzy models—we have no such guarantee: there is no
reason to expect that I(P)(I(a1 ), . . . , I(an )) = x in the base model just in
case the measure over classical extensions assigns x to the set of models in
which P(a1 , . . . , an ) is true. So let us suppose for a moment that in the base
model, the extension of ‘is bald’ assigns the referent of ‘Bob’ the value 0.3,
while ‘Bob is bald’ is true in 40 per cent of the classical extensions of the
base model. Now let’s ask: how true do we want to say the sentence ‘Bob
is bald’ is? Is it 0.3 true, or 0.4 true? It seems clear that the value assigned
93
by the base model is the one that captures the sentence’s degree of truth,
and that the other value—the one assigned by the measure over classical
extensions of the base model—while it may (or may not) be interesting
in some other way, simply doesn’t have much to do with truth as we
know it. That much seems very clear when the values assigned by the
base model and by the measure over extensions diverge—but of course,
the moral carries over to the case where the values are the same (whether
by luck or because we impose it as a constraint on acceptable measures
that I(P)(I(a1 ), . . . , I(an )) = x in the base model just in case the measure
assigns x to the set of models in which P(a1 , . . . , an ) is true). Thinking
otherwise would be like thinking that the tachometer in my car measures
the frequency to which my car radio is tuned—as long as (whether by
accident or by rigid disciplining of my driving and/or listening habits) the
read-out on the tachometer is the same as the read-out on the radio tuner.
Thus, while degree supervaluationism seems like a neat idea at first sight,
it is ultimately a deeply unhappy one. Furthermore, the moral carries over
to the original form of supervaluationism (in which both the base model
and the supervaluation are partial or three-valued). The supervaluationist
faces a choice regarding the assignment of truth values to atomic sentences:
this can be done by the base model, or by the supervaluation. When we
originally noted this fact, we said that it might seem to be of little import,
because the two methods are guaranteed (in the case of the original form of
supervaluationism) to give the same results. But we can now see that neither
choice is satisfactory. We cannot have the truth values of atomic sentences
being assigned by the supervaluation, for this violates our most basic
intuitions about truth—in particular, about the way truth is determined
by the meanings of our words together with the way the world is. But
then if the truth values of atomic sentences are assigned by the base model,
while those of compound sentences are assigned by the supervaluation,
we have the bizarre situation in which (for example) the individual truth
assessments of two atomic sentences operate in a completely different way
from the truth assessment of their conjunction. Thus, the core idea of
supervaluationism—of any kind—is an unhappy one.
2.4.2 Subvaluationism
Given a base interpretation—many-valued, or partial two-valued—we
have seen two non-recursive ways of assigning truth values to compound
94
⁶⁸ A subvaluationist approach to vagueness is presented in Hyde 1997; for some further discussion
see Akiba 1999; Hyde 1999; Beall and Colyvan 2001; and Hyde 2001.
⁶⁹ A logical consequence relation |= is explosive just in case for all A and B, {A, ¬A} |= B; it is
paraconsistent just in case it is not explosive (sometimes this is called weak paraconsistency, with strong
paraconsistency consisting in the fact that A ∧ ¬A B). A logic is paraconsistent just in case its logical
consequence relation is paraconsistent. A theory T is inconsistent just in case for some A, {A, ¬A} ⊆ T;
it is trivial just in case for all B, B ∈ T. Thus, paraconsistent logics provide the basis for inconsistent but
non-trivial theories. See e.g. Priest and Routley 1989.
⁷⁰ This was the first formal system of paraconsistent logic (originally presented in 1948). See Hyde
1997, 648, and Jaśkowski 1969.
95
⁷¹ Multiple-conclusion consequence may be thought of as follows. Where and are sets of wfs,
|= just in case in every model in which every member of is true, some member of is true.
Where contains only one wf, this reduces to the usual notion of single-conclusion consequence.
(In the case of singleton sets of wfs, we typically just write the wf, without set-brackets around it; for
example we write {A, ¬A} |= B rather than {A, ¬A} |= {B}.)
⁷² In the alternative way of proceeding, where the base model is two-valued, suppose that A
and ¬A are assigned neither value in the supervaluationist case, and are assigned both values in the
subvaluationist case.
96
language and the world. All these views deny any indeterminacy in the
relationship between language and the world; all the indeterminacy is in the
world itself. This is because they posit a unique intended base interpretation
which has some form of inherent indeterminacy (different forms in the
different kinds of base interpretation we have considered—but always
some form). True, they also posit a multiplicity of classical models which
extend this base interpretation—but these classical models are not admissible
interpretations of the language. Rather, they are admissible extensions of
the uniquely correct base interpretation of the language. They play an
auxiliary role in calculating the semantic value of compound wfs on the
base interpretation—they have no direct semantic significance. In short, the
multiplicity of admissible extensions does not introduce semantic indeterminacy.
This is in sharp contrast to the plurivaluationist picture, to which we
turn now.
are as a matter of fact true). But plurivaluationists do not go this far. They
do not jettison the idea of correct or intended interpretations altogether:
they just say that in general there is not a unique correct or intended
interpretation. The way that a given speech community uses a language
places constraints on what is an acceptable interpretation of that language.
For example, we apply the name ‘Helen Clark’ to a certain person, so on
acceptable interpretations, this person is the referent of that name; we apply
the term ‘apple’ to certain objects, so on acceptable interpretations, these
objects are in the extension of that predicate; and so on. Now the classical
view holds that the set of acceptable interpretations has one member, which
we called the intended/correct/etc. interpretation. The plurivaluationist
denies this—but retains the notion of an acceptable interpretation. Thus,
we have a privileged class of interpretations, but not a single interpretation
privileged above all others.
One natural motivation for plurivaluationism would be a fondness for
classical interpretations, coupled with a conviction that epistemicism cannot
adequately answer the location problem. The plurivaluationist can regard a
unique intended classical interpretation as a goal which we never actually
reach: our usage rules out certain interpretations as incorrect, but never
manages to narrow down the remaining set of acceptable interpretations to
just one member.
Now to return to our question about truth simpliciter. We have a class
of acceptable interpretations. Suppose that a given sentence is true on
all of them. Then in an obvious sense it does not matter that we have
many acceptable interpretations and not just one: we can still say that our
sentence is true. Similarly, if our sentence is false on all the acceptable
interpretations, then again it does not make any difference that we have
many acceptable interpretations and not just one, and we can say that
our sentence is false. But if our sentence is true on some acceptable
interpretations and false on others, then it does matter that we have many
acceptable interpretations and not just one: we can say neither that our
sentence is true simpliciter, nor that it is false simpliciter.
This may sound like supervaluationism: we have a range of acceptable
classical models, and our sentence comes out true/false simpliciter if it is
true/false on all of them. Plurivaluationism is, however, crucially different
from supervaluationism. This has not been appreciated in the literature: the
100
two positions have never been distinguished.⁷³ Contrast the following two
quotations:
Broadly speaking, supervaluationism tells us two things. The first is that the
semantics of our language is not fully determinate, and that statements in this
language are open to a variety of interpretations each of which is compatible with
our ordinary linguistic practices. The second thing is that when the multiplicity of
interpretations turns out to be irrelevant, we should ignore it. If what we say is
true under all the admissible interpretations of our words, then there is no need to
bother being more precise.
(Varzi 2003b, 14)
In van Fraassen’s original terminology, for each acceptable U-model U, the function
that assigns to each sentence its truth value in U is called a classical valuation, while
the function that assigns the value true to those sentences true in every acceptable
U-model is called a supervaluation.
(McGee 1997, 154 n. 12)
⁷³ Even Varzi 2007, §1, which seeks to lay out all the options, does not make this distinction.
⁷⁴ This is not intended as a criticism of Varzi: his usage is not out of line with the vagueness literature.
I have singled out Varzi only because the passage quoted gives such a clear description of one of the
two views that I want to distinguish.
⁷⁵ Because supervaluationism and plurivaluationism have not been distinguished in the literature, it
is impossible to say of previous authors who espouse views in the ballpark of either of these positions
whether they should be seen as advocating supervaluationism, plurivaluationism, or some (possibly
confused) mixture of the two. Hence I shall make no attempt to classify existing ‘supervaluationist’ views
with respect to this distinction. Three views should be mentioned in connection with plurivaluationism,
however. One is the partial denotation view of Field 1973, 1974. Field clearly espouses the idea (central
to plurivaluationism) that the various classical models in his view play the role of equally acceptable
interpretations of the language. (This contrasts with the supervaluationist view of classical models as
admissible extensions of a unique intended non-classical interpretation.) However, Field does not draw
the distinction between regarding the truth condition ‘true if true on all models’ as a mere manner
of speaking (as in the plurivaluationist view), as opposed to a further piece of semantic machinery
(as in the supervaluationist view), and so ultimately it is not clear where to classify Field’s view. The
second view which should be mentioned is that of Przełȩcki 1976. Przełȩcki is explicit that (as in
101
But how exactly are the views different? Many views countenance a range
of classical models: this is common to supervaluationism, plurivaluationism,
and contextualism (see §2.6), among others. What distinguishes each of
these views is the role it gives to these classical models, in conjunction with
any further semantic machinery that the view introduces.
Consider supervaluationism. As I was at pains to point out, the classical
models are admissible extensions of a unique intended partial interpretation
of the language. They do not give the content of the language in its
actual, unprecisified state: they interpret the language as it would be were it
precisified. They do not stand in a direct interpretational relationship to the
language. Rather, they serve as calculating devices which yield an assignment
of truth values to (actual, vague) sentences of the language. This assignment
of truth values, together with the other assignments made in the base
interpretation—not in its classical extensions—play the interpretational
role: they are what constitute a description of the actual semantic state of
the language. In the supervaluationist view, each sentence is in a unique
semantic state, arrived at via a supervaluation over many classical models.
The truth value of a sentence—i.e. the one assigned to it, on the base
interpretation, by the supervaluation (as opposed to any of its truth values
on the various classical extensions of this base interpretation)—represents
the one actual semantic state of the vague sentence. This truth value merely
happens to be arrived at via a description of idealized precisifications which
misrepresent the actual, vague state of the language.
Plurivaluationism, by contrast, has no further semantic machinery, besides
the classical models. In the final story about the actual semantic state of the
language, there is no other, non-classical model: there are only the classical
plurivaluationism, not supervaluationism) his view involves no non-classical semantic machinery, only a
multiplicity of classical models, and he makes the point (crucial to plurivaluationism: see §2.5.1—which
was written before I read Przełȩcki’s paper) that on his view ‘‘reality is sharp—composed of exact
structures. It is the connection between language and reality that is fuzzy—linking language not to
a single exact structure, but to a whole class of such structures’’ (pp. 379–80). The classification of
Przełȩcki as a plurivaluationist is rendered problematic, however, by his claim (p. 380 n. 3) that Fine’s
(supervaluationist) view is similar to his: whether this foils the classification depends upon just how
similar he thinks Fine’s view is to his own. The third view is the account of abstract mathematical
discourse presented by Islam 1996. Like Field, Islam makes it clear that the various classical models in
his view play the role of equally correct interpretations of the language, and, like Przełȩcki, he makes
it clear that his view involves no non-classical semantic machinery. Unlike the plurivaluationist view
of vagueness, however, it is no part of Islam’s view that a multiplicity of correct interpretations goes
hand in hand with speakers failing to specify exactly what their terms are to mean: on his view, talking
relative to multiple interpretations is often exactly what we want to do.
102
ones. Together, they give the content of the language in its actual, vague
state. Each of them is an interpretation of the language, not a mere extension
of the one uniquely correct interpretation. (This is why I have used
different terms for the classical models in the context of the different views:
admissible extensions versus acceptable interpretations.) In the plurivaluationist
view, there is no further semantic story, beyond that told by the many
acceptable interpretations. Here we have indeterminacy of meaning: there are
many acceptable interpretations, and that’s all there is to say. Thus each
sentence is not in a unique semantic state: semantic states are individuated
by interpretations, and each sentence has many of them. On this view, the
actual semantic state of the sentence is indeterminacy of truth value: there
are many acceptable assignments of truth value to the sentence, and nothing
to decide between them. That is a complete and accurate description of the
actual semantic state of the sentence. Rather than indeterminacy, we might
just as well describe this as plurality or multiplicity of meaning. The essential
point remains: each sentence is not in a unique semantic state, as semantic
states are individuated by interpretations, and each sentence has many of
them, with nothing to decide between them.
Figure 2.2 is an attempt to give visual form to the distinction between
supervaluationism and plurivaluationism. Suppose the request is made to
describe the semantic state of a given sentence. The plurivaluationist can say only:
it is true on this interpretation, false on this one, and so on, through all
the acceptable interpretations. This is because these interpretations are the
only pieces of semantic machinery she has. Her picture is of a language which
simultaneously has many classical interpretations, and whose semantic
state is thereby indeterminate (or equivalently, which is thereby in many
semantic states simultaneously). The supervaluationist, by contrast, will
describe the sentence as being in a unique, determinate semantic state. For
in her picture, she has further semantic machinery, over and above the classical
models: partial interpretations, which assign truth values to compound
sentences via supervaluations over classical models. For her, the sentence
is not in an indeterminate first-order (classical) semantic state—it is in a
perfectly determinate higher-order (non-classical) semantic state.
Note that while there are important analogies between plurivalu-
ationism, supervaluationism, and possible-worlds semantics for modal
languages—analogies which have, for example, enabled formal results
from modal logic to be applied in discussions of supervaluationism in the
103
supervaluationism
plurivaluationism
language language
key:
classical non-classical interprets/ extends
model model gives the meaning
Figure 2.2. Plurivaluationism and supervaluationism.
language
key:
The crucial point to note is that this view of vagueness fits with pluri-
valuationism, not supervaluationism. We saw in §2.4.3 that on the super-
valuationist approach, vagueness—or at least indeterminacy—is found in
the world itself, not in the relationship between language and the world.
As far as the issue of worldly vagueness is concerned, the supervaluationist
approach over a certain sort of base interpretation is in exactly the same
boat as the recursive approach over that sort of base interpretation: the base
interpretation brings with it worldly indeterminacy (e.g. partial properties),
and the superstructure cannot then take it away. The supervaluational
approach differs from the recursive approach in positing worldly indeterm-
inacy and a ‘non-local’ (non-truth-functional) account of how compound
sentences get their truth values.
The plurivaluationist, by contrast, locates vagueness entirely in the
relationship between language and the world. The plurivaluationist model
theory gives us a picture on which each n-place predicate is associated
with many n-place relations (one in each acceptable interpretation), each
of which is inherently precise. The idea is that when we use a vague
predicate, we do not refer to a unique relation, but all the relations we
simultaneously refer to are inherently precise. Suppose ‘Bob is tall’ comes
out true on some acceptable interpretations of ‘tall’ and false on others.
Then we can say neither that ‘Bob is tall’ is true nor that it is false. But,
unlike on the supervaluationist approach, this does not mean that the property
of tallness is inherently indeterminate. The point is rather that there is no
unique property of tallness. There are many properties, each corresponding
to an acceptable extension of ‘tall’, and all precise, in the sense that each
object either definitely possesses or definitely fails to possess each property.
⁷⁷ In fact, we employ this way of speaking even if most persons believe the government will be
voted out: it does not matter if a few don’t. To get the analogy with plurivaluationism right, we need
to suppose for a moment that we say ‘‘The man on the street ’s’’ only when all persons on the street
. Cf. n. 78 below.
110
1. those who want to say that a sentence is neither true nor false when it is
true on some acceptable interpretations and false on others (cf. the basic
form of supervaluationism)
2. those who want to say that a sentence is half true when it is true on
half of the acceptable interpretations and false on the other half (cf. the
degree form of supervaluationism)
3. those who want to say that a sentence is both true and false when
it is true on some acceptable interpretations and false on others (cf.
subvaluationism).
111
2.5.4 Pragmatism
There are several approaches to vagueness to be found in the writings of
David Lewis. One of them is distinctive (or at least apparently so—but
see on) in that it locates vagueness not in semantics, but in pragmatics.
Lewis sets up a framework in which a language (in one sense) is a
function which assigns meanings to sentences and their components: the
meaning of a sentence is a function from possible worlds to truth values;
the meaning of a predicate is a function from possible worlds to sets of
possible objects; and so on. Lewis then has a story to tell about how a
particular population P comes to use a particular one of these languages £.
⁷⁸ There are many other possibilities besides these three. For example, we can also imagine a
plurivaluationist who wants to say that a sentence is true/false when it is true/false on most acceptable
interpretations. Cf. n. 77 above, and Lewis 1983 [1979], 244. Another view worth mentioning is that
of Braun and Sider 2007, according to which a sentence is true just in case it expresses a unique
proposition which is true. On this approach, we would say that a sentence which has many admissible
interpretations is untrue, regardless of its truth values on its various admissible interpretations.
112
⁷⁹ See also Lewis 1969, 200–2: ‘‘I think we should conclude that a convention of truthfulness
in a single possible language is a limiting case—never reached—of something else: a convention of
truthfulness in whichever language we choose of a tight cluster of very similar possible languages.
The languages of the cluster have exactly the same sentences and give them corresponding sets of
interpretations; but sometimes there are slight differences in corresponding truth conditions . . . . Our
actual language is like a resonance hybrid of the possible languages that make it up.’’ Lewis does not
appear to be very attached to this pragmatic view. Elsewhere he writes: ‘‘we have so far been ignoring
the vagueness of natural language. Perhaps we are right to ignore it, or rather to deport it from semantics
to the theory of language-use. We could say, as I do elsewhere, that languages themselves are free of
vagueness but that the linguistic conventions of a population, or the linguistic habits of a person, select
not a point but a fuzzy region in the space of precise languages. However, it might prove better to
treat vagueness within semantics, and we could do so as follows’’ (Lewis 1983 [1970], 228). Lewis then
proceeds to present a degree-theoretic form of supervaluationism (see §2.4 above). Burns 1991, 1995,
on the other hand, defends the pragmatic view as the correct theory of vagueness.
⁸⁰ ‘‘By a language, I mean an interpreted language. So what I call a language is what many logicians
would call a language plus an interpretation for it. For me there cannot be two interpretations of the
same language; but there can be two languages with the same sentences’’ (Lewis 1969, 162).
113
way of saying that our practice does not fix a unique correct interpretation
of our (uninterpreted) language, but only a cluster of such interpretations.
And that is plurivaluationism! Pragmatism, then, is simply a stylistic variant
of plurivaluationism.⁸¹
2.6 Contextualism
While they differ in many ways from one another, the views we have
looked at so far can all be regarded as having one thing in common:
they offer synchronic accounts of vagueness. They hold that in order to see
what is distinctive of vague language, we need to look at two areas: the
nature of its interpretations (classical, gappy, many-valued, etc.) and the
number (one or many) of these interpretations which are singled out as
correct or acceptable. No mention is made here of how (if at all) which
interpretation(s) is intended changes over time or with context. No doubt
the proponents of the views we have examined would be prepared to admit
that such changes do, or at least might, occur: for example, it is plausible
that ‘today’ uttered today has a different referent from the one it had when
uttered yesterday. But the proponents of these views will not think that
such changes have anything in particular to do with vagueness. This is
where contextualism is distinctive. The basic idea behind contextualism is
that vagueness is a diachronic phenomenon, which only emerges when we
consider the semantic state of a language over time (or more generally,
over multiple instances of interpretation). It is to be accounted for not
(primarily) in terms of the nature and number of correct interpretations
(at any time), but in terms of how the intended interpretation(s) changes
over time. Hence the opening sentence of Tappenden (1993): ‘‘This paper
develops one aspect of a program aimed at studying normative constraints
on the evolution of language’’, and Soames (1999, 209): ‘‘We need a dynamic
model of vagueness’’ (my emphases).
⁸¹ In fact, the interpretations in Lewis 1983 [1970] are more complicated than the classical
interpretations countenanced by the plurivaluationist, so we need to be a little more subtle—but my
basic point remains. The point can be more accurately stated thus: relative to a given syntax and a given
sort of interpretation, plurivaluationism and pragmatism are simply stylistic variants of one another. Keefe
1998a argues (independently) that pragmatism either fails or collapses into supervaluationism (see also
Keefe 2000, ch. 6). I agree with many of her points, but wish to stress that in my terms, pragmatism is
the same as plurivaluationism, not supervaluationism.
114
Consider a vague predicate: say, ‘is tall’. Here is one way in which the
extension of this predicate might be thought to vary with context. Suppose
I say ‘Bill Bradley is tall’. If we are engaged in a discussion of current
and former basketball players, my claim would seem to be false, because
Bill Bradley is below the average height of current and former basketball
players. If we are engaged in a discussion of past presidential candidates,
my claim would seem to be true, because Bill Bradley is among the tallest
of past presidential candidates. Plausibly, what is going on here is that ‘is
tall’ has a different extension in the different contexts: in the first context
it excludes persons who are not of a height significantly greater than the
average height of current and former basketball players; in the second
context it excludes persons who are not of a height significantly greater
than the average height of past presidential candidates. Many would agree
with (something like) this story. Yet clearly this story does not get to the
heart of the issue of the vagueness of ‘is tall’—if in fact it touches this issue
at all. There are two ways of seeing this. First, the vagueness of ‘is tall’
emerges within a fixed context of the sort just considered: for example, we
can easily generate a Sorites series for the predicate ‘is tall (for a basketball
player)’, and likewise for the predicate ‘is tall (for a presidential candidate)’.
Second, the phenomenon under consideration affects precise terms, such as
‘act of parliament’. Assuming that the above story about ‘is tall’ is plausible,
it is also plausible that ‘is an act of parliament’ has a different extension when
uttered in a discussion of English law from the one it has when uttered
in a discussion of Australian law. Yet there is no vagueness about whether
something is an act of the English parliament or whether something is
an act of the Australian parliament. It thus seems clear that ‘large-scale’
contextual variation of extension of the sort just discussed is different from
vagueness, which arises even within a fixed (large-scale) context. Accounts
of vagueness should therefore kick in and do their distinctive work after the
(large-scale) context has been fixed.
So contextualism about vagueness is not the view that ‘tall’ has a different
extension when we are discussing basketball players from when we are
discussing jockeys, that ‘bald’ has a different extension when we are looking
for a male model for a shampoo commercial from when we are looking for
a male model for a scalp oil commercial, and so on. That thesis—about
what I am calling large-scale contextual variation—is not an alternative to
any of the views we have discussed, but a near truism which proponents of
115
⁸² Given (as discussed in the opening paragraph of this section) that contextualism differs from the
views we have considered on a dimension orthogonal to any of the dimensions on which those views
differ from one another, this is not surprising.
⁸³ Contextualist treatments of vagueness include Kamp 1981; Tappenden 1993; Raffman 1994, 1996;
Soames 1999, ch. 7; Fara 2000; and Shapiro 2006. Tappenden, Soames, and Shapiro explore versions
of the strong Kleene recursive truth gap approach, while Fara is sympathetic towards classical models
as the basis for contextualism.
⁸⁴ This formulation is due to Shapiro 2006, who calls this thesis ‘open texture’.
⁸⁵ Shapiro 2006 calls this thesis ‘judgement-dependence’.
116
important aspect to the contextualist story. It is not only the object a that
moves into the extension of P when a competent speaker judges that Pa;
other objects move too, according to the rules of adjustment associated with
P.⁸⁶ For example, if I decide for present purposes in the course of some
conversation (for example, about what T-shirt sizes to order, where the
options are S, M, L, and XL) that Bill is tall, then it thereby becomes
true that Bill is tall, and it also becomes true of anyone who is negligibly
different in height from Bill that he or she is tall (i.e. an interpretation
which assigns an extension to ‘is tall’ which assigns 1 to all these people
thereby becomes the intended interpretation of our language, at this point
in the conversation or discourse).⁸⁷ Note that this further adjustment just
happens: I do not have to explicitly consider these other people, and say
that they are tall. I simply consider Bill, and say he is to count as tall; this
affects the context, and we get a new correct interpretation in which not
just Bill, but those negligibly different in height from him, are assigned 1 by
the extension of ‘is tall’. It is as if Bill is bound to those who are negligibly
different from him in height: when I move him into the positive extension
of ‘is tall’ (i.e. into the set of things sent to 1 by the extension of ‘is tall’),
those very similar to him move with him.
The contextualist idea of deciding borderline cases in a partial interpret-
ation one way or the other is related to, but importantly different from,
the supervaluationist’s idea of extending a partial interpretation. For the
supervaluationist, an extension decides all borderline cases of all predicates
one way or the other, resulting in a classical interpretation. In the contex-
tualist picture, we decide one borderline case of one predicate one way or
the other. The rules of adjustment ensure that other borderline cases of this
predicate thereby get decided too—but in general, not all the borderline
cases of this predicate get decided, and in general, other predicates are not
affected.⁸⁸ Thus after the adjustment, we still have a partial interpretation,
not a classical one.
What is the advantage of adding a contextualist story on top of the strong
Kleene recursive truth gap view? I shall argue in Chapter 4 that a major
problem with all versions of the truth gap view (and of the three-valued
view) is that they do not allow for a gradual or jolt-free transition between
the clear cases and countercases: they posit two sharp semantic jumps in
a Sorites series for a predicate F —one between the objects of which the
predicate is true and those of which it is neither true nor false, and one
between the objects of which the predicate is neither true nor false and
those of which it is false.⁸⁹ One apparent advantage of contextualism over
the (non-contextualist) strong Kleene recursive truth gap view is that it can
explain why we do not think there are such sharp transitions (even though
in fact there are, in any context).⁹⁰ Suppose that in the current context, Bob
is the last tall man: in the interpretation which is correct at this stage of the
conversation or discourse, Bob is sent to 1 by the extension of ‘is tall’, while
everyone shorter than Bob is sent nowhere or to 0. If I wanted to identify
the boundary between the tall and the borderline cases as lying between
Bob and the next man, I would have to say ‘Bob is tall, and the next man
is a borderline case’. But now we have a problem: when I say ‘Bob is tall’,
the context shifts in such a way that everyone negligibly different from Bob
is also sent to 1, and this includes the next man. So I generate a new context
in which what I set out to say is not true: in the new context, Bob is tall,
and so is the next man—he is no longer a borderline case. The rules of
adjustment ensure that when I classify Bob as tall, the next man goes with
him. Thus I can never state where the boundary between the positive and
borderline cases (or negative and borderline cases) is without generating a
new context in which the boundary is somewhere else. We can never nail
down the borderline: just as we are about to get it in our sights, it shifts. The
contextualist can thus explain why we do not think there are sharp semantic
jumps in a Sorites series for a predicate F (even though there are such jumps,
in any context, on this view), in a way that the basic truth gap view cannot.
The response to the Sorites paradox on the part of the recursive truth
gap view was as follows: the error in the paradoxical reasoning is that some
of the Sorites conditionals are not true; the reason we are taken in by the
paradox is that none of the conditionals is false, either. The contextualist
⁸⁹ As I have mentioned, this jolt problem is one of two problems which are commonly run together
under the heading ‘higher-order vagueness’. The other one is the location problem: the problem of
how the positions of these jumps could be determined by our usage of vague terms. This, as I have
argued and will argue further, is much less of a problem for the truth gap and three-valued views.
⁹⁰ I say the advantage is apparent, because I shall argue in Ch. 4 that contextualism does not solve
the jolt problem.
118
inherits this solution,⁹¹ but can add something more as well. First, not only
are none of the Sorites conditionals false (in any context), but furthermore,
if we think about any of the conditionals, it is liable to become true. Recall
that if we say that Bill is tall, it thereby becomes true that Ben is tall also
(where Ben is just after Bill in the Sorites series). Now of course if we
merely entertain the conditional ‘If Bill is tall, then Ben is tall’, we do not
make it true; but if we say for the sake of argument that Bill is tall, and
then consider whether, given that Bill is tall, Ben is tall, we do make the
conditional true. This provides an added reason why we tend to think that
the Sorites conditionals are true, and hence get taken in by the paradox.
Second, the contextualist has a response to the dynamic version of the Sorites
paradox (also known as the ‘forced march’ Sorites paradox (Horgan 1994)).
Suppose we are walked along our Sorites series for F, and asked of each
object in the series whether it is F, and then walked back the other way,
and asked the same question of each object again. It is very likely that the
point at which we stopped saying ‘Yes’ on the way out would be further
along the series than the point at which we started saying ‘Yes’ on the way
back. This behaviour might seem rather difficult to explain on the recursive
truth gap view, but it is easily explained by the contextualist: as we classify
an object as F, we thereby make it true that that object and all others very
similar to it—including the next object in the series—are F. We thus push
the boundary between the F’s and the borderline cases out before us as we
go—and on the way back, we push the boundary back the other way.⁹²
Let us now consider other forms of contextualism. If we take a non-
recursive truth gap view—say, supervaluationism—as our foundation, we
get a very similar view to the one just presented. At any stage of the
conversation or discourse, one partial interpretation is the intended one;
speakers have the freedom to classify borderline cases of a predicate one
way or the other, and so classifying them changes which interpretation is
intended in such a way that the classification is true. The only difference from
⁹¹ In fact, not all contextualists accept their inheritance. For example Soames 1999, 214 ff. offers a
rather more complex solution to the regular Sorites paradox, based upon the solution to the dynamic
Sorites paradox provided by his contextualism (see on).
⁹² There is clearly a problem looming here when we get all the way through the borderline cases:
when we classify the last borderline case as P, we thereby push the next object in the series into the
positive extension of P also. But the next object in the series was not a borderline case of P —it was a
clear negative case—and Freedom and Power tell us only that we can make borderline cases into positive
or negative cases. For an objection along these lines directed specifically at Soames’s contextualism, see
Robertson 2000; for a reply see Soames 2002.
119
⁹³ Goguen 1968–69, 351 considers a view according to which the algebra of truth values may vary from
context to context. This is much more radical than fuzzy contextualism in my sense, according to which
the intended interpretation may vary from context to context, but is always a fuzzy interpretation—i.e.
is always based on the same underlying algebra of fuzzy truth values.
⁹⁴ This might not in fact involve a change of intended interpretation if a already was P (unbeknownst
to the speaker) and the speaker asserted Pa, or if a already was not P (unbeknownst to the speaker) and
the speaker asserted ¬Pa. But even in these cases, it might still involve a change of interpretation, if
a was very close to the border: the rules of adjustment might then shift the border, even though the
border does not cross a.
120
world itself. The answer is that this depends entirely upon what sorts
of interpretations are chosen as the basis for the contextualist view in
question. Worldly vagueness is a matter of the internal nature of our
models; semantic indeterminacy is a matter of the relationship between a
language and its models. A model is a picture of what the world is like.
Classical models, for example, tell us that all the properties and relations
in the world are precise: for each property and each object in the world,
the object either possesses that property (outright) or does not possess
it (at all). Fuzzy models, for example, tell us that the world contains
inherently vague properties and relations: there are objects and properties
such that—out there in the world, quite independently of human thought
and language—that object possesses that property to an intermediate
degree. So if a contextualist builds her view on classical foundations, she
posits no worldly vagueness; if she builds her view on fuzzy foundations,
she does posit worldly vagueness; and so on for the other sorts of models
she might take as the basis for her view. If there is worldly vagueness in the
base models, the contextualist superstructure cannot take it away; if there is
no worldly vagueness in the base models, the contextualist superstructure
cannot introduce it.
Semantic indeterminacy enters when a view tells us that which object or
property is picked out by a name or predicate is indeterminate: there are
multiple, equally correct answers. In other words, semantic indeterminacy
enters when a view tells us that instead of one correct interpretation, a
language has (at some time) many acceptable interpretations. This is what
plurivaluationism tells us; it is not what contextualism tells us (except for
contextualism built on a plurivaluationist foundation). The contextualist
tells us that meaning is variable —it is one unique way (there is one
intended interpretation) at any given time, but different ways at different
times—not that it is indeterminate. (Contextualism built on plurivaluationist
foundations tells us that meaning is both indeterminate—at a time—and
variable—over time.)
1. ∀n (If a man with n hairs is bald, then a man with n + 1 hairs is bald).
2. A man with 0 hairs is bald.
3. ∴ ∀n (A man with n hairs is bald).
We also accept that 2 is true and 3 is false. Thus we conclude that 1 is false,
and we accept its negation. However, reasoning intuitionistically, we are
not then forced to conclude:
4. ∃n (A man with n hairs is bald, and a man with n + 1 hairs is not bald)
⁹⁵ For further discussion of Putnam’s view see Schwartz 1987, 1990; Rea 1989; Schwartz and Throop
1991; Putnam 1991; Williamson 1996a; and Chambers 1998. For a different intuitionist approach to
vagueness, see Wright 2001.
124
¹ Cf. Suppes 1957, 151: ‘‘Many textbooks . . . promulgate the four traditional ‘rules’ of definition:
1. A definition must give the essence of that which is to be defined. 2. A definition must not be circular.
3. A definition must not be in the negative when it can be in the positive. 4. A definition must not be
expressed in figurative or obscure language.’’ 1 and 3 come under my heading ‘fundamental’; 2 and 4
under ‘useful’.
130
nevertheless, when all the pros and cons of the various theories are weighed
up, one theory comes out ahead of the others on balance (cf. Keefe 2000).
The strong form of cost/benefit analysis offers no way forward at this
point in the debate, because while some theories are generally regarded as
untenable, there is a range of live alternatives in the literature which have
been developed and defended to such a level of sophistication that it is
simply not credible to say that all but one of them is untenable. In short,
while the strong form of cost/benefit analysis would be great if it worked,
in fact it simply fails to yield a winner in the case of theories of vagueness.
The moderate form of cost/benefit analysis also offers no clear way forward.
It suffers from two problems. First, it is inherently limited to yielding a
provisional result at best: for new objections and defences are coming to
light in the journals at a great rate. Second, two authors can weigh up the
same theories—considering the same objections to, and replies on behalf
of, each—and yet come to a different conclusion concerning which theory
is best on balance. Thus the cost/benefit model desperately needs to be
supplemented by theory-neutral criteria for weighing up theories: and we
have no such criteria.
Related to this second point, it seems that the search for criteria for
weighing up the costs and benefits of rival theories of vagueness cannot
succeed independently of the project of giving a fundamental definition
of vagueness—and so moderate cost/benefit analysis is not, in the end,
an alternative to my own strategy. Let me explain. In the debate on
vagueness, one often hears arguments such as the following: ‘‘My theory is
better than yours, because mine retains the validity of the law of excluded
middle, while yours does not.’’ Now in fact the weight of this argument
is entirely unknown, unless we have some idea of the defining features of
vagueness. For suppose it is the case that the essential feature of vague
predicates is that they admit borderline cases (in some specific sense), and
suppose that no theory which retains excluded middle can allow for the
existence of predicates which have borderline cases (in this sense). If that
were the case, then far from being a problem, abandoning excluded middle
would be a necessary condition on any correct theory of vagueness. Now
I am not saying that these claims about borderline cases and excluded
middle are true. I am simply saying that we cannot assess the relevance
of retaining excluded middle until we have investigated such questions.
A fundamental definition of vagueness will answer these questions. If
131
² For example, Greenough 2003 asserts that by characterizing vagueness from a neutral standpoint,
we ‘‘can at least ensure that we are all talking about the same thing from the outset in our inquiry into
the nature and source of vagueness’’ (p. 235), and claims that if we failed to be neutral in our initial
characterization of vagueness, then there would ‘‘be a very real sense in which there would be no
disagreement about the character of vague language at all since each partisan would mean something
different by the predicate ‘is vague’ ’’ (pp. 238–9). On a similar note, Bueno and Colyvan 2006 write:
‘‘A definition of vagueness should not prejudice the question of how best to deal with it’’ (p. 1); ‘‘What
we’re after is a definition of ‘vagueness’ that does not beg any questions about how vagueness is best
treated’’ (p. 4); and ‘‘without a definition of ‘vagueness’ it is not even clear that the various theories are
theories of the same phenomenon’’ (p. 5). And Williamson 1994, 2 writes that ‘‘we can agree to define the
term ‘vagueness’ by examples, in order not to talk past each other when disagreeing about the nature
of vagueness’’. See also Shapiro 2006, 1.
³ I do not mean here to be criticizing the authors cited in n. 2, for I am not claiming that they
intended ‘definition’ in the sense of ‘fundamental definition’.
⁴ In order to get the point of these examples, this needs to be understood as the (actual) chemical
composition of cows’ milk. In fact, milk contains less lactose (C12 H22 O11 ) than this, and some other
things besides water and lactose, but I have left out the other ingredients for the sake of simplicity.
132
than having a genuine disagreement about the nature of water. But if they
agree on the surface characterization of ‘water’ as the clear liquid that falls
from clouds (etc.), and then one says that the fundamental definition of
water is that it is H2 O, and the other says that the fundamental definition
of water is that it is a mixture of 90 per cent H2 O and 10 per cent
C12 H22 O11 , then there is no reason at all to think that they do not have
a genuine disagreement about the nature of water. In this second case,
we should not think that they are talking past one another; we should
think that at least one of them is just wrong about what water really is.
Now in the vagueness debate, all parties do indeed need to agree on the
surface characterization of ‘vagueness’, in order to avoid talking past one
another. But this requirement seems to sort itself out very smoothly in
practice, to the extent indeed that it hardly seems necessary to point it out:
those who are talking about something under the heading ‘vagueness’ that
does not admit of borderline cases, draw blurred boundaries, and generate
Sorites paradoxes, are in fact rapidly identified, and confusion is avoided.⁵
On the other hand, all parties do not need to agree on the fundamental
definition of vagueness. In fact, it is hard to see how that definition
could be theory-neutral, given that the different rival theories of vagueness
cannot all be correct. Whatever the fundamental nature of vagueness, it
surely cannot be compatible with what epistemicists, supervaluationists,
and degree theorists (for example) say about vague predicates, for these
theorists say incompatible things.
Given that the fundamental definition of vagueness does not have
to be compatible with all theories of vagueness, we cannot rule out a
proposed definition on the grounds that it conflicts with some existing
theory of vagueness. How, then, do we judge proposed definitions? What
should make us think that one fundamental definition of vagueness is
correct and another incorrect? The answer is simple. We should judge
a definition by its clarity and rigour, and by how satisfying and how
unified an account it yields of the various aspects of vagueness, including
(especially) the three properties that feature in the surface characterization
of vagueness.
⁵ For example, some writers in other disciplines—e.g. linguistics and engineering—sometimes talk
about other phenomena—e.g. lack of specificity or contextual reference-shifting—under the heading
‘vagueness’; but philosophers easily recognize (see e.g. Greenough 2003 n. 27) that these authors are
lumping distinct phenomena in with the one that we investigate under the heading ‘vagueness’.
133
The predicate ‘is schort’ has borderline cases: all persons between four and
six feet in height (inclusive). If asked whether such a person is schort,
we would react with a hedging response. It is quite unclear whether
‘x is schort’ is true or false—or neither, or something else—when x is
between four and six feet in height. Unlike the ordinary predicate ‘is short’,
however, ‘is schort’ doesn’t seem to be genuinely vague. This is because
while there are borderline cases for ‘is schort’, there is no unclarity about
where these borderline cases begin and end. The delineation between
the things which clearly are schort, the things which clearly are not, and
the borderline cases, is too sharp for ‘is schort’ to count as genuinely
⁶ Obviously the conjunction of the three—i.e. our surface characterization of vagueness—is too
motley a thing to serve itself as a fundamental definition.
⁷ Cf. Fine’s ‘‘nice1 ’’ (1997 [1975], 120), Soames’s ‘‘smidget’’ (1999, 164), and Tappenden’s ‘‘tung’’
(1993, 556).
134
vague. Thus, merely possessing borderline cases is not enough for true
vagueness.⁸
One possible response to this problem is to retain the definition according
to which a predicate is vague if and only if it has borderline cases, but
to posit an additional phenomenon called higher-order vagueness, over and
above (first-order) vagueness, and to say that ordinary vague predicates
are also higher-order vague, whereas ‘is schort’ is only first-order vague.
A predicate is first-order vague iff it has borderline cases; a predicate is
second-order vague if it has borderline cases of borderline cases (things
of which it is unclear whether the predicate applies to them or they are
borderline cases, and things of which it is unclear whether the predicate
does not apply to them or they are borderline cases); a predicate is third-
order vague if it has borderline cases of borderline cases of borderline
cases; and so on. I shall explain why I think this approach is mistaken
in §3.5.5.
There is another problem for the borderline case definition. If it is
to yield a clear and perspicuous account of vagueness, it must be sup-
plemented by a definition of borderline case (Bueno and Colyvan 2006,
3). We have been working with a rough and ready characterization
of borderline cases, according to which a borderline case of a pre-
dicate is a thing to which it is unclear whether or not the predicate
applies⁹—and where the presence of such unclarity is indicated by our
tendency to give a hedging response when asked whether the predic-
ate applies to the thing in question.¹⁰ This is fine as part of a trio of
informal characterizations which together delineate the class of vague
predicates. It would not be fine as part of a fundamental definition
of vague predicates as predicates which admit of borderline cases. For
clearly there are many precise predicates which have borderline cases
in our sense simply because we are ignorant about or uncertain as to
whether these predicates apply to certain objects. For example, when
asked whether the predicate ‘is travelling faster than any polar bear trav-
elled on 11th January 1904’ applies to the car we are driving in, we
would certainly react with a hedging response—and yet this predicate
is just as precise as the predicate ‘is travelling faster than 31 kilometres
⁸ This point has been made independently by a number of authors, including Sainsbury 1991, 173,
Keefe and Smith 1997a, 15, Fara 2000, 47–8, and Shapiro 2006, 1 n. 1; cf. also Eklund 2001, 376.
⁹ Cf. Williamson 1994, 2, and Keefe and Smith 1997a, 2. ¹⁰ Cf. Fara 1997, 81; 2000, 76.
135
per hour’. One way to remedy this situation would be to add to our
definition of a borderline case that the uncertainty involved must not be
a matter of ignorance (Peirce 1902; Black 1997 [1937], 71). Of course,
our definition of vagueness would then rule out the epistemic theory of
vagueness in advance—but I have already said that I do not think that a
definition of vagueness need be neutral between all possible theories of
vagueness. However the proposal does face a critical problem, which is
that it clearly cannot be stating the fundamental fact about vague predicates.
If we are told that vague predicates admit of cases where we are uncertain
whether the predicate applies, and where this uncertainty is not a matter
of ignorance, then we have been told only part of the story about what
vagueness is, supplemented by a story about what it is not. We immediately
want to ask, ‘‘Then of what is our uncertainty a matter? Why are we
uncertain?’’ It is only an answer to this further question that would get to
the heart of vagueness. Giving rise to borderline cases—in the sense under
discussion—can be regarded only as a symptom of vagueness, rather than
as constitutive of vagueness. We could remedy this problem by saying that
an object x is a borderline case of the predicate P if there is no fact of the
matter whether P applies to x (Sainsbury 1995), or if ‘x is P’ is neither true
nor false (Fine 1997 [1975]; Tappenden 1993), or if neither ‘x is P’ nor ‘x is
not P’ are true (Shapiro 2006, 2).¹¹ However, we would still be left with
the original problem, illustrated by ‘is schort’. I could have explicitly said,
when I introduced this predicate into the language, that it is neither true
nor false of things between four and six feet in height,¹² or that there is
no fact of the matter whether it applies to such things—rather than simply
remaining silent about these cases. Either way, ‘is schort’ still would not
be vague.¹³
¹¹ Cf. also Schiffer 2000. Again, this would rule out certain theories of vagueness in advance
(epistemicism and theories which say that vague predicates are both true and false of their borderline
cases), but I do not regard this as a problem per se.
¹² Cf. Sainsbury’s ‘child∗ ’ 1991, 173.
¹³ The issue of how to define ‘borderline case’ brings out a problem with Hyde’s 1994 defence of
the borderline case definition of vagueness against the ‘is schort’ problem. Hyde argues that the notion
of borderline case is itself a vague notion, and hence has borderline cases: thus, when we say that a
predicate is vague if it has borderline cases, we are not leaving open the possibility of a predicate which
is vague, and yet for which there is a sharp jump between the definite cases (or definite non-cases) and
the borderline cases. The problem with this approach becomes clear when we ask what ‘borderline
case’ could mean, on Hyde’s approach. Under the operational definition of a borderline case as a case
over which we hedge, clearly ‘borderline case of P’ will have borderline cases for many predicates P.
However, we have already seen the problems with plugging this definition of ‘borderline case’ into a
136
definition of vagueness. Under the definition of ‘borderline case of P’ as an object x such that ‘x is P’ is
neither true nor false, on the other hand, we can accept that there are borderline cases of borderline cases
only if we suppose that the language in which we describe the semantics of vague predicates is itself
vague. I argue in §6.2.1 that we should not go down this route. (For other objections to Hyde, see Tye
1997 [1994], who argues that the notion of a borderline case is not as vague as Hyde claims, and Varzi
2003a, who argues that there is a circularity in Hyde’s arguments. For replies to these objections, see
Hyde 2003.)
¹⁴ Frege was fully aware of the limitations of his metaphor: ‘‘this is admittedly a picture that may be
used only with caution’’ (Beaney 1997, 259).
¹⁵ Compare Johnston’s 1989, 1993 missing explanation argument against dispositional theories of
value and colour. Compare also my argument above against the Peirce/Black version of the borderline
case definition of vagueness.
¹⁶ Although see the end of §3.5.4 below for discussion of the view that not all vague predicates
give rise to Sorites paradoxes. Note that while this view is automatically ruled out by a strict wielding
of our surface characterization of vague predicates as those which admit of borderline cases, draw
blurred boundaries and generate Sorites paradoxes, it is not ruled out by a more subtle wielding of
that characterization which allows a few (but not too many) exceptional vague predicates, i.e. which
allows us to count a few predicates as vague if they do not generate Sorites paradoxes, but do admit of
borderline cases and do draw blurred boundaries, and are relevantly similar to some other predicates
which satisfy all three parts of the surface characterization.
¹⁷ There is another problem for this definition—paralleling the secondary problem for the borderline
case definition noted above—which is that it needs to be supplemented by a definition of Sorites
argument (Bueno and Colyvan 2006, 5).
137
where τ is a variable which ranges over the set of truth states {true, determ-
inately true, not true, not determinately true, determinately determinately
true, . . . }, α and β are variables which range over actual and counterfactual
cases, v is a function from actual and counterfactual cases to non-negative
real numbers, such that the truth of S depends only on the value of v,
c is some small positive real number, and Ks abbreviates ‘It is known by
speaker s that’.) Greenough recognizes that as stated, this definition will
¹⁸ Cf. Lewis (see the passages cited in §2.5.1) and Weatherson (see the passage quoted in §6.1.3); cf.
also Fine 1997 [1975], 120.
138
¹⁹ Note that this is a criticism of Greenough’s definition qua fundamental definition of vagueness; I
am not claiming that Greenough intended his account to amount to a fundamental definition, rather
than a surface characterization (see in particular Greenough 2003, 237 n. 2).
²⁰ Eklund 2005, 32–3 independently makes a similar objection to Greenough, and to the character-
ization of vagueness in terms of quandaries in Wright 2001.
139
3.3 Closeness
For any set S of objects, and any predicate F —vague or precise—a
competent user of F can discern relationships of closeness or nearness or
similarity amongst the members of S: closeness or nearness or similarity in
the respects that are relevant to—or determine—whether something is F
(for short, ‘F-relevant respects’). For example, consider the term ‘red’, and
the set of all visible objects in the room. As a competent user of the term
‘red’, you will automatically discern relationships of closeness or similarity
amongst these objects in red-relevant respects. For example, the orange
things in the room are closer in red-relevant respects to the red things than
are the green things. This is not to say that orange things are more red than
are green things: neither orange things nor green things are red at all.²³
Rather, it is to say that in the respects that determine whether something is red,
orange things are closer—or more similar—to red things than are green
things. Think of a colour wheel—or better, a colour solid—and imagine
locating each visible object in the room on the point of the colour wheel
that has the same colour as that object. The closeness relationships amongst
²² Another definition of vagueness which has been discussed—but not defended—in the literature
holds that a vague predicate is a predicate for which mathematical induction fails. This suggestion is
made, and convincingly criticized, by Bueno and Colyvan 2006, 4–5. Cf. Dummett 1997 [1975], 102–3.
²³ I mean here that ‘a is red’ and ‘a is green’ are both clearly false, when a denotes (say) a ripe orange.
It is also the case, of course, that (some) green things are not at all reddish, while (all) orange things
are reddish (and yellowish): for there are unique green hues, whereas all orange hues are binary (see e.g.
Hardin 1988, 39).
141
Blue pencils are not very close in blue-relevant respects—there are quite
a few pencils in between them in the tin—even though both are clearly
blue pencils.
Now consider a non-colour predicate, and indeed a non-vague predicate:
for example, ‘weighs over one kilogram’.²⁵ This term is likewise associated
with closeness relationships on the set of visible objects in the room—but
not the same closeness relationships as just considered. The black grand
piano in the corner and the black fragment of ash on the window sill are
very close in red-relevant respects, but not so close in respects relevant to
whether a thing weighs over one kilogram, while the blue and red pen
caps on the table are not very close in red-relevant respects, although they
are very close in respects relevant to whether a thing weighs over one
kilogram. Imagine locating every object in the room on a long line, with
an object weighing x kilograms being placed x metres from the beginning
of the line. The closeness relationships amongst these objects in respects
relevant to whether a thing weighs over one kilogram correspond to the
relationships of spatial closeness amongst them when they are located on
the line in this way.
The similarity relationships corresponding to other words are often not
as easy to visualize as the ones associated with colour terms and the ones
associated with the predicate ‘weighs over one kilogram’, but they are
always just as apparent to users of these terms. Consider, for example, the
predicate ‘heap’. A twenty-grain pile of sand is close to a twenty-one-grain
pile of sand in respects relevant to whether something is a heap; a pile
of ten olives is further away; but not as far away as an armchair, which
itself is reasonably close (in heap-relevant respects) to a dishwasher. (In the
respects relevant to whether something is a heap—e.g. size, the extent
to which the thing is composed of separate smaller things held together
only by gravity, shape²⁶—the armchair and the dishwasher are not very
different.) These things are apparent to any competent user of the predicate
‘heap’, even though there is no familiar object—such as a colour wheel
or a line—onto which we can map things in order to represent their
²⁵ I assume for the sake of example that weight terms such as ‘kilogram’, length terms such as ‘foot’
and ‘metre’, and other such terms, are precise—as opposed to predicates such as ‘heavy’ and ‘tall’,
which are vague. If you think that really, the former too are somewhat vague (see Russell 1997 [1923],
63), then please just treat them as precise for the purposes of the examples, or else substitute other
examples for mine.
²⁶ Cf. Hart 1992, 3.
143
²⁷ Of course, whether or not someone is nearly home depends upon how far away he was when
he set out: if Bill is travelling home from across the world, then when he is 10 kilometres away, he is
nearly home. Thus, when I said ‘‘Consider, for example, the predicate ‘nearly home’ ’’, I should have
said ‘‘Consider, for example, the predicate ‘nearly home’, as used on a particular occasion’’. Recall the
point about ‘large-scale’ contextual variation in §2.6.
144
²⁸ Here and in what follows I omit the superscript F on the relation symbol in order to indicate
F
generality of the schematic sort; i.e. I am here asserting transitivity for all relations ≤.
²⁹ Such a general theory would undoubtedly yield particular judgements about relative closeness
concerning which competent speakers have no intuitions. This does not cast doubt upon the idea that
there is a correct general theory to be given. Compare the case of the colour solids: not every instance
of the relation ‘a is closer to c than b is, in respect of colour’ that they yield is one which we can
perceive to hold. Nevertheless, we can perceive enough to ground a meaningful distinction between
correct and incorrect general theories.
145
³⁰ The ordinary notion of absolute similarity in F-relevant respects may well be vague, for many
predicates F. However, this would simply mean that we must regard the absolute closeness relations
which will figure in my account of vagueness as precisifications of their ordinary counterparts.
³¹ As already stated, I do not think that the task of giving a general theory which codifies our ideas
about closeness of objects in respects relevant to whether something is F —for a given predicate F —is
easy; nor do I think it is hard simply because it involves filling in a lot of tedious details, the general
framework in which the details are to be placed being clearly understood. But lest the reader suspect
that the task is in fact impossible, I suggest the following as a possible starting point in giving a general
theory of closeness of objects in respects relevant to whether a thing is, say, a table. First we determine
the relevant respects. Solidity, flatness of upper surface, and possession of legs, for example, are
relevant respects, while colour, monetary value, and location, for example, are not. Then we associate
each respect with a numerical scale. (Where this cannot yet be done—e.g. if one of the respects is
colour—we must analyse further, in order to reach respects which can indeed be associated with a
linear scale.) This gives us a vector space, each object corresponding to a vector whose coordinates
are the numbers with which the object is associated on each numerical scale. If we then define an
appropriate norm on this vector space, we get a metric space. Relative closeness relationships may then
be extracted via the idea that x is at least as close to z as y is, just in case the distance between x and z
is less than or equal to the distance between y and z. Absolute closeness relationships may be extracted
via the selection of a particular number d, the idea being that x and y are very close just in case the
distance between them is less than d. I stress that this is not intended as a glib dismissal of the difficulties
associated with constructing a general theory of closeness in, say, table-relevant respects: it is merely
intended as an antidote to the view that the very idea of constructing such a theory is absurd. Note also
that nothing in what follows depends upon our possessing such a general theory for any predicate F.
146
6 ft in height and that Ben is exactly 6 ft in height need not be—one might
be true and the other false (‘is exactly 6 ft in height’ is precise). In general,
if objects x and y are very similar in all respects relevant to whether or
not something is F, then if F is a vague predicate, the two claims that x is
F and that y is F must be very similar in respect of truth, while if F is a
precise predicate, they need not be. I call this the closeness picture of vague
predicates: closeness of x and y in F-relevant respects makes for closeness
of ‘Fx’ and ‘Fy’ in respect of truth.³²
One useful way to think of this is as a weakening of the idea that
vague predicates are tolerant. To say that F is tolerant is to say that very
small differences between objects in F-relevant respects never make any
difference to the application of the predicate F. Closeness says that very
small differences between objects in F-relevant respects never make a
big difference to the application of the predicate F —but (in contrast to
tolerance) they may make a very small difference.³³
I explain in §3.5 how my definition satisfies our desiderata for a correct
definition of vagueness. In the present section, I formulate and explain
the definition in detail. The final version of my proposed definition of
vagueness will be presented in §3.4.4. Here is a first version. A predicate F is
vague just in case it satisfies the following condition, for any objects a and b:
Closeness If a and b are very close in F-relevant respects, then ‘Fa’ and
‘Fb’ are very close in respect of truth.³⁴
³² Here, and in some other places in this book, I use ‘x’ (‘y’, ‘a’, ‘b’, etc.) both inside and outside
quotation marks, and sometimes in addition I quantify such expressions, in which case the quantifier(s)
bind occurrences of ‘x’ (‘y’, ‘a’, etc.) both inside and outside quotation marks. I regard such usage as
shorthand: it is sloppy, but I think it increases readability. Here is a key to interpreting the shorthand.
Shorthand: Closeness of x and y in F-relevant respects makes for closeness of ‘Fx’ and ‘Fy’ in respect
of truth. Longhand: Closeness of x and y in F-relevant respects makes for closeness of ‘F x’ and ‘F y’ in
respect of truth, where ‘x’ is a singular term which refers to x and ‘y’ is a singular term which refers to y.
NB: Furthermore, I often omit quotation marks, if it is obvious that an expression is being mentioned
rather than used.
³³ For more on the relationship between Closeness and Tolerance, see §3.5.1.
³⁴ Recall n. 32. This is shorthand for: (for any objects a and b, and any singular terms ‘a’ and ‘b’) if a
and b are very close in F-relevant respects, and ‘a’ refers to a and ‘b’ refers to b, then ‘F a’ and ‘F b’ are
very close in respect of truth.
147
about the former, but nothing so far about the latter. The first thing to
say is that it is a particular instance of the more general notion of closeness
in respect of a property: the instance in which the property is truth. Note
that there is a crucial difference between similarity in respects relevant to the
possession of a property P —that is, in P-relevant respects—and similarity
in respect of P itself. Consider, for example, Bill, who is 6 ft 4 in., and
Ben, who is 7 ft 4 in. Bill and Ben are not very close in respects relevant to
possession of the property tallness. After all, they differ in height by a full
foot—and this is a big height difference between persons. Nevertheless,
they are very close in respect of tallness itself: they both quite definitely possess
this property.
We need an understanding of closeness in respect of truth such that
two sentences may be close in respect of truth while not being identical in
respect of truth—or else Closeness will simply reduce to Tolerance. Fuzzy
semantic theory provides one such understanding: assuming that sentences
are assigned truth values in the interval [0, 1], we can say that two sentences
S and T are very close in respect of truth if |[S] − [T]| ≤ 0.01. Note that
this is not a definition of ‘close in respect of truth’, but a sufficient condition
for two sentences to be close in respect of truth. For a start, it might be
that sentences whose (fuzzy) truth values are further apart than this are still
close in respect of truth. Second—and more fundamentally—I take it that
the notion of two sentences being close in respect of truth is a (relatively)
primitive one which pre-exists—and indeed provides one of the central
motivations for—the formulation of the fuzzy semantic framework. Why
are degree theorists unsatisfied with the classical semantic picture? One
fundamental thought which they want to capture is that there are sentences
whose truth status lies between true and false, and furthermore (for the
previous thought does not get us further than the idea of three truth values,
or two values and a gap) that we can make comparisons within the class
of such intermediate sentences, potentially without limit (i.e. we can say
that this one is truer than that one, which is less true than this other one,
which itself is less true than the original one, which is nevertheless not
as true as this other one, and so on). But even that does not exhaust the
motivation of many degree theorists: for it gets us only to a set of degrees
of truth with a certain order structure, whereas it is quite natural to think
that there is, in addition, a meaningful notion of distance between truth
148
0
1 50,000 100,000
Men x
Figure 3.1. Ways of assigning truth values to sentences about baldness.
values.³⁵ One way to bring this out is to consider the assignment of degrees
of truth to sentences about the objects in a Sorites series—say, sentences
of the form ‘Man 1 is bald’, ‘Man 2 is bald’, . . . , ‘Man 100,000 is bald’,
where the numbering of the men in the series goes by the numbers of
hairs on their heads. Consider the systems of assignments of truth values to
these sentences shown in Figure 3.1 (each of the eleven curves represents
one system of assignments). If we think that the only thing that matters
about our set of degrees of truth is its order structure, then we will have
to think that all these systems represent exactly the same hypothesis about
the meaning of ‘is bald’—for the ordering of sentences with respect to how
true they are is uniform across the systems. It is quite natural to think,
however, that these systems of assignments represent different hypotheses
about the meaning of ‘is bald’. For example, one might naturally say
something along these lines: ‘‘If the system of assignments represented by
the straight line in the middle is correct, then the truth value of ‘This man
is bald’ changes uniformly between man 1 and man 100,000. If the system
of assignments represented by the curve on the right is correct, then at first
each hair removed hardly changes the truth value of ‘This man is bald’
at all—although it does change it a little bit—until we get to about man
95,000, whereafter each hair removed makes a big difference to the truth
value of ‘This man is bald’. If the system of assignments represented by
the curve on the left is correct, then at first each hair removed makes a big
difference to the truth value of ‘This man is bald’, until about man 5,000,
whereafter each hair removed hardly changes the truth value of ‘This man
is bald’ at all, although it does change it a little bit.’’
This second way of looking at the matter makes use of a notion of
distance between truth values—for example, in the idea of a uniform change
of truth value—and furthermore, it makes use of the idea that there is a
distance between truth values which is very small. This comes out in the
parts about single hairs making a ‘big difference’ (here the distance between
the truth values involved is not very small) or ‘hardly changing’ the truth
value (here the distance between the truth values involved is very small).
This notion of a very small difference between truth values hangs together
with the notion of closeness in respect of truth: where there is a big difference
in the truth values of two sentences, the sentences are not very close in
respect of truth; where there is hardly any change in truth value from one
sentence to another, the sentences are very close in respect of truth.
I am not saying that all degree theorists are motivated by the idea that
not only does the ordering of truth values matter, but also, there is a
meaningful notion of distance between truth values—and furthermore, a
meaningful notion of a distance between truth values which is very small
(this notion going hand in hand with the idea of two sentences being
very close in respect of truth). Some degree theorists are explicit that only
ordering matters to them.³⁶ On the other hand, the majority of degree
theorists have taken the real interval [0, 1] as their set of truth values,
without any caveat to the effect that they are only concerned with its
order structure, and hence that they consider any mapping of sentences
to truth values to be completely interchangeable with any other mapping
resulting from it by composing it with an order-preserving transformation
of the interval [0, 1]—and this indicates that they had in mind some metric
intuitions, as well as merely ordering intuitions, when they thought that
we should countenance degrees of truth. In any case, what I am saying is
that the second motivation for introducing degrees of truth makes sense.
That is, it makes sense to think that there could be predicates which differ
precisely in that an assignment something like the curve on the left in
Figure 3.1 is correct for one of them, while an assignment something like
³⁶ See in particular the many-valued systems of Post 1920, 1921 (Malinowski 1993, ch. 6, gives an
overview). Cf. also the passage from Goguen 1968–9 quoted on p. 297, and Goguen 1979, 59; Hájek
1999, 162–3; and Weatherson 2005, but note that he is not a degree theorist in my sense: ‘‘I do not
have the concept of an intermediate truth value in my theory’’ (p. 53).
150
the curve on the right is correct for the other.³⁷ I am also saying that
considering this second way of looking at things is one way to get a feel
for the notion of closeness in respect of truth that figures in the Closeness
definition.
At this point the suspicion may arise that there is going to be something
fishy about my proposed method of arguing that the correct theory of
vagueness must countenance degrees of truth. Is it not somehow question-
begging to argue for this conclusion via a definition of vagueness which
employs a notion (closeness in respect of truth) which, while admittedly not
defined in terms of degrees of truth, nevertheless figures in the motivation
of many degree theorists? No: there will be no question-begging. First,
I shall not assume that non-degree theorists cannot accept the Closeness
definition. Rather, I shall argue in Chapter 4 that, given the existence of
a Sorites series for the predicate F, one cannot accept the claim that F
conforms to Closeness unless one countenances degrees of truth. Second,
my argument for the correctness of the Closeness definition takes the
following form. In §3.5, I show that predicates which conformed to this
definition would appear to have just the features that the predicates we call
‘vague’ actually do have. This provides a prima facie reason for believing
that the predicates we call ‘vague’ do in fact have the same underlying
nature as predicates which conform to the Closeness definition—that is,
it provides a prima facie reason for thinking that Closeness is the correct
definition of vagueness. But only a prima facie reason: for there might
be another proposed definition which also has the feature that predicates
which conformed to it would appear to have just the features that the
³⁷ Weatherson (2006, 12) gives the following example: ‘‘Consider the predicate is very late for the
meeting. At least where I come from, a person who is roughly ten minutes late is a borderline case of
this predicate. But which side of ten minutes late they are matters. (In what follows I make some wild
guesses about how numerical degrees of truth, which aren’t part of my preferred theory, should operate.
But I think the guesses are defensible given the empirical data.) If Alice is nine and three-quarters
minutes late, and Bob is ten and a quarter minutes late, then the degree of truth of ‘Alice is very
late’ will be much smaller than the degree of truth of ‘Bob is very late’. The later you are the truer
‘you are very late’ gets, but crossing conventionally salient barriers like the ten minutes barrier matters
much more to the degree of truth than crossing other barriers like the nine minutes thirty-three
seconds barrier.’’ Now just as Americans and Australians, for example, have great fun discovering that
they mean different things by some terms (e.g. ‘sloppy joe’), we can imagine someone from another
country going to the place Weatherson comes from and saying ‘‘Wow, that’s interesting: where I come
from, it’s the twenty minute mark that really matters’’, and someone else saying ‘‘Really? Fascinating.
In my country, ‘you are very late’ just gets truer at a uniform rate as time passes—there are no
sudden increases.’’ My claim is simply that it makes sense to imagine these sorts of differences between
predicates.
151
⁴⁰ A basis (or base) for a topology T on S is a set B of open subsets of S such that every open subset
of S is a union of sets in B .
⁴¹ Recall that f : S → T is continuous just in case the pre-image of every open set in T is open in
S. If every subset of S is open, this will automatically be the case, regardless of what f and T are like.
⁴² Bringing in the two-place absolute similarity relation does not seem to help. If we stipulate that an
F
open set X must satisfy the condition that if x ∈ X and x ≈ y then y ∈ X, then whenever the domain
D consists of a Sorites series for F, the only open sets will be ∅ and D; i.e. we end up with the trivial
F
topology. If we stipulate that an open set X must satisfy the condition that for all x and y in X, x ≈ y,
then except in the special case where the domain consists of objects all of which are very similar to one
another in F-relevant respects, the open sets will not be closed under unions, and hence we do not get
a topology at all.
154
⁴³ I said that sometimes our similarity judgements arise from our discernment of an underlying metric
structure. In these cases the continuity proposal is on safe ground. For a metric yields a topology—or
we could simply work with the metric directly, using the metric space definition of continuity rather
than the more general topological definition.
⁴⁴ It would be no better simply to stipulate that every predicate is precise relative to a discrete
domain—for ‘is bald’ is vague relative to the discrete domain just considered, and ‘is large’ as a
predicate of natural numbers is vague.
155
small change in input produces at most a small change in the value of the
function), but not in the final official definition of continuity, which says
roughly that for any positive-sized target area in the codomain (whether or
not we would ordinarily regard it as ‘small’) we can find a positive-sized
launch area in the domain (which again need not be ‘small’ in the ordinary
absolute sense) such that everything sent by the function from that launch
area lands in that target area. The notion of absolute closeness of items in
the domain, and of values in the codomain, plays no role here. This is a
good thing in the context of a highly general mathematical definition, but
it is not a good thing in a definition of vagueness—for the reasons we
have seen (and see further n. 57 below). On the discrete domain of men
considered in the previous paragraph, the predicate ‘has 100 or less hairs
on his head’ comes out as precise by the Closeness definition, because its
characteristic function assigns True to man 100 and False to man 101, who
differ by just a hair. These two men are very close, in absolute terms, in
hair count, and hence in respects relevant to whether someone has n or
less hairs on his head (for any n). Yet our predicate is flat-out true of one
of them, and flat-out false of the other. That makes the predicate precise
by my definition—and that is clearly the right result, in contrast to the
continuity view, which, as already discussed, deems the predicate vague.⁴⁵
⁴⁵ Thanks to John Cusbert for helpful feedback on the material in this section.
⁴⁶ Thanks to Kit Fine for the first example, and to Brian Weatherson for the second.
156
§3.4.2. Finally, what about the idea that there might be vague objects?
I argue elsewhere that for there to be vague objects is for certain special
properties or relations to be vague: for example, the part–whole rela-
tion, the existence-at-a-world relation (which holds between an object
and the worlds at which it exists),⁵² and the occupation relation (which
holds between an object and the spacetime points it occupies).⁵³ There-
fore the Closeness definition applies, by extension, to vague objects:
their vagueness is a matter of the vagueness of certain properties and
relations, and the vagueness of these is handled by the Closeness def-
inition.
The Closeness definition does not cover everything to which the term
‘vague’ has ever been applied. I regard this as a virtue, not a vice. As
discussed in §3.1, a definition of vagueness should get to the essence of the
phenomenon that philosophers have been investigating under the heading
‘vagueness’. It should classify as vague something which has at some time
been called ‘vague’ by someone only if that thing is relevantly like central
examples of the core phenomenon under investigation. Weatherson (2006,
3, 11) objects to the Closeness definition on the grounds that it does not
seem to apply to predicate modifiers such as very—i.e. it does not classify
them as vague or as non-vague. Weatherson is not arguing that ‘very’ is
vague; but he does think that the question whether it is vague or not is
a good one, and that it is a requirement on a definition of vagueness that
it allow the question to be asked. Now at first sight it might indeed seem
as though ‘very’ could well be vague in the sense in which we are interested,
and so should be handled by my account. After all, isn’t ‘very tall’ vague
in the same sort of way as ‘tall’? Yes, the latter seems true—but it does
not imply the former. We have intuitions concerning the vagueness of
‘tall’, the vagueness of ‘very tall’, and perhaps the relationship between
these (‘very tall’ might perhaps seem to be less vague than ‘tall’). But do
we have any intuition concerning the vagueness of ‘very’, all by itself? I
don’t; contra Weatherson, I cannot even make clear sense of the question
whether ‘very’ (by itself) is vague. The genuine question in this area seems
to be: What in general is the relationship between the vagueness of F and
the vagueness of ‘very F’? The Closeness definition has the resources to
⁵² See Smith 2005a. ⁵³ But not the identity relation; see Smith 2008.
159
⁵⁴ That is, they have the same truth value, or they both lack a truth value. Note that where Wright
is concerned with applying predicates, Tolerance is concerned with the truth of sentences which express
such applications; this difference has no significance for what follows.
⁵⁵ If a Sorites series consists of objects x1 , . . . , xn , then Closeness tells us that Fxi and Fxi+1 must
always be very similar in respect of truth—which is quite compatible with Fx1 being true simpliciter
and Fxn being false simpliciter. Note, however, that given bivalence, Closeness reduces to Tolerance. So
Closeness plus bivalence plus a Sorites series leads to contradiction. This issue will be discussed in detail
in Ch. 4.
⁵⁶ I shall discuss the examples in Wright 1975, 331–8 (see also Forbes 1983, 239–41 for a hearty
endorsement of Wright’s views). Wright 1987, §VII introduces a number of qualifications, and draws
(somewhat) less far-reaching conclusions.
161
⁵⁷ Recall the discussion in §3.4.3 of the idea of defining a vague predicate as one whose characteristic
function is continuous. We can now see a further reason why the continuity proposal is inferior to
the Closeness proposal. The predicate ‘is six∗ feet in height’, whose characteristic function assigns 0 to
persons whose height differs from 6 ft by 1 nm or more, 1 to persons whose height is exactly 6 ft,
and in the intervals (6 − 1 nm, 6 ) and (6 , 6 + 1 nm) changes smoothly from 0 to 1 and from 1 to 0
respectively—and hence is vague according to the continuity definition, but non-vague according to
the Closeness definition—is just as unusable from the point of view of casual observation as the precise
predicate ‘is exactly six feet in height’. The fact—which presumably underlies their ubiquity—that
vague predicates are essentially useful in contexts of rough and ready observation is thus intimately
linked to the inclusion in their definition of an absolute notion of similarity (where two things are very
similar in some respect precisely if they are not distinguished in that respect in contexts of casual observation)
which would be out of place in a highly general, abstract definition of a mathematical concept such as
continuity.
163
given set of objects. Children, for example, will quite naturally sort their
coloured pencils into a rainbow pattern, without instruction, and without
an actual rainbow pattern to copy: they can just see that red and orange
are closer together than red and green (that is, they are closer together in
respect of colour, and hence belong closer together in space, when the
pencils are sorted in their tin). Now, the ostensive teaching of colour words
can proceed by the indication of paradigms and foils for each colour, with
the (implicit) instruction that ‘red’, for example, applies to an object in
proportion to its similarity to the red paradigms. Thus, if an object is very
close to a red paradigm, it is red, or as good as for all practical purposes;
if it is very close to an orange paradigm, it is orange, or as good as for all
practical purposes; and if it is somewhere between red and orange, then
it is to some extent red and to some extent orange. This account of the
learning of colour terms in fact fits perfectly with our actual practice with
these terms. We do not know how to describe the colour of every object
we see, just as the story I have told would predict. The instructions do not
fix a particular way of describing every object we might come across: they
fix what to say about things very close to a paradigm, but what to say about
other objects is more open—although it is not totally open. Things would
not work the way they actually do if an undetectable change from a red
paradigm could make it false to say ‘That’s red’; but things would work
exactly as they do now if such negligible changes made it negligibly less true
to say ‘That’s red’.
Thus, predicates which satisfy Closeness (but not Tolerance) would
exhibit precisely the features Wright notes. Wright’s examples do not,
then, establish that vague predicates are tolerant: the examples are just
as compatible with the view that vague predicates are intolerant (small
changes need not make no difference) but conform to Closeness (small
changes never make a big difference). Given also that we believe that
some things are red and some not, that some men are bald and some
not, and so on, one conclusion we might draw is that vague predicates
are tolerant, and subject to inconsistent requirements of use. Of course,
the more reasonable conclusion to draw is that vague predicates are not
subject to inconsistent requirements: they do not conform to Tolerance;
they simply conform to Closeness. Accepting Closeness takes us far enough
along the road to Tolerance to capture the intuitions which Wright
uses to motivate Tolerance, without taking us so far as to run into
164
The problem with this argument is that Wright conflates precise matching
and indistinguishability. Certainly, if colour predicates are observational, then
they must apply equally to objects which are indistinguishable by ordinary
observational means. They need not, however, apply equally to objects
which precisely match. This is because objects which precisely match (i.e.
which cannot be distinguished in colour on the basis of direct comparison)
need not be observationally indistinguishable. For ordinary observational
means of distinguishing two objects extend beyond direct comparison of
them: one can compare each with a third object, compare them under
different coloured lights, and so on.
⁵⁸ Wright’s argument here has roots in Dummett 1997 [1975] and in Russell 1997 [1923], 64.
165
What about the predicate ‘looks red’, as opposed to ‘is red’? Here there
are two responses that we might make. First, there is the response that
parallels the one given in the case of ‘is red’. Just as ‘is red’ need not apply
to exactly the same extent to objects which match precisely, if these objects
are distinguishable by ordinary observational means, so too for ‘looks red’.
But this option is not quite as attractive in the case of ‘looks red’: it is more
natural to say that adjacent patches in the Sorites series are ever so slightly
different in respect of redness than it is to say that they look ever so slightly
different. However, there is a second possible response, based on Raffman
(2000). We need not deny that if two objects match, then when they are
viewed side by side in direct comparison, they look the same (in respect of
colour). Rather, we claim that whether or not two things look the same
(and likewise, whether or not they look red) depends upon the context in
which they are viewed. Viewed one pair at a time, adjacent patches in the
Sorites series look the same, and ‘looks red’ applies to each to exactly the
same extent. Viewed together in series, the first and last patches are easily
distinguishable in colour: the first patch looks red and the last patch looks
orange. No problem: for when viewed in the context of the whole series, it
is not the case that both members of every adjacent pair look the same, and
it is not the case that ‘looks red’ applies to each to exactly the same extent.
Do not try to focus on two adjacent patches that do not look exactly alike,
however: as soon as you look at the pair alone, its two component patches
do look exactly the same as one another. Whilst it is not plausible to suggest
that the context of viewing alone might affect the colour of an object, it is
plausible to suggest that the context of viewing affects the look of an object:
the look of a thing depends not only on the lighting and so on, but upon
what else is in the visual field when the thing is observed. Thus, the fact
that adjacent patches look the same when viewed alone provides no basis
for concluding that if the first patch in the series looks red, then so does
the last.
practical purposes we can just say that b is a heap too. On the Tolerance
reading, the second premiss expresses (a consequence of) the claim that
‘heap’ (or in general F) conforms to Tolerance. Two piles of sand a and
b which differ by just a grain are very similar in heap-relevant respects;
thus the two claims ‘a is a heap’ and ‘b is a heap’ must be exactly the same
in respect of truth. So if a is a heap, then b is a heap too: not just for all
practical purposes, but without qualification. On the Tolerance reading,
the conclusion follows from the premisses. On the Closeness reading, it
does not: each successive statement ‘this is a heap’ (said after removing one
grain) must be very similar in respect of truth to the one before, but need
not be exactly the same in respect of truth; and so by the end, the final
statement may be simply false.
Suppose we accept that a predicate F satisfies Closeness, but not Toler-
ance. The very fact that we accept Closeness will mean that in many ordinary
circumstances, we act as if we believe Tolerance. For Closeness tells us
that a negligible or insignificant difference between a and b in F-relevant
respects makes for at most a negligible or insignificant difference between
Fa and Fb in respect of truth—and a negligible or insignificant difference
is one which we are entitled to ignore for practical purposes. So for
practical purposes, when there is a negligible or insignificant difference
between a and b in F-relevant respects, we will simply ignore any differ-
ence between Fa and Fb in respect of truth, and so treat them as being
identical in respect of truth. This practice is licensed by our acceptance of Close-
ness. This explains why someone who accepts Closeness, but not Tolerance,
will find the Sorites paradox compelling. Of course, she accepts the first
premiss—everyone does.⁶⁰ She accepts the second premiss—taken as an
expression of Tolerance—because, as we have seen, someone who accepts
Closeness will in ordinary circumstances regard Tolerance as an acceptable
(and indeed useful: we needlessly clutter up our thought if we do not
ignore negligible things) approximation of what she accepts. But then the
unacceptable conclusion follows: for the argument is valid when the second
premiss is read as expressing Tolerance. So that is why the paradox has
force. But we can also explain why someone who accepts Closeness, but
not Tolerance, will regard the paradox as ultimately mistaken, even though
initially compelling. Tolerance is an acceptable approximation of Closeness
⁶¹ Thanks to John Cusbert for first suggesting a similar analogy in this context, and for helpful
feedback on the material in this section.
170
⁶³ Note that if a predicate F is vague according to the Closeness definition, then there is an
F-connected, F-diverse set S of objects such that F satisfies Closeness over S. But an F-connected,
F-diverse set S of objects does not automatically yield a Sorites series for F. An F-diverse set S is one
for which it is not the case that for every a and b in S, Fa and Fb are very similar in respect of truth.
Thus an F-diverse set need not contain an object which is clearly F and an object which is clearly not
F —whereas a Sorites series begins with an object which is clearly F and ends with an object which is
clearly not F.
173
⁶⁴ For a similar view see Keefe 2000, 31. Cf. also Burns 1995, 29.
174
predicates which are vague but not higher-order vague (on the borderline
case conception) are not vague at all (in the ordinary sense).⁶⁵
Another advantage of the Closeness definition of vagueness is, then,
that it incorporates the residual intuitions which the borderline case
definition leaves out, and which lead—when vagueness is defined in
terms of possession of borderline cases—to the problem of higher-order
vagueness. The Closeness definition thus gives us a clear understanding
of the intimate relationship between vagueness and so-called higher-
order vagueness: properly understood, the phenomenon of higher-order
vagueness is simply part of vagueness.⁶⁶
⁶⁵ Cf. Sainsbury 1991, who rejects the borderline case definition and characterizes vagueness in
terms of what Sainsbury calls ‘boundarylessness’. He writes: ‘‘The phenomena which, from a classical
viewpoint, lead to notions of ‘higher-order vagueness’ are accounted for by boundarylessness. But these
phenomena are not bolt-on options; they are integral to the very nature of vagueness’’ (p. 179).
⁶⁶ Note that a predicate which satisfies Closeness is not necessarily radically higher-order vague, i.e.
vaguen for all n (see p. 189 for further discussion of this point). Thus I am not forced to disagree (or
agree) with Burgess 1990, 1998, who argues that higher-order vagueness peters out at a finite level.
Nor am I required to rebut Wright’s 1992 argument to the conclusion that higher-order vagueness
is paradoxical (for replies to Wright see Edgington 1993 and Heck 1993). This is because Wright’s
argument turns on a certain principle concerning the logic of the definitely operator. Higher-order
vagueness has traditionally been discussed in terms of such an operator—but the Closeness definition
involves no such operator, and hence, while it accommodates the intuitions which motivate the
conception of ‘higher-order vagueness’ discussed in this section, it does not inherit the problems
associated with this operator.
4
Accommodating Vagueness
4.1 Epistemicism
No matter how sophisticated her epistemology, as long as her semantics is
classical, the epistemicist cannot accommodate predicates which satisfy the
Closeness definition—that is, given the correctness of the Closeness defin-
ition, she cannot allow that there are any vague predicates. Suppose that we
have a Sorites series x1 , . . . , xn for the predicate F. If bivalence—thought
of as the claim that not only are there only two truth values, but also that
every declarative sentence has exactly one of them—is true, then Closeness
must be violated here. Fx1 is true. x1 and x2 are very similar in F-relevant
respects, so if F conforms to Closeness, then Fx1 and Fx2 must be very
176
False
k
Points x on the strip (red points at left)
Figure 4.1. The strip according to the epistemicist.
similar in respect of truth. Given bivalence, the only way this can be the
case is if Fx2 is true too. Similarly on down the series. But by the end,
when we get to xn , it must be the case that Fxn is false. Thus if Closeness
is to be respected, we need to abandon bivalence. Another way to make
the point is to note that given bivalence, Closeness reduces to Tolerance:
the only way in which two sentences can be very similar in respect of
truth is by having the same truth value. So to capture Closeness without
Tolerance—and hence without incoherence—we must reject bivalence.
Before considering possible epistemicist responses to this objection, let
us consider a concrete example. (We shall return to this example below;
it will help us to focus ideas as we discuss the various different theories of
vagueness.) Consider a strip of paper which is red at the left end and orange
at the right end, and in between changes colour continuously from red to
orange (so that points on the strip which are very close together in space
are very similar in respect of colour). Now consider the sentence ‘Point x
is red’ for each point x on the strip. For a point near the left end of the
strip, the corresponding sentence is straightforwardly true, and for a point
near the right end, the corresponding sentence is straightforwardly false.
For points in the middle, the truth values of the corresponding sentences
are unclear. According to the epistemicist, this unclarity is purely epistemic:
every one of the sentences is either true or false. Figure 4.1 represents the
situation according to the epistemicist. This flouts Closeness: the classical
truth values True and False are not close (in respect of truth), and yet
according to the classical semantic theory which underlies the epistemic
theory of vagueness, there are points k (shown on the diagram) and k +
177
(a point to the right of, and arbitrarily close to, k) whose colours are
very close together, while the sentence ‘Point k is red’ is True and the
sentence ‘Point k + is red’ is False. This sort of flouting of Closeness is
unavoidable for the epistemic theory.
This objection to epistemicism is often phrased as follows: Contrary to
what the epistemic theory tells us, there is no last point on the strip which
is red. This is an unfortunate way of phrasing a good objection.¹ Point k is
indeed the last point which is red (or else k + is the first point which is
not red), according to the epistemicist. But this is not the problem with the
epistemic account. On the contrary: if there were no last red point (and no
first non-red point), then all the points in the series would be red—even
the points at the end, which are clearly orange. This is absurd, and thus it is
in fact the denial of the claim that there is a last red point or a first non-red
point that is problematic. The real problem with the epistemic picture is
that point k is a jump point of the characteristic function of the predicate
‘is red’ (the function which assigns to each point x on the strip the truth
value of the claim ‘Point x is red’)—i.e. a point at which the value of the
function suddenly jumps significantly—while k is not a location at which
adjacent points on the strip have significantly different colours. Thus, the
idea that if points x and y are very close in respect of colour, then the
sentences ‘Point x is red’ and ‘Point y is red’ should be very close in respect
of truth, is violated.
The epistemicist might respond to this objection in several ways. First,
she might note that I have been assuming that there are Sorites series
associated with the vague predicates under discussion. If, however, there
is no Sorites series for the predicate F, then my demonstration that the
epistemicist account of the meaning of F violates Closeness does not go
through. This would not get the epistemicist very far, however. For even
if there are some vague predicates with no associated Sorites series, these
are a special case (cf. §3.5.4). Certainly, the epistemicist cannot allow that
the standard examples of vague predicates—for which there are associated
Sorites series—are vague (i.e. satisfy Closeness), and this is problem enough.
Second, the epistemicist might object to the use of a continuum of points
in the example of the coloured strip. She might attempt to argue that in
¹ I am being charitable. Some people who say this about epistemicism may well be phrasing a bad
objection accurately!
178
False
k
Patches x (red patches at left)
Figure 4.2. The patches according to the epistemicist.
all actual cases, we deal with finite, or at least countable, sets of objects.
Setting aside the question of whether this is plausible, it would not help
the epistemicist if it were true. Consider a finite series of distinct uniformly
coloured patches (rather than a continuous strip of paper), with each patch
very similar in colour to the next—but not indistinguishable from it—and
the first patch red and the last patch orange (Fig. 4.2). Given certain natural
assumptions discussed in §3.4.3, the epistemicist can correctly claim that
the characteristic function of ‘is red’ in this situation is continuous. Yet while
point k is not now a point of discontinuity of the characteristic function of
the predicate ‘is red’ (as it was in the case of the continuous strip), it is still
a jump point —that is, a point at which the value of the function suddenly
jumps significantly—while k is not a point at which adjacent patches have
significantly different colours. Thus Closeness is still violated. (Recall the
discussion in the second half of §3.4.3, where I explicitly took cases of
this sort as reason to define vagueness in terms of Closeness rather than
continuity.)
The epistemicist’s next move might be to argue that her framework can
accommodate vague predicates in the following way. Closeness is satisfied
by constant characteristic functions: functions which map every object to
one value, in this case to True or to False. However, saying that vague
predicates are true (or false) of everything is absurd. One of our most basic
intuitions—one which contributes crucially to making Sorites paradoxes
compelling—is precisely that some things are heaps and some are not,
that some things are tall and some are not, that some things are red and
some are not, and so on. The current proposal retains the intuitive idea
179
that one grain of sand never makes the difference between a full-fledged
heap and a full-fledged non-heap, but only by saying either that everything
is a full-fledged heap or that nothing is—that is, at the expense of denying
either the claim that one grain of sand is not enough to make a heap or the
claim that 10,000 grains of sand is enough to make a heap. In any case, the
proposal falls foul of the final version of the Closeness definition in §3.4.4: a
predicate F is vague iff there is some F-connected, F-diverse set S of objects
such that F satisfies Closeness over S. If the characteristic function of F is
constant, then there are no F-diverse sets, and so F is not vague according
to the Closeness definition. (Indeed, part of the point of modifying the
initial version of the Closeness definition was precisely to avoid classifying
predicates which are true, or false, of everything as vague.)
As her next move, the epistemicist might claim that for any vague
predicate F, the structure of relationships of similarity in F-relevant respects
is impoverished, in the following ways. First, the ternary relation ‘x is at
least as close to z as y is, in F-relevant respects’ divides any set of objects
into just two subsets: the F’s and the non-F’s. The members of any one
of these two subsets are indistinguishable in terms of this relation: for any
two objects v and w in one of these subsets, the relation holds with v in
one of its three places (and objects a and b in the other two places) if and
only if it holds with w in that place (and a and b in the other two places).
The relation distinguishes v and w only if v is in one subset and w is in
the other. Second—and more important in the present context—x is very
close to y in F-relevant respects if and only if x and y are in the same subset,
i.e. if and only if x and y are both F’s, or both non-F’s. The latter move
means that epistemicism no longer violates Closeness: although it posits a
jump in a Sorites series between the last F and the first non-F —it is true
of the former that it is F, and false of the latter—this is now compatible
with Closeness, because these two objects are not very close in F-relevant
respects.
Under this proposal, whether or not something is F becomes an F-
relevant respect—that is, a respect relevant to whether or not that thing
is F. This is something I earlier ruled out (see Ch. 3, n. 47)—for it leads
to circularity and ungroundedness in the application of the predicate F to
any object. Even apart from leading to this undesirable result, however,
the proposal under consideration can be seen to be untenable by direct
inspection of what it says about familiar cases. Consider piles of sand, for
180
example. The proposal tells us that as we remove grains from a large pile
of sand, we make no difference at all to the pile, in any respect relevant
to the application of the predicate ‘heap’—until we remove the single
grain which renders the heap a non-heap, at which point we make a
big difference to the pile both in respects relevant to the application of
‘heap’ and in respect of application of the predicate ‘heap’. This is absurd.
Removing a grain of sand from a heap does make a difference to it, in
a respect relevant to the application of the predicate ‘heap’. It might not
make it a non-heap—that is, it might not stop the predicate applying.
Indeed, it might not affect the application of the predicate at all. But it
does change the heap in a respect that is relevant to the application of the
predicate. That, when we think about it, is undeniable. Indeed, if it was
not the case, then Sorites paradoxes would never be compelling: for the
paradox turns precisely on having a series of objects in which adjacent
members—every pair of adjacent members—are very similar in respects
relevant to the application of the predicate in question.²
Next, rather than try to accommodate it, the epistemicist might claim
that the Closeness definition of vagueness is (subtly) incorrect. Of course,
this reply is no good by itself: she must propose an alternative definition. I
have defined vagueness in terms of the following principle:
Closeness If a and b are very similar in F-relevant respects, then ‘Fa’ and
‘Fb’ are very similar in respect of truth.
The epistemicist might claim that the true idea to be captured is somewhat
different:
JA-Closeness If a and b are very similar in F-relevant respects, then
‘Fa’ and ‘Fb’ are very similar in respect of justified assertibility (i.e. to
whatever extent it is justifiable (reasonable, warranted . . . ) to assert
(believe, judge . . . ) that Fa, it is to a very similar extent justifiable to
assert that Fb).
² Note that I am ruling out as absurd and untenable the denial of (A): in a Sorites series, every object
is very similar to its neighbour(s) in respects relevant to the application of the predicate in question. I
am not ruling out as absurd and untenable the denial of (B): in a Sorites series, every object is very similar
to its neighbour(s) in respect of application of the predicate in question. To say that any two objects
which differ by just one grain of sand are very similar in heap-relevant respects is not to say that there
is no grain of sand that takes us all the way from full-fledged heaps to full-fledged non-heaps. I am, in
effect, in the process of arguing in this chapter that (B) must be true, on the grounds that given (A)
(which, I have just been saying, is obviously true), denying (B) means denying that vague predicates
satisfy Closeness; but I do not regard denying (B) as absurd.
181
Thus the epistemicist’s claim would be that the true idea to be captured by
a theory of vagueness is an epistemic/pragmatic one, rather than—as the
Closeness definition would have it—an alethic/metaphysical one.
As things currently stand in the literature, epistemicists have not shown
that their theories can accommodate predicates which satisfy JA-Closeness.
This does, however, seem plausible. Williamson (1994, 244–7) argues for
the following principle:
If it is reasonable to believe that n grains make a heap, then it is not
reasonable to believe that n − 1 grains do not make a heap.
This is not yet JA-Closeness. But let us set aside the issue of how good
Williamson’s arguments for this principle are, and the issue of getting
from this principle to JA-Closeness (or of getting to JA-Closeness in
some other way), and grant for the sake of argument that the epistemicist
can accommodate JA-Closeness. The point that I wish to stress is that
JA-Closeness does not yield an adequate definition of vagueness.
There are at least two problems with defining vagueness in terms of
JA-Closeness: this definition does not explain why vague predicates draw
blurred boundaries, or why they are Sorites-susceptible. Recall that a
definition of vagueness is supposed to provide a statement of the essence
of vagueness, in light of which we can understand why vague predicates
behave in the characteristic ways they do. Defining vagueness in terms
of Closeness meets this desideratum, as we saw in the previous chapter.
Defining vagueness in terms of JA-Closeness does not: a predicate which
satisfied JA-Closeness would not thereby draw blurred boundaries, or
generate a compelling Sorites paradox.
Consider first the point about blurred boundaries. Suppose that we
have a Sorites series for F, and someone asserts that Fxi and not Fxi+1 ,
where xi and xi+1 are adjacent items in the series. Intuitively there
is a problem here. The epistemicist now tells us that this person has
violated epistemic or pragmatic norms: she has unjustifiable beliefs, or she
has asserted something which she does not believe. But the problem seems
deeper: there is a problem with the very idea that there is a sharp jump
between the F’s and the non-F’s, not merely with the idea that someone
might believe or assert that the jump is in this or that particular place.
Someone who asserts that such a jump does exist, in such-and-such place,
seems not merely to have violated epistemic or pragmatic norms, but to
182
have violated the norm of truth. This comes out clearly if we imagine
someone guessing that the jump between the F’s and the non-F’s comes
between xi and xi+1 . Someone who makes such a guess seems to have
misunderstood the nature of vagueness just as much as someone who makes
the corresponding assertion. Yet the guesser—unlike the asserter—has not
violated JA-Closeness. Thus JA-Closeness does not capture the ordinary
conception of vagueness. Epistemicism plus JA-Closeness yields a view on
which the extensions of vague predicates are sharp in themselves, but we
cannot know or justifiably believe that their borders are in this or that
particular place—and so they appear blurry to us. But the problem with the
above guess seems to be that it could not be true: because vague predicates
have genuinely blurry boundaries—blurry in themselves—not sharp but
unknowable ones. Phenomenologically, there is an enormous difference
between a proposition such as ‘The jump from baldness to non-baldness
comes at the four hundredth hair’ and a proposition such as ‘The least upper
bound of velocities reached by polar bears on 11th January 2004 was 31.35
kilometres per hour’, which may well be true—and could certainly be
guessed to be true, if one wanted—but should not be asserted or accepted,
as no one could justifiably believe it to be true. Satisfaction of JA-Closeness
is not, then, of the essence of vagueness.
In contrast to the foregoing, Greenough (2003, 272–4) writes: ‘‘Argu-
ably, our naïve intuitions concerning vagueness are not sophisticated
enough to make the distinction between tolerance and epistemic tolerance,
between lacking sharp boundaries and lacking known boundaries . . . argua-
bly the phenomenological data merely supports a thesis of epistemic
tolerance and not a thesis of boundarylessness: close inspection simply
shows that there is no clear or known boundary . . . not that there is no
boundary.’’ But if there is indeed an argument to back up Greenough’s
position here, it would seem that the argument must rest on a conflation
of the content of an utterance and its assertibility conditions. Suppose a
witness tells us that there was no one there. This is incompatible with
the claim that there was someone there. The barrister responds ‘‘Ah, you
are saying that from where you were positioned you could not see any-
body’’—a claim which is compatible with there having been someone
there. This is clearly an insidious move: the witness said what he said
because, from where he was positioned, he could not see anybody; but
this is not what he said, which is that there was nobody there. Likewise in
183
the case of vagueness. We can all agree that the ordinary speaker says that
there is no sharp jump from red to orange because she cannot see a sharp
jump; but it is illegitimate to move from this to the claim that what she
says is that there is no clearly visible jump from red to orange. What any
ordinary speaker, when faced (say) with an unobscured, not-too-distant
rainbow—and whilst wearing her glasses or contact lenses and not under
the influence of drugs or alcohol, etc.—would say is not the absurdly
timid ‘‘I cannot perceive a clear boundary between red and orange’’,
but the stronger claim that there is no sharp boundary. This should not
be controversial. If ordinary speakers made only the weaker claim, then
epistemicism would not have been met with howls of disbelief from
both its opponents and some of its proponents³ when it first came onto
the philosophical scene, and furthermore Frege—to whom the blurred
boundaries idea can be traced—would not have thought that vagueness
has no place in a logically perfect language (i.e. a language with, among
other things, a bivalent semantics). (A reminder about the dialectic here: I
am not objecting to epistemicism on the grounds that it is counter-intuitive;
I am objecting to the definition of vagueness in terms of JA-Closeness,
on the grounds that it does not capture an essential feature of the ordin-
ary conception of vagueness: namely, that vague predicates draw blurred
boundaries.)
Second, the point about Sorites-susceptibility. Suppose that we have
a predicate F which conforms to JA-Closeness, but not Closeness. (If a
predicate conforms to Closeness, then it will also conform to JA-Closeness,
given some very natural assumptions. We know that a predicate which
conforms to Closeness will be Sorites-susceptible. What we want to know
now is whether JA-Closeness has the advantages of Closeness. Thus we
need to suppose JA-Closeness by itself.) Suppose that we have a Sorites series
for F. Will the corresponding Sorites paradox be compelling? We have no
reason to think so. Given JA-Closeness, and the fact that adjacent objects
in the Sorites series are very close in F-relevant respects, we know that
for any object in the series, to whatever extent it is justifiable, reasonable,
or warranted to assert, believe, or judge that it is F, it is to a very similar
extent justifiable, reasonable, or warranted to assert, believe, or judge that
³ See e.g. Williamson 1994, p. xi: ‘‘This book originated in my attempts to refute its main
thesis. . . . For years I took this epistemic view of vagueness to be obviously false, as most philoso-
phers do.’’
184
the next object is F. But we have no reason at all to conclude from this
that our Sorites premiss is true, i.e. that for any object in the series, if it is
F, then so is the next object. Whether or not this is compelling depends
crucially on why we think JA-Closeness holds for F. If it holds because F
satisfies Closeness, then the Sorites premiss does become compelling, as we
have seen—but as we have also seen, this is of no help to the epistemicist in
the present context. If, on the other hand, F satisfies JA-Closeness because
F works in the way in which the epistemicist thinks vague predicates
work—i.e. it draws sharp but unknowable boundaries—then obviously
we have no reason at all to accept the Sorites premiss, and so the Sorites
argument will not be compelling in the slightest. But this shows that
JA-Closeness—by itself—yields no account of Sorites-susceptibility. If we
define vague predicates in terms of JA-Closeness, we are left with no
understanding of why they generate Sorites paradoxes.
What about the epistemicist’s proposed solution of the Sorites paradox
(§2.1)? The idea was that the paradox is mistaken because the inductive
premiss is false—there is a sharp cut-off between the F’s and the non-
F’s—and that we are nevertheless taken in by the paradox because our
ignorance of where the cut-off is makes us think that there is no cut-off
at all: that is why we are inclined to accept the inductive premiss, even
though, according to the epistemicist, it is false. We can see now that this
solution is not robust, in the following sense: it provides an account of how
the paradox could be compelling and yet mistaken which is compatible
with the truth of epistemicism; but it does not provide an account of how
the paradox could be compelling to someone who had come to accept
epistemicism. If we become epistemicists, we should no longer find the
paradox compelling at all.⁴ It is therefore part of the epistemicist’s account
that we were taken in by the paradox only because we did not realize
how vague predicates work. We must, then, have thought they worked
differently from the way the epistemicist says they work. So how did we
think they work? Well, not according to JA-Closeness: we did not think that
JA-Closeness was the defining feature of vague predicates, because if we
⁴ If we define vagueness in terms of Closeness, on the other hand, we have a robust explanation of
Sorites-susceptibility: for we saw that accepting Closeness licenses being taken in by the paradox (and
also finding it ultimately unconvincing). I think this gets the phenomenology of the Sorites paradoxes
right: their pull is more resilient than the epistemicist can allow. Thanks to Mark Colyvan for helpful
discussion here.
185
had thought that, we would not—as we have just seen—have found the
Sorites paradox compelling. This brings us back to the point that the
definition of vagueness in terms of JA-Closeness does not capture a second
essential feature of the ordinary conception of vagueness: namely, that
vague predicates generate compelling Sorites paradoxes.
Finally, in light of the problems with the JA-Closeness definition of
vagueness, the epistemicist might try to repackage her view as as error
theory of vagueness. The idea would be that the correct way to define
vagueness is indeed in terms of Closeness; but in fact there are no predicates
which satisfy Closeness—i.e. there are no vague predicates. There are only
predicates which satisfy JA-Closeness, and of which the epistemicist theory
is correct—i.e. predicates which draw sharp but unknowable boundaries.
Thus the view is not that ‘tall’, ‘bald’, etc. are vague and work the way the
epistemicist says they do—vagueness being characterized not by Closeness
but in some other way. That line of approach has just been discussed,
and rejected. The current approach is to agree that vagueness is correctly
defined in terms of Closeness, and yet to maintain that ‘tall’, ‘bald’, etc.
do not satisfy Closeness—for they work the way the epistemicist says they
do—and hence are not vague.
Why would anyone adhere to such a view? The view that Closeness is of
the essence of vagueness is a result, in part, of reflection on intuitions about
canonical examples of vague predicates, such as ‘tall’ and ‘bald’. Turning
around at the end of such reflection and saying that these canonical examples
are in fact not vague at all—indeed, that no predicates are vague—thus
leaves one in an uneasy position. The only legitimate reason for believing
such a view would be that the alternatives—the theories of vagueness
which do accommodate Closeness—are no good.⁵ For example, one might
have general reasons for thinking that no non-classical semantic theory
is acceptable—and to the extent that these reasons were good ones, this
would mitigate the unease of the error theory under consideration. The
key point is that one could not adequately defend this sort of error theory
without attacking its rivals: for if there is an alternative view which is
quite acceptable in itself and which allows us to accommodate both the
view that vagueness is correctly defined in terms of Closeness and the view
that apparently canonical examples of vague predicates really are vague,
then this will, by default, be more attractive than a view which accepts
that vagueness is correctly defined in terms of Closeness, but denies that
apparently canonical examples of vague predicates are vague. The way to
argue against an epistemicist error theory of the sort under consideration is,
then, to argue in favour of an alternative theory of vagueness which does
accommodate Closeness. I turn to the task of defending such an alternative
theory in Part III.
⁶ These are separate points: two cats in opposite corners of a small room are as far apart as can be,
but they are not very far apart.
187
False
l1 l2
Points x on the strip (red points at left)
Figure 4.3. The strip according to the gap view.
Truth value of ‘Point x is red’
True
False
l1 l2
Points x on the strip (red points at left)
Figure 4.4. The strip according to the third value view.
gap view, and Figure 4.4 represents the situation according to the third
value view. According to these views, point l1 is the last point which
is clearly red, and point l2 is the first point which is clearly non-red.
But, as in the case of the epistemicist, this is not the problem. The
real problem is just as before. Points l1 and/or l2 are jump points of
the characteristic function of the predicate ‘is red’ (the function which
assigns to each point x on the strip the truth value of the claim ‘Point
x is red’)—that is, points at which the value of the function jumps
significantly—while they are not locations at which adjacent points have
significantly different colours. Thus, the idea that if points x and y are very
close in respect of colour, then the sentences ‘Point x is red’ and ‘Point y
is red’ should be very close in respect of truth, is violated at one or both of
l1 and l2 .
188
m
Points x on the strip (red points at left)
Figure 4.5. The strip according to the fuzzy view.
Can the fuzzy theory accommodate Closeness? Yes. The fuzzy theory
takes as its set of degrees of truth the real numbers between 0 and 1
inclusive. Now there are certainly pairs x and y of reals in [0, 1] such that if
sentence S’s truth value is x and sentence T’s truth value is y, then S and T
are very similar in respect of truth (e.g., let |x − y| = 0.01). Thus, the fuzzy
framework can accommodate Closeness: it has a sufficiently rich structure
of truth values to allow arbitrarily small steps in F-relevant respects to
correspond to arbitrarily small steps in truth.
Consider our strip of paper again. Figure 4.5 represents the situation
according to the fuzzy view. If we are accustomed to phrasing what was in
fact a good objection to the epistemic, truth gap, and third value accounts
of vagueness in a misleading way—i.e. in terms of last points rather than
jump points—then we might think that this objection applies also to the
fuzzy account of vagueness. For point m is indeed the last point which is
definitely red, according to the fuzzy view. But this is not a problem—it
never was. The true problem was that in previous cases, points k, and l1
and/or l2 , were jump points of the function which assigns truth values to
sentences concerning points on the strip—i.e. points at which the value of
the function jumps significantly—and in the present case, m is not such a
jump point. (The value of the function does change at m—but the change
is gradual.) The existence of a last red point does not, in and of itself,
necessitate a violation of Closeness.⁷
⁷ Some might still think that the existence of a last red point is a problem for the fuzzy view. This
objection will be discussed in §6.2.
189
m m′
Patches x (red patches at left)
Figure 4.6. The patches according to the fuzzy view.
⁸ This is why I said in Ch. 3, n. 66, that a predicate which satisfies Closeness is not necessarily
radically higher-order vague.
⁹ Strictly speaking, we could say that S and T are very close in respect of truth if S’s truth value is
1 and T’s is 1 − , but not if S’s truth value is, say, 0.5 and T’s is 0.5 − . I can, however, see no
motivation for such a view.
¹⁰ Given the way the case was set up, we may assume without loss of generality that divides
evenly into 1.
190
¹¹ Recall from Ch. 2, n. 40, however, that we do not want to take the rationals between 0 and 1 as
our set of truth values.
191
4.3 Supervaluationism
The (standard) supervaluationist view cannot accommodate vagueness
(defined in terms of Closeness). My comments at the beginning of §4.2
about truth gaps and third truth values made no reference to whether
the truth values/gaps of compound sentences were assigned recursively, or
according to the supervaluationist method: those comments apply equally
to supervaluationist and recursive versions of truth gap and three-valued
views. Furthermore, the comments apply to any third truth status, whether
it is a truth gap, a third value, or a truth glut—thus subvaluationism is in
the same boat as supervaluationism here.
Nevertheless, the supervaluationist view uses more machinery than
the recursive view—in particular, the admissible classical extensions of
¹² A little while ago I mentioned a set of truth values of the form {0, , 2, . . . , 1 − 2, 1 − , 1}.
This fits the present pattern: = n −1 1 .
¹³ Of course I do not know what the relevant number n is, but that was not the problem. The
problem was supposed to be that any choice of number of truth values is arbitrary. I have just argued
that there is one choice of number that is not arbitrary. It is irrelevant in this context that I do not
know which number it is.
192
¹⁴ Here, and elsewhere, for ease of exposition I assume symmetry of the third status with respect to
truth and falsity (see p. 186). More cautiously, the point is that a sentence with the third status cannot
be very similar in respect of truth both to true and to false sentences.
¹⁵ The above is the situation regarding sentences whose truth values are given by the supervaluation.
If atomic sentences get their truth values directly from the base model, not from the supervaluation (see
p. 78), then even this much isn’t true. If ‘Bill is bald’ and ‘Ben is bald’ get their truth values from the
base model, and ‘Bill is bald’ is true and ‘Ben is bald’ lacks a truth value, then even if these sentences are
classically true/false on almost exactly the same admissible extensions of the base interpretation, they
are still not similar even in respects relevant to truth, let alone in respect of truth.
193
of ‘higher-order vagueness’. Can these help with the jolt problem? No: as
we shall now see, they do not help at all. The basic supervaluationist claim
is that a sentence is true/false (i.e. in the intended base interpretation) just
in case it is (classically) true/false in all the admissible classical extensions of
that base interpretation. This leads to a situation whereby in a Sorites series
for a vague predicate F, there are objects x in the middle of the series such
that Fx is neither true nor false, and—the important point in the present
context—there is a semantic jolt as we progress down the series from an
object a such that Fa is true to an adjacent object b such that Fb is neither
true nor false. Thus we have two objects a and b which—being adjacent
in a Sorites series for F —are very close in F-relevant respects, and yet Fa
and Fb are not very close in respect of truth. Now the fundamental idea
behind all the supervaluationist treatments of higher-order vagueness is to
assert that there are borderline cases of admissible extensions. Here is how this
would seem to help with our present problem:
There could thus be a borderline case admissible specification.¹⁶ Sentences that are
false in such a specification but true in all definitely admissible specifications will be
borderline cases of ‘true on all complete and admissible specifications’; and ‘ ‘‘p’’ is
true on all complete and admissible specifications’ will then be indeterminate. The
truth-conditions provided by supervaluationism will not then definitely determine
any truth-value for p or definitely determine that the sentence lacks a truth-value.
For the indeterminacy over whether the truth-condition is fulfilled implies that we
should not conclude that p is true, but nor should we call p neither true nor false
as we would if the condition (determinately) failed to be fulfilled. The truth-value
status of p (whether it is true, false or lacks a value) remains unsettled. And sharp
boundaries between the true predications and the borderline cases of ‘tall’ are
avoided. . . . A sorites series can be described as starting with true predications,
having borderline ones in the middle, and false ones at the end but being such that
there are no sharp boundaries between those categories.
(Keefe 2000, 203)
In fact, however, when we begin to spell out the basic idea (that there are
borderline admissible extensions) in any detail, we see that it fails to solve
the jolt problem. There are two main ways to spell out the idea. One way,
advocated by Keefe (2000, ch. 8, §I), is to keep the standard supervaluationist
story just as it was originally told, but then to say that certain predicates in
¹⁷ Fine’s approach was later taken up by Williamson 1994, §5.6; 1999; 2003, 698–9. Williamson
1994, 160–1 presents reasons for thinking that this second approach needs to be supplemented by the
idea that supervaluationism ‘‘must conduct its business in a vague metalanguage’’ (p. 161), but the latter
idea—which, as I have already said, I reject—is an optional addition to the second approach, rather
than an inherent part of it.
195
fact that DA is neither true nor false; third-order vagueness consists in the
fact that DDA is neither true nor false; and so on. If R were transitive and
symmetric, as well as reflexive, then the logic of D would be S, and while
it might be the case that A was true at some points and not others, in this
case DA , DDA , and so on would be false at all points. However given
that R is not transitive, it might be the case that not only A , but also DA ,
DDA , and so on are true at some points and false at others.
There is a deep conceptual problem with this proposal. The relative
admissibility relation R between boundaries is supposed to model the idea
that each precisification makes a ruling as to which other precisifications
are admissible. But the background set on which R is defined—the set of
points in the frame, the set of boundaries—is already the set of admissible
precisifications. We cannot allow every boundary into our frame. That is, we
cannot allow every classical model to be the first level of some boundary
in our frame. For if we did, then every D-free compound formula which
was not a classical logical truth or logical falsehood would come out as
neither true nor false in the base model—including, for example, statements
expressing penumbral connections. So we must restrict the classical models
which can be the first level of a boundary in our frame to the admissible
extensions. But then it is entirely irrelevant if a boundary comes along later
and tells us that it does not regard one of these models as admissible. ‘So
what?’, we should reply: ‘What does your opinion matter?’ For the set of
all classical models which figure as the first level of some boundary in our
frame simply is the set of admissible extensions, and whether or not some
boundary agrees is beside the point. But of course, the set of all classical
models which figure as the first level of some boundary in our frame is
a unique, precise set—and so the machinery of boundaries has ultimately
taken us no way towards implementing the initial idea that there should be
borderline admissible extensions.
Even setting this problem aside, however, the current proposal does not
help with the jolt problem. The problem was that there are adjacent objects
a and b in the Sorites series for F such that Fa is true and Fb is neither true
nor false (or has the value ∗). We are now told that this is in fact all right,
because while DFa is true, it is not the case that DFb is false: rather, DFb is
neither true nor false. But this does not help at all with the problem that
Fa and Fb are not very close in respect of truth—because it does not alter
the truth values of Fa and Fb. On the modified supervaluationist proposal, it
197
is still the case that Fa is true and Fb is neither true nor false (or has the
value ∗)—and so it is still the case that these two sentences are not very
close in respect of truth, even though a and b are very close in F-relevant
respects. The modified supervaluationist approach is simply on entirely the
wrong track (as far as solving the jolt problem goes). For we cannot render
two sentences close in respect of truth by changing the truth values which
we assign to other sentences. Distributing our original three truth values to
more sentences (ones containing ‘definitely’) does nothing; what we need
to do is countenance more truth values, and distribute them to our original
sentences (Fa, Fb, and so on).¹⁸
That, of course, is exactly what the degree-theoretic form of super-
valuationism does. For reasons that will now be obvious, the form of
supervaluationism where the base interpretation is fuzzy and we have a
measure over the set of admissible classical precisifications of the base inter-
pretation can accommodate vagueness (defined in terms of Closeness)—that
is, it does solve the jolt problem. On this sort of view, there need be no
point in the Sorites series for F at which we encounter adjacent objects a
and b such that Fa and Fb are not very close in respect of truth.
4.4 Plurivaluationism
I have said that the problem of ‘higher-order vagueness’ divides into two
problems: the location problem and the jolt problem. I have also argued
that epistemicism—based as it is on the view that each vague discourse has
a unique intended classical interpretation—can solve neither of these prob-
lems. The location problem—how could our practice with vague terms
single out one classical interpretation as uniquely correct?—was discussed
in Chapter 2. The jolt problem was discussed in §4.1: the epistemicist must
posit a sharp jump in a Sorites series between some persons a and b, who
are adjacent in the series and thus very similar in F-relevant respects, by
deeming Fa True and Fb False—and thus cannot accommodate vagueness.
I argued in Chapter 2 that plurivaluationism—the view that each vague
¹⁸ Similar comments apply to Fine’s other proposal for handling higher-order vagueness in the
supervaluationist framework, which is to distribute the three supervaluationist truth values not to
sentences containing the new operator ‘definitely’, but to sentences containing new truth predicates (Fine
1997 [1975], 146–7).
198
4.5 Contextualism
It is evident that whether or not a predicate satisfies Closeness at any
given point in time or in any given context turns entirely on the nature of
the base models on which the contextualist story is overlaid. For what
is distinctive about the contextualist story is what it says about how the
intended interpretation changes over time, from context to context. Within
a given context, the contextualist’s semantic story is identical to that of
the proponents of the non-contextualist version of the semantic picture
which goes with the type of base model (classical, partial, three-valued,
fuzzy, etc.) chosen by the contextualist. Thus, for reasons seen in earlier
²² I am writing in terms of contextualism built over a base of partial interpretations, but my comments
apply, mutatis mutandis, to all non-degree-theoretic forms of contextualism.
204
that on any occasion on which I use a vague term such as ‘tall’, I cleanly
divide the domain of discourse into three sets, in such a way that some
objects which are very close in tall-relevant respects are not assigned nearby
values by the characteristic function with which ‘tall’ is associated on that
occasion of use. The insult is that when I turn to the offending cases, the
context shifts in such a way that these cases no longer offend—now other
cases offend, but if I turn to them, the offence will be yet elsewhere, and so
on. Contextualism plus TA-Closeness yields a view on which, rather than
drawing blurred boundaries, the extensions of vague predicates are sharp but
shifty: at any instant we have a sharp boundary to the extension of a vague
predicate (the injury), but the boundary moves if we try to get a fix on
where it is (the insult).
Taking her cue from the passage from Greenough (2003) quoted on
p. 182 above (in the discussion of my criticism that epistemicism plus
JA-Closeness allows only sharp but unknowable boundaries, not genuinely
blurry boundaries), the contextualist might respond that ordinary speakers
cannot tell the difference between sharp but shifting boundaries and
genuinely blurred boundaries. However, this would not be plausible at
all. If blurriness and shiftiness were indistinguishable to ordinary speakers,
then we could not think—as surely we do—that the boundary of our
visual field is not just shifty (i.e. we cannot focus on it) but blurry as
well. Phenomenologically, there is an enormous difference between the
(blurry) boundary of the red region of the rainbow and the (sharp but
shifty) boundary between, say, the water and the concrete boat ramp as
the waves lap against it. It would be absurd to accuse ordinary speakers of
being unable to make this sort of distinction.
Soames (1999, 216–17) considers a similar objection to the one I have
just made, and responds: ‘‘The critic wrongly takes the inability to display
a sharp dividing line between things that are F and things for which the
predicate is undefined, according to some conversational standard, to imply
that there is no sharp dividing line, according to that standard. This ignores
the dynamic feature of the model. Once this feature is recognized, the
objection is defused.’’ But my objection is not defused by this: for the
objection does not ignore the dynamic feature of the contextualist view;
rather, the objection is precisely that the contextualist view ignores the
static aspects of vagueness. The idea that vague predicates draw blurred
boundaries concerns any particular use of a vague predicate, not a collection
205
of uses over time. According to this idea, a vague term is one which does
not draw sharp boundaries—not one which draws a succession of different
sharp boundaries in such a way that we can never focus attention on
just one of them. The TA-Closeness definition does not, then, capture
an essential feature of the ordinary conception of vagueness—the blurred
boundaries idea.
Nor does it capture the idea that vague predicates generate Sorites
paradoxes. Suppose that we have a predicate F which conforms to
TA-Closeness, but not Closeness. (If a predicate conforms to Close-
ness, then it will also conform to TA-Closeness. We know that a predicate
which conforms to Closeness will be Sorites-susceptible. What we now
want to know is whether TA-Closeness has the advantages of Closeness.
Thus we need to suppose TA-Closeness by itself.) Suppose that we have a
Sorites series for F. Will the corresponding Sorites paradox be compelling?
We have no reason to think so. Given TA-Closeness, and the fact that
adjacent objects in the Sorites series are very close in F-relevant respects,
we know that if we were to assert of some object in the series that it
is F, then we could not truthfully assert, in the same breath, that the
next object was not F. Thus TA-Closeness does yield an understanding
of why the forced march Sorites is compelling. But it yields no account
of why the regular Sorites argument is compelling. For, given only TA-
Closeness, we have no reason at all to think that the Sorites premiss is
true: i.e. that for any object in the series, if it is F, then so is the next
object. Indeed, if we have come to accept the contextualist story, we
will see immediately that this is false (even though we cannot point to
a pair of objects in the series which provides a counterexample). Thus
TA-Closeness is no better than JA-Closeness when it comes to yielding
an understanding of why vague predicates generate compelling Sorites
arguments.
Non-degree-theoretic versions of contextualism might, at this point, try
to stay in the game as error theories. The idea would be that the correct way
to codify our ordinary conception of vagueness is indeed in terms of the
Closeness definition. But lo and behold, there are no predicates out there
which satisfy Closeness—i.e. there are no vague predicates. There are only
predicates which satisfy TA-Closeness, and of which a non-degree-theoretic
contextualist theory is correct—i.e. predicates that draw a succession of
different sharp boundaries in such a way that we can never focus attention
206
on just one of them. Exactly the same comments apply to this proposal as
applied to the proposal for an epistemicist error theory in §4.1.
The upshot of the present chapter is as follows. If we understand ‘vague’
in terms of the Closeness definition, then if we want to say that there
exist vague predicates with associated Sorites series, we must adopt a
theory of vagueness that countenances degrees of truth. Furthermore, we
should understand ‘vague’ in terms of the Closeness definition. We saw the
advantages of doing so in the previous chapter; in the present chapter we
have seen that rival definitions do not share these advantages.
P A RT III
Degrees of Truth
This page intentionally left blank
5
Who’s Afraid of Degrees
of Truth?
¹ In this chapter and the next I generally write in terms of the fuzzy view, rather than large finite
degree theories, simply because, as noted on p. 190, I find the idea that there are continuum-many
truth values more appealing than the idea that there are only finitely many. Most everything I say
in these chapters could be applied—in some cases with minor changes of wording or detail, in the
remaining cases as is—to large finite degree theories.
² Ramsey 1990 [1926], 83 is often cited as an authority here—e.g. by Haack 1996 [1980] and
Sainsbury (1986, 97)—although the textual evidence does not seem so clear to me.
³ The identification of this picture as a motivating force behind the general opposition to degrees of
truth is due to Gideon Rosen, in conversation. (He identified the picture, but did not endorse it.)
211
The second version of the objection is weaker: rather than claiming that
the idea of degrees of truth makes no sense, the objector claims not to
know what degrees of truth are, and puts the onus on the degree theorist
to explain them properly. For example, Fara (2000, 54) complains of ‘‘the
absence of some substantial philosophical account of what degrees of truth
are’’. Given the generally sketchy level of description of the fuzzy view
in philosophical discussions, this objection is fair enough. At this point in
the present book, however, the objection has already been answered. To
recapitulate: The classical picture begins with the idea that objects may
possess or lack properties, and sentences may be true or false—and if (for
example) Bob possesses the property baldness, then ‘Bob is bald’ is true.
These ideas are captured by positing two truth values, 0 and 1: properties
are modelled by subsets, which are functions from a background set of
objects to the set of values; sentences are assigned the value 0 or 1 according
to whether they are false or true; and the link between truth and property
possession is captured in the model theory. Now the fuzzy picture begins
with the thought that there is something wrong here: the classical picture is
adequate in the realm of precise discourses such as mathematics (for which
the picture was developed), but not in contexts in which there is vagueness.
The thought is that some objects possess some properties to intermediate
degrees: for example, Bob is neither bald simpliciter nor non-bald simpliciter,
he is somewhere in between. From here, the basic correspondence intuition
about truth tells us that if I say that Bob is bald, my statement will be neither
true simpliciter nor false simpliciter —the truth of my statement mirrors Bob’s
baldness. Now all these ideas are captured by positing a continuum of
truth values, [0, 1], in place of the two classical truth values, {0, 1}, and
then modifying the rest of the classical picture in ways dictated by this
basic alteration. The result is the fuzzy picture: properties are modelled
by fuzzy subsets, which are functions from a background set of objects to
the set of values; sentences are assigned a greater or lesser value according
to whether they are more or less true; and the link between truth and
property possession is captured in the model theory—for example, if Bob’s
degree of baldness is 0.3, then ‘Bob is bald’ is 0.3 true. Thus, in place of the
classical picture, we get a richer picture, motivated by the basic idea that
outside precise realms such as mathematics, objects may possess properties
to intermediate degrees, in between complete possession and complete
lack thereof.
212
Suppose that now the objection is raised, ‘‘But what are these degrees
of truth?’’ In this context, what we are being asked is, ‘‘What are these
truth values of which you speak?’’ Now often, amongst philosophers, talk
of the truth values of sentences is simply a stylistic variant of talk of whether
sentences are true. This is fine in itself, but it should not blind us to the
fact that truth values, properly so-called, are not mere façons de parler:
they are objects. There is a particular point to positing these objects: we
wish to bring certain useful mathematical machinery—most notably the
machinery of functions—to bear on the analysis of phenomena such as truth
and validity. Using this machinery, we can achieve a very elegant and useful
picture of language and its relationship to the world, and hence we do not
baulk at positing the objects required to get the picture off the ground:
we need objects to serve as the arguments and values of functions, and in
particular, we need certain objects called truth values. Depending upon our
antecedent ideas about the phenomena we wish to model—for example,
whether properties may be possessed to intermediate degrees—we will
posit different sets of truth values, with different structural properties.
Now, to return to our question as to what these truth values are: this is
not a special question for the fuzzy theorist. If we want to ask, ‘‘What are
these fuzzy truth values—these degrees of truth?’’, then we should also
ask, ‘‘What are these classical truth values, 1 and 0?’’ In both cases, the
answer is that the truth values are elements in a particular sort of algebraic
structure—and what matters is the structural properties of the latter, not
the intrinsic nature of its elements.
There are alternative pictures available here. According to the Fregean
picture, the classical truth values are two definite, particular objects. What
are these objects like in themselves? Well, we just don’t care—we do
not ask this question. It is not that there is no correct answer—it is just
that beyond the structural properties of the set of truth values, we do not
know what they are like; but nor do we need to know. This goes not
just for classical truth values, but for fuzzy ones as well: they are particular
objects, but the question as to what intrinsic properties these objects have
(as opposed to what structural properties the set of all of them has) is beside
the point. The fuzzy theorist no more owes us an answer to this question
than the classical theorist owes us an answer to the corresponding question
about the classical truth values. According to the structuralist picture, on
the other hand, there is not a unique set of objects which are our truth
213
values. Rather, there are many different sets of objects, any of which is
suited to play the role of the truth values, and none of which is singled
out as the unique player of this role. What suits a set of objects to play
this role is precisely that it has the right structural properties. Thus, once
again, beyond the structural properties of the set of truth values, we should
not ask what the truth values are like in themselves: not because, as in the
Fregean picture, the answer exists but is unimportant, but because what is
true of the truth values in general is what is true of any and all sets of objects
which are suited to play the role of the truth values, and this extends no
further than structural properties.⁴
The Boolean algebra {0, 1} of classical truth values (whether thought of
in the concrete Fregean way or in the abstract structuralist way) serves very
well as the foundation of an elegant and useful theory which captures the
classical intuitions. The fuzzy theorist asserts that the Kleene algebra [0, 1] of
fuzzy truth values (whether thought of in the concrete Fregean way or in the
abstract structuralist way) serves very well as the foundation of an elegant and
useful theory which captures the fuzzy intuition that, outside mathematics,
objects may possess properties to intermediate degrees, and hence that
sentences may be true to intermediate degrees. Now there can certainly
be legitimate disagreements concerning which intuitions are correct and
hence worthy of being captured in our formal picture, and legitimate
disagreements concerning how best to capture a given set of intuitions.
Disagreements of the latter sort may lead to the positing of new algebras of
truth values with associated set theories and logics. But once such an algebra
of truth values has been described, we have said all that needs to be said
about what the truth values are—and hence, in the fuzzy case, about what
degrees of truth are. Any further question as to what the truth values are like
is out of place: all we need to know (and, if we are structuralists, all there
is to know in general) is the structural properties of the set of all of them.
The third version of the objection is that, rather than being confused
or inadequately explained, the idea of degrees of truth is unmotivated. In
particular, the claim is that the fuzzy theorist is confused on a particular
⁴ The structuralist view of truth values is analogous to the view of numbers presented in Benacerraf
1965. There is also a hybrid picture, in which truth values are particular objects—but shadowy,
insubstantial ones, with no properties aside from those they have in virtue of being part of a certain sort
of structure. Thanks to Amitavo Islam, in conversation and in his 1996, for clarification of the issues in
philosophy of mathematics alluded to in this paragraph.
214
point, and that once the confusion is cleared up, we see that we have no need
for degrees of truth. This objection occurs in a number of places; I shall take
the presentation in Keefe 1998b as my starting point, because it is particularly
clear.⁵ Keefe begins by noting that ‘‘In many paradigm cases of a vague
predicate F there is a corresponding measurable attribute related to F in such
a way that the truth-value status of Fx . . . is determined by x’s quantity of
that attribute. For example, the truth-value status of ‘a is tall’ is determined
by, or supervenes on, a’s height . . . ; similarly for the relation between ‘a
is hot’ and a’s temperature’’ (p. 575). I agree: these underlying attributes
associated with F are what I have called F-relevant respects.⁶ Keefe continues:
But although the measure of the underlying quantity may determine the applicability
of the vague predicate, it does not follow that this measure is reflected in non-
classical numerical truth-values. . . . Are degree theorists thus mistaken in claiming
that vague predicates come in degrees? I suggest that there is a sense in which F
can be said to come in degrees—call it coming in degreesm —whenever there
is a measure of the attribute F-ness, and where things have different degreesm
of F-ness by having more or less of the attribute. The degreem of heat of an
object will be a matter of its quantity of heat and we happen to call the measure
degrees Celsius. . . . But the fact that many vague predicates come in degreesm is not
enough for the degree theorist, who needs there to be implications for truth-values
or degrees of truth, so that if F comes in degrees, predications of F can be true
to intermediate degrees . . . coming in degreesm is not the sense of ‘‘coming in
degrees’’ required by the degree theorist.
(1998b, 575–6)
Again, I agree: the distinctive claim of the fuzzy theory (and of degree-
theoretic treatments of vagueness in general) is that there are degrees of
truth of predications of F, as well as degreesm of F. So far, however, we
have no objection to the fuzzy view: we have a warning not to confuse
degrees of truth with degreesm , but we have no argument to the effect that
the fuzzy theory is involved in such confusion. In fact the fuzzy view—at
least as I have presented it—is not confused on this point at all. Recall:
Closeness If a and b are very similar in F-relevant respects, then ‘Fa’ and
‘Fb’ are very similar in respect of truth.
⁷ Keefe has warned against confusing or conflating degreesm and degrees of truth. I have agreed that
this confusion is to be avoided, and have avoided it. So far, then, there is no problem for the fuzzy view.
However, some fuzzy theorists do apparently conflate the things Keefe warns us to keep apart. Keefe
1998b, 576–7 attributes the following argument to Forbes 1983, 241–2; see also Forbes 1985, 170:
Consider a pair of people, a and b, such that
This argument does not go through, for the reasons Keefe notes: with the ‘degreesm ’ sense of ‘degree’
in play, (2) follows from (1), but (4) does not follow from (2); whereas with the ‘degrees of truth’ sense
of ‘degrees’ in play, (4) follows from (2), but (2) does not follow from (1).
216
truths about, for example, comparative relations, this is no more than a measure of
an attribute related to, or underlying, the vague predicate.
(Keefe 1998b, 575)
⁸ The following discussion applies, mutatis mutandis, to all such pairs of words, e.g. ‘loud’ and
‘louder’, ‘heavy’ and ‘heavier’, etc.
⁹ I.e. h(y) ≤ h(x) and h(x) h(y). (Recall that H is equipped with an ordering relation ≤.)
217
tall
persons
O (persons)
H (heights) R (reals)
key: h f
sufficient height. Implicit in the word ‘sufficient’ here is the idea that T
should be closed upwards: for any x and y in H, if x ≤ y and x ∈ T, then
y ∈ T. This immediately gives us an important relation between ‘tall’ and
‘taller’: for any x and y in O, if x is taller than y and y is tall, then x is tall.¹⁰
So far, so good—and no degrees of truth in sight (only degreesm ). But
(as discussed in Chapter 4) there is something wrong with this model: it
ignores the vagueness of ‘tall’. If two objects a and b in O are very close in
respect of height, then ‘a is tall’ and ‘b is tall’ should be very close in respect
of truth. In the picture outlined above, however, assuming that O contains
a series of objects ranging from one which is not tall to one which is tall, in
very small steps of height, there will be a pair of things a and b in O whose
heights are very close, one of which is tall and the other not—i.e. ‘a is
tall’ is true, and ‘b is tall’ is false. Thus the proposed picture does not allow
for the vagueness of ‘tall’. In response to this problem, the fuzzy theorist
proposes that we replace the classical subset T of H with a fuzzy subset T,
and modify the requirement that T be closed upwards to the requirement
that for any x and y in H, if x ≤ y then x’s degree of membership in T
¹⁰ One might think that nothing is tall simpliciter, but rather tall for an F. In order to accommodate
this observation, we would need—instead of a single distinguished subset T of H —different subsets
TF for different kinds F of thing, with x being tall for an F just in case h(x) ∈ TF .
218
0
[0, 1]
O (persons)
H (heights)
key: h f T
R (reals)
Figure 5.2. Tall and taller (II).
is less than or equal to y’s degree of membership in T. (See Fig. 5.2. For
clarity, [0, 1] is drawn separately from R.) Now, ‘a is tall’ will be true to
whatever degree h(a) is in T, and thus we have the following important
relation between ‘tall’ and ‘taller’: for any x and y in O, if x is taller than y,
then the degree of truth of ‘x is tall’ is at least as great as the degree of truth
of ‘y is tall’. We now have the resources to accommodate the vagueness of
‘tall’ (if a and b in O are very close in respect of height, then it can now be
the case that ‘a is tall’ and ‘b is tall’ are very close in respect of truth), and
we are not committed to the idea that if a is taller than b, then ‘a is tall’ is
truer than ‘b is tall’. That is, we can also accommodate the idea that while
Kareem Abdul Jabbar is taller than Larry Bird, ‘Kareem Abdul Jabbar is tall’
is not truer than ‘Larry Bird is tall’, for both sentences are 1 true (i.e. true
to degree 1).¹¹
¹¹ Kareem Abdul Jabbar is 7 ft, 2 in. in height, and Larry Bird is 6 ft, 9 in. in height. I owe this
example to Scott Soames.
219
such that for any object x in O, j(x) = αh(x); thus if h(y) < h(x), then it is also the case that j(y) < j(x).
For more details on measurement theory see §6.1.5.
¹⁵ See e.g. McGee and McLaughlin 1995, 237, and Williamson 1994, 118.
221
ideas yield the same consequence relation (Priest 2001, 216–17)—and it is not classical (e.g. disjunctive
syllogism is not valid). Other ideas have also been explored in the literature: for example that an
argument is valid if the degree of falsity (i.e. one minus the degree of truth) of its conclusion can never
exceed the sum of the degrees of falsity of its premisses (Edgington 1992).
¹⁷ Proof. Let |=c be the classical consequence relation and |=f the fuzzy consequence relation.
(i) |=c B ⇒ |=f B . Given a fuzzy model Mf = (M, If ), we can construct a corresponding classical
model Mc = (M, Ic ) which (so to speak) treats numbers ≥0.5 as 1 and numbers <0.5 as 0. More
precisely, Ic assigns the same referents as If to singular terms, and where If assigns the n-place
predicate R the function f : M n → [0, 1], Ic assigns R the function f : M n → {0, 1} defined thus:
∀x ∈ M n , if f (x) ≥ 0.5 then f (x) = 1, and if f (x) < 0.5 then f (x) = 0. By a straightforward induction
on complexity of formulae we can show that for any fuzzy model Mf and any closed wf A , if
[A ]Mf < 0.5 then [A ]Mc = 0, and if [A ]Mf > 0.5 then [A ]Mc = 1. Now suppose there is a fuzzy
model Mf on which every member of is > 0.5 true and B is <0.5 true; then on the corresponding
classical model Mc , every member of is true, and B is false. So if there is no classical model of the
latter sort, there is no fuzzy model of the former sort. (ii) |=f B ⇒ |=c B . A classical model is (a
special case of ) a fuzzy model (one in which only the extreme values 0 and 1 get assigned). If there is
no fuzzy model on which every member of is > 0.5 true and B is <0.5 true, then a fortiori there is
no classical model on which every member of is true and B is false.
¹⁸ This does not mean that the fuzzy consequence relation is not transitive. It is transitive. (It must be,
because it is the same relation as the classical consequence relation, and that is transitive!)
223
¹⁹ Williamson 1994, 124 also claims that this is an intuitive view of the Sorites.
²⁰ See n. 57 below for an important point of clarification about the notion of a statement being
‘assertion grade’.
225
of language. Can the idea that truth comes in degrees be made to mesh with
important developments in other areas of philosophy of language, where
bivalence has hitherto been taken for granted? Or is the degree-theoretic
approach to vagueness a gated community in theory space?
At first sight, things look good for degree theories. The degree-theoretic
framework, as presented in §2.2.1, is designed to be interoperable with any
view which has at its heart ordinary logical, semantic, and set-theoretic
machinery. As we saw, any theory built on classical logic and/or set
theory can be seen as querying the underlying algebra of classical truth
values to return values for this or that function, and to return the res-
ults of applying the operations , , and to some given values. If
we unplug the Boolean algebra of classical truth values and plug in the
Kleene algebra of fuzzy truth values instead, then the higher-level theories
will, to be sure, get different answers to their questions; but the crucial
point is that everything will still work smoothly. Thus, to take just one
type of example, there is no problem extending degrees of truth from
the first-order semantics employed in discussions of vagueness to frame
semantics for modal logics, and to extensions and variations thereof such
as are involved in Stalnaker–Lewis-style treatments of counterfactuals and
in Kaplan-style treatments of indexicals. In the modal case, for example,
a predicate is assigned an intension: a function which assigns to each pos-
sible world a crisp subset of the domain of that world. To accommodate
degrees of truth, we simply suppose that the function assigns to each
possible world a fuzzy subset of the domain of that world.²¹ In the case
of Kaplan-style semantics, for example, we countenance both contexts of
utterance and circumstances of evaluation, and each expression is assigned
both a content (a function from circumstances to extensions/referents)
and a character (a function from contexts to contents). To accommod-
ate degrees of truth, we simply suppose that the extension assigned by
a content is a fuzzy (rather than crisp) subset of the domain of the
circumstance.
So far so good, but there is at least one theoretical viewpoint in
philosophy of language which has great explanatory power and hence has
gained considerable currency and found many applications, and which
has at its heart more than ordinary logical, semantic, and set-theoretic
the proposition asserted.²³ Each person has her own context set—her own
set of worlds which she regards as live options. However, ‘‘it is part of the
concept of presupposition that a speaker assumes that the members of his
audience presuppose everything that he presupposes’’ (Stalnaker 1999, 85).
We say that a context is non-defective if each participant does in fact have
the same context set.
Much of this is standard logico-set-theoretic fare—but the notion of
assertion poses special problems. For once we have degrees of truth in the
picture, we should also countenance degrees of assertion—i.e. degrees
of confidence of assertion—and corresponding degrees of belief.²⁴ For
suppose we have a Sorites series leading from tall men down to short
men. Suppose also that we have accepted a degree-theoretic account of
vagueness—so we think that ‘This man is tall’ goes gradually from 1
true, said of men at the beginning of the series, down to 0 true, said
of men at the end. Then what attitude should we adopt to (the pro-
position expressed by) ‘This man is tall’ as we consider various men
in the series? Surely we should go from being fully committed to the
proposition, both in thought and in talk, at the beginning of the series,
to fully rejecting it, in thought and in talk, by the end of the series,
via a gradually changing series of intermediate states of partial belief
and partially confident assertion, which decrease in degree of confid-
ence as we progress down the series. Here is Schiffer (2000, 223–4) on
this issue:
Sally is a rational speaker of English, and we’re going to monitor her belief states
throughout the following experiment. Tom Cruise, a paradigmatically non-bald
person, has consented, for the sake of philosophy, to have his hairs plucked
from his scalp one by one until none are left. Sally is to witness this, and will
judge Tom’s baldness after each plucking. The conditions for making baldness
judgments—lighting conditions, exposure to the hair situation on Tom’s scalp,
Sally’s sobriety and perceptual faculties, etc.—are ideal and known by Sally to be
such. . . . Let the plucking begin.
Sally starts out judging with absolute certainty that Tom is not bald; that is, she
believes to degree 1 that Tom is not bald and to degree 0 that he is bald. This state
of affairs persists through quite a few pluckings. At some point, however, Sally’s
²³ By ‘intersection’ here I mean ordinary set-theoretic intersection, for context sets and propositions
are sets of worlds.
²⁴ Cf. Sainsbury 1986.
228
Other things that we might say about the case—things that would avoid
admitting that Sally has degrees of belief—are (i) that Sally fully believes
that Tom is not bald until a particular hair is removed, from which point
on she fully believes he is bald; (ii) that Sally fully believes that Tom is
not bald until a particular hair is removed, at which point she enters an
indeterminate state in which she does not believe (to any degree, even 0)
that Tom is not bald and does not believe (to any degree, even 0) that Tom
is bald, and then when another particular hair is removed Sally comes to
fully believe that Tom is bald; and (iii) that Sally does not have attitudes
towards propositions such as ‘Tom is bald’, but only towards propositions
such as ‘Tom is bald to degree x’, each of which she either fully believes or
fully rejects. The problem with these approaches is that they do not fit the
phenomena. Contra (i) and (iii), Sally certainly seems to be unsure as to what
to believe and say about Tom’s baldness, at various points in the process;
and contra (ii), she does not have one catch-all ‘confused state’, which she
enters, remains in, then leaves: rather, she seems clearly to become less and
less sure that Tom is not bald, and then later more and more sure that he is.
The proponent of (iii) may reply that Sally’s qualified assertion that Tom
is bald—behaviour which seems clearly to indicate that there is some P
such that Sally is unsure whether P —should in fact be understood as a
full-on assertion that Tom is bald to an intermediate degree. But whether
or not this is plausible, (iii) faces other problems too, most notably that it
leads to a strange separation between truth, on the one hand, and belief
and assertion, on the other: we have a semantics which assigns degrees of
truth to atomic propositions such as ‘Tom is bald’, but we are then told we
229
uncertainty, she believes both of the propositions ‘Sali is tall’ and ‘Sali is
bald’ to degree 0.5. Suppose also that Sally regards these two propositions
as independent: supposing one to be true would have no bearing on her
beliefs about the other. Then, for familiar reasons, she should believe ‘Sali is
tall and bald’ to degree 0.25. Now suppose that midway through Schiffer’s
experiment, when Sally’s degree of belief that Tom is bald is 0.5, she
also believes to degree 0.5 that Tom is tall—on the basis of looking at
him and seeing that he is a classic borderline case of tallness.²⁵ Then what
should be her degree of belief that Tom is tall and bald? The answer 0.5
suggests itself very strongly: certainly the answer 0.25 seems wrong. If you
don’t think so, then just add more conjuncts (e.g. funny, nice, intelligent,
cool, old—where Sally knows of Sali only that he is not a borderline
case of any of them, and of Tom that he is a classic borderline case of
all of them): the more independent conjuncts you add, the lower the
uncertainty-based degree of belief should go, but this is clearly not the case
for the vagueness-based degree of belief (Schiffer 2000, 225; MacFarlane
2006, 221–2).
A second thought—Schiffer’s—is that there are two kinds of degree
of belief: uncertainty-based degrees of belief, or SPB’s (‘standard partial
beliefs’), and vagueness-based degrees of belief, or VPB’s (‘vagueness-related
partial beliefs’). In Schiffer’s view, we have two distinct systems of degrees
of belief: an assignment of SPB’s to propositions, which obey the laws of
probability, and an assignment of VPB’s to propositions, which obey the
laws of standard fuzzy propositional logic.²⁶ But there is a grave problem
for any proposal which posits two different systems of degrees of belief,
where it is allowed that a subject may have a degree of belief of one kind
of strength n in a proposition P and a degree of belief of another kind of
strength m = n in the same proposition P. The problem is that the very idea
of degree of belief is made sense of via the thought that a degree of belief
that P is a strength of tendency to act as if P. As Ramsey (1990 [1926], 65–6)
puts it:
the degree of a belief is a causal property of it, which we can express vaguely
as the extent to which we are prepared to act on it. . . . it is not asserted that a
²⁵ Suppose, for the sake of the example, that Tom Cruise is borderline tall.
²⁶ VPB(¬p) = 1 − VPB(p), VPB(p ∧ q) = min{VPB(p), VPB(q)}, and VPB(p ∨ q) = max{VPB(p),
VPB(q)}.
231
belief is an idea which does actually lead to action, but one which would lead to
action in suitable circumstances . . . The difference [between believing more firmly
and believing less firmly] seems to me to lie in how far we should act on these
beliefs.
But one simply cannot have two different strengths of tendency to act as if
P, in a given set of circumstances. Consider, for example, the proposition
that Fido is dangerous. When Fido enters the room, one will do some
particular thing: for example, sit still or jump and run. When Fido looks
at one, one will do some particular thing: for example, tremble or offer
him some beef jerky. When Fido barks, one will do some particular
thing: for example, scream; and so on. One cannot both back away
slowly and run screaming (at the same time), and it cannot both take
Fido getting within two metres of one to make one run away and require
Fido getting within one metre to make one run. So one cannot both
tend strongly to act as if Fido is dangerous and tend weakly to act as if
Fido is dangerous—at least, not if there is to be any sort of transparent
relationship between these tendencies and the way one actually acts. But
given that a degree of belief just is a strength of tendency to act, this
means that one cannot have two different degrees of belief in the same
proposition.
The proponent of two kinds of degrees of belief might offer a number of
responses here. (1) She might deny that there is a transparent relationship
between tendencies to act and the way one actually acts. So, in the case
of Fido, one might have both a strong tendency to act as if Fido is
dangerous and a weak tendency, and these interact so as to make one
behave in particular ways in particular situations (ways that we would
like to describe as indicating that one has a mid-strength tendency to act
as if Fido is dangerous—although on the current proposal, we cannot
straightforwardly say this). But for this view to get off the ground, we
would need to be told how exactly degrees of belief of the two sorts
combine to produce certain behaviour; furthermore, the view threatens to
make it impossible for us ever to know (even roughly) someone’s degree(s)
of belief in a given proposition. (2) She might say that although there
are indeed two kinds of degrees of belief, they always have the same
strength, for every proposition. But clearly this would run us headlong
into the problem discussed above, that partial beliefs arising from vagueness
232
do not and should not behave in the same ways as partial beliefs arising
from uncertainty. (3) She might deny that degrees of belief are to be
understood in terms of strength of tendency to act. But any view which
disconnects degree of belief from tendency to act threatens to undermine
the utility of the notion of degree of belief, and furthermore any candidate
replacement proposal—for example, the view that the difference between
believing more firmly and believing less firmly is a matter of strength of
feeling²⁷—would seem to face the very same problem (one cannot have
two different intensities of feeling about one proposition). (4) She might
claim that one never has both kinds of degree of belief in the same
proposition at the same time. For suppose, for reductio, that you have an
uncertainty-related degree of belief of 0.3 that Dobbin wins the race and a
vagueness-related degree of belief of 0.5 that Dobbin wins the race. How
could you have acquired both these beliefs? In order to acquire the first, you
would need to lack evidence concerning who wins. In order to acquire
the second, you would need to have all the relevant evidence, and see
that it—i.e. the world itself—leaves it unsettled who wins.²⁸ So clearly
you could not have both these degrees of belief at once. There are still
problems for this view, however. First, we need to be told how to reason
with several propositions—and compounds thereof—in some of which
we have degrees of belief of one type, and in others of which we have
degrees of belief of the other type. Second, what justifies saying that we
have here two non-interacting systems of degrees of belief, rather than one
system, which assigns degrees to all propositions, but where these degrees
behave differently in different situations (e.g. sometimes they obey the laws
of probability, sometimes they do not)?
This is the third possibility regarding the relationship between vagueness-
based degrees of belief and uncertainty-based degrees of belief: the
suggestion that what we have is one univocal notion of degree of
belief—one single system of assignments of degrees of belief to propos-
itions—but where the degrees assigned sometimes behave in accordance
²⁷ This is the view with which Ramsey contrasts his own view, in the discussion quoted earlier.
²⁸ I am imagining a case where due to the vagueness of the boundaries of horses, two horses are
equally good candidates for having crossed the line first. In practice this would no doubt be deemed
a tie, but imagine that we are examining very high-resolution pictures of the finish, and that we are
interested not in the practical question of distributing winnings, but purely in the question of which
horse in fact crossed the line first.
233
with the laws of probability, and sometimes do not. This is the sort of view
I shall advocate in the next section.²⁹
²⁹ Apart from my own view, another view which fits the description just given is that of Field
(2000). Field supposes that an agent has a probability function P over propositions; he supposes also
that the language includes a determinately operator D; and he then proposes that the agent’s degree
of belief Q(α) in any proposition α is given by Q(α) = P(Dα). Thus my degree of belief that α
is my subjective probability that determinately α. It may sound, then, as though we do have two
different systems of degrees of belief: P-values and Q-values. But Field says that only Q-values are
to be thought of as degrees of belief: ‘‘P should be thought of as simply a fictitious auxiliary used
for obtaining Q’’ (p. 16); ‘‘P [should] not be taken seriously: except where it coincides with Q, it
plays no role in describing the idealized agent’’ (p. 19). Field’s proposal, however, is of no use to us
in our project of exploring how degrees of belief arising from degrees of truth are related to degrees
of belief arising from uncertainty: the proposal does not employ degrees of truth, and it is hard to
see how to add them. I also have some other worries about Field’s proposal. One worry concerns
the appearance of a primitive determinately operator within the contents of beliefs. Another worry
concerns the downgrading of P: I think Field takes this too far. In my proposal (§5.3.2), subjective
probabilities do play an important role in describing an agent, but they are not to be identified with
degrees of belief. Field, on the other hand, seems to be in the grip of the view that if subjective
probabilities are allowed into the picture at all (as anything beyond fictitious auxiliaries), then they
will automatically grab the mantle ‘degrees of belief ’. A third worry concerns the consequence of
Field’s view (see his p. 17) that Q(α ∨ ¬α) is always 1: I think (and my view will capture this
idea) that if Sally has vagueness-based degrees of belief of 0.5 that Tom is tall and of 0.5 that Tom
is not tall, then she should also have a degree of belief of 0.5 that Tom is tall or not tall (cf. §5.5
below).
234
is in that set, and assigning a set of worlds measure 0 means that you are
absolutely certain the actual world is not in that set—the three probability
axioms are well-motivated:
P1. For every set A ⊂ W , P(A) ≥ 0
P2. P(A ∪ B) = P(A) + P(B) provided A ∩ B = ∅
P3. P(W ) = 1
(2) At each possible world, each proposition has a particular degree of
truth. Thus we may regard each proposition S as determining a function
S : W → [0, 1], i.e. the function which assigns to each world w ∈ W
the degree of truth of S at w.³⁰ The relationships between the functions
associated with various propositions will be constrained in familiar ways
by the logical relationships between these propositions: thus, for example,
(S ∨ T) (w) = max{S (w), T (w)}, (S ∧ T) (w) = min{S (w), T (w)}, and
(¬S) (w) = 1 − S (w).
(3) We have a measure over worlds (the agent’s epistemic state P) and
functions from worlds to real numbers (each proposition S). Thus S is a
random variable, and I propose that we identify the agent’s degree of belief
in S with her expectation (aka expected value) of S.³¹
To get a feel for the proposal, consider the case where there are
finitely many possible worlds. One’s probability measure—which assigns a
probability to each set of worlds—is in this case determined, via the addit-
ivity axiom P2, by the values assigned to singleton sets: P({w1 , . . . , wi }) =
P({w1 }) + . . . + P({wi }). So, one assigns each world a degree of likelihood:
a number indicating how likely one thinks it is that that world is the actual
world. Each world w itself assigns each proposition S a degree of truth
S(w). Now, my degree of belief in S is my expectation for S, i.e. my
expected value for S’s degree of truth. Let us denote this E(S). In this
finite case, it can be calculated thus, where w1 . . . wn are all the possible
worlds:
E(S) = P({w1 }) · S(w1 ) + . . . + P({wn }) · S(wn )
³⁰ For the sake of simplicity of presentation, I may sometimes conflate S and S , i.e. write of a
proposition as being a function from worlds to degrees, rather than as determining such a function.
³¹ Further to n. 30: there can be two distinct random variables which have the same value at every
point in the sample space. Of course, considered as functions they will be identical—but I mean that
they could still be considered as distinct random variables which happen to coincide. For example, on
a given sample space it may turn out that the height of each person in millimetres is the same as her
bank balance in dollars.
235
³² It is important to note that I am not claiming that two persons who have the same degree of
belief that S will behave in the same ways (or have the same tendencies to behave in certain ways). I am
claiming that they will have the same tendency to act as if S —and whether a person’s behaving in a
certain way constitutes her acting as if S depends on her preferences (desires, utilities) and on her other
beliefs. For example, let S be the proposition that there is an especially fragrant rose in Bob’s garden.
For a rose-fancier, approaching Bob’s garden might constitute acting as if S, whereas for a person with
an aversion to roses—or a rose-fancier with false beliefs about the location of Bob’s garden—moving
away from Bob’s garden might constitute acting as if S.
³³ I am making the assumption here that your preferences regarding team members can be summed
up thus: ‘The taller the better.’ If, on the other hand, you wanted only very tall players—so you are
236
The proposal also has the desired feature that sometimes degrees of belief
behave like probability assignments, and sometimes not. Before showing
this, I shall generalize the picture presented above. For so far we have con-
sidered only the special case where we have finitely many possible worlds,
but of course we cannot, in general, suppose that there are only finitely
many possible worlds—indeed, we cannot suppose that there are only
countably many. But if there are uncountably many possible worlds, then
(i) we cannot assume that the agent’s probability measure is defined on all
subsets of the space of possible worlds,³⁴ and (ii) we cannot assume that every
proposition determines a measurable function from worlds to truth values,
i.e. a random variable. We shall handle this situation in the standard way. In
regards to point (i), we suppose there to be a family F of subsets of the space
W of all possible worlds which is a σ-field, i.e. it satisfies the conditions:
1. W ∈ F .
2. For all A ∈ F , A ∈ F .
3. For any countable number of sets A1 , . . . , An in F , n An ∈ F .³⁵
Our probability measure will be defined on F , i.e. it will assign probabilities
to sets in F , and not to other subsets of W ; the sets in F will be called the
measurable sets of possible worlds.³⁶ In regards to point (ii), for a function
S from worlds to the reals to be measurable, i.e. a random variable, it
must satisfy the condition that for any real x, {w ∈ W : S(w) ≤ x} ∈ F . If
such a function is bounded, it will have a well-defined expectation E(S).
All propositions are functions from worlds to [0, 1], and hence bounded.
As for the condition that they be measurable, we henceforth restrict our
attention to propositions which meet it. This means that we consider only
propositions S such that it makes sense to ask, ‘‘How likely do you take it
to be that this proposition has a truth value within such-and-such limits?’’
just as averse to signing up a borderline tall person as to signing up a short person—then signing up
P would not constitute acting as if P is tall; rather, it would constitute acting as if P is very tall (recall
n.32 above). In that case, in the situation described—where your expectation that Bob is tall is 0.5, and
your expectation that Bob is very tall is 0—you would have no tendency to sign up Bob. Thanks to
Wlodek Rabinowicz for helpful comments here.
³⁴ In fact, we might not want to assume this anyway—see objection 10 in §5.3.3.
³⁵ Note that by the De Morgan laws, we could equivalently replace union with intersection in
condition 3.
³⁶ Once we have made this alteration to our setup, it is standard also to change axiom P2 so that
it applies not just to unions of two sets, but ∞to unions
of∞countably many sets: i.e. for any countable
collection {Ai } of pairwise disjoint sets, P n=1 An = n=1 P(An ).
237
With the general picture now in place, we can make the following
definitions:
Definition (vagueness-free situation). An agent is in a vagueness-free
situation (VFS) with respect to a proposition S iff there is a measure 1
set T of worlds (i.e. a set T such that P(T) = 1) such that S(w) = 1
or S(w) = 0 for every w ∈ T. (That is, the agent may not know for
sure whether S is true or false, but she does absolutely rule out the
possibility that S has an intermediate degree of truth: for she is certain
that the actual world is somewhere in the class T, and everywhere in
T, S is either 1 true or 0 true.)³⁷ An agent is in a VFS with respect to
a set of propositions if she is in a VFS with respect to each of the
propositions in .
Definition (uncertainty-free situation). An agent is in an uncertainty-
free situation (UFS) with respect to a proposition S iff there is a measure
1 set T of worlds and a k ∈ [0, 1] such that S(w) = k for every w ∈ T.
(That is, it is totally ruled out that S has a degree of truth other than k:
for the agent is certain that the actual world is somewhere in the class T,
and everywhere in T, S is k true.)³⁸ An agent is in a UFS with respect
to a set of propositions if she is in a UFS with respect to each of the
propositions in .
We can now establish four results which show when degrees of belief
behave like probability assignments and when they do not.³⁹
Proposition (degrees of belief equal probabilities in VFSs). If an agent
is in a VFS with respect to S, then E(S) = P({w : S(w) = 1}).
Proposition (degrees of belief equal degrees of truth in UFSs). If an
agent is in a UFS with respect to S, then E(S) equals the degree of truth
which the agent is certain S has.
³⁷ Note that one is in a vagueness-free situation with respect to S, in my sense, if one is subjectively
certain that S does not have an intermediate degree of truth. Of course, it might still be the case that in
fact S does have an intermediate degree of truth. So a ‘vagueness-free situation’ might better be called a
‘perceived vagueness-free situation’. I shall continue to use the former term, however, for the sake of
brevity. Thanks to Peter Milne here.
³⁸ If there is such a k, it is unique. A world in which S is k true is not a world in which S is m true
for any m = k. So if there are k and m with k = m such that there is a measure 1 set T of worlds with
S(w) = k for every w ∈ T and a measure 1 set U of worlds with S(w) = m for every w ∈ U, then T
and U are disjoint, and so P( T ∪ U) = 1 + 1 = 2 by P2, violating P3.
³⁹ The proofs of the following four propositions are straightforward and are left as exercises for the
interested reader.
238
⁴⁰ The closure requirement is no restriction, because if one is in a VFS with respect to a class of wfs,
then one is in a VFS with respect to the closure of that class under the operations of forming wfs using
our standard propositional connectives: this follows because whenever the component wfs are 1 true
or 0 true, so are the compounds.
⁴¹ There are several possible definitions of ‘tautology’ in fuzzy logic. All we need for the proof is
something they all agree on, viz. that a tautology never gets the value 0.
⁴² There are several possible definitions of ‘mutually exclusive’ in fuzzy logic. All we need for the
proof is something they all agree on: viz. that two mutually exclusive propositions never both get the
value 1.
⁴³ Again, the closure requirement is no restriction, because if one is in a UFS with respect to a class
of wfs, then one is in a UFS with respect to the closure of that class under the operations of forming
wfs using our standard propositional connectives: if one is certain that S is m true and that T is n true,
then one is certain that S ∨ T is max{m, n} true, that S ∧ T is min{m, n} true, and that ¬S is 1 − m
true.
239
thinks it is that the actual world is in that set) and the objective facts about
how true each proposition is in each world. Specifically, the agent’s degree
of belief in a proposition is the agent’s expected value for its degree of
truth: roughly, the average of its truth in all the worlds the agent has not
ruled out, weighted according to how likely the agent thinks it is that
each of those worlds is the actual one. In some situations, the agent will
have ruled out vagueness: she may not know which world is actual, but
she is certain that in the actual world, some propositions of interest are
either fully true or fully false. In such situations, her degrees of belief will
behave just like probabilities (propositions 1 and 3). In other situations,
the agent will be free of uncertainty with respect to some propositions of
interest: she is certain of exactly how true they are in the actual world. In
such situations, her degrees of belief will behave just like degrees of truth
(propositions 2 and 4). In situations which are neither vagueness-free nor
uncertainty-free—that is, where the agent is unsure of the truth values
of some propositions of interest, and cannot rule out vagueness, that is,
cannot rule out that they might have intermediate degrees of truth—her
degrees of belief in those propositions need not behave like probabilities
or degrees of truth. (In situations which are both uncertainty-free and
vagueness-free—that is, the agent knows of each of the propositions in
question that it is 1 true, or that it is 0 true—degrees of belief behave
both like probabilities and like degrees of truth. This is possible because
the behaviours of probabilities and degrees of truth coincide in this special
case.) In all cases, I maintain that an agent’s expectation of a proposition
S’s degree of truth is an accurate measure of her tendency to act as if S, and
this is why I identify degrees of beliefs with expectations.
My proposal contrasts with views such as the following:
Let our degrees of belief be represented by a probability measure, P, on a standard
Borel space (, F, P), where is a set, F is a sigma-field of measurable subsets of
, and P is a probability measure on F.
(Skyrms 1984, 53)
[By a reasonable initial credence function C] I meant, in part, that C was to
be a probability distribution over (at least) the space whose points are possible
worlds and whose regions (sets of worlds) are propositions. C is a non-negative,
normalized, finitely additive measure defined on all propositions.
(Lewis 1986 [1980] 87–8)
240
The crucial difference between these views and mine is that they equate
an agent’s degrees of belief directly with her subjective probabilities.⁴⁴
My view, on the other hand, countenances the subjective probability
measure—it models the agent’s epistemic state—but regards degrees of
belief as resultants of this state and degrees of truth. In the sort of cases
Skyrms and Lewis were considering, in which bivalence was assumed, this
difference makes no difference (propositions 1 and 3). However, if we
want to add degrees of truth to the mix, then we will run into all sorts
of problems if we have already identified degrees of belief with subjective
probabilities—for, as we saw at the outset, degrees of truth also give rise to
degrees of belief, but these degrees of belief do not behave like probabilities.
On the other hand, if we identify degree of belief with expectation of truth
even in the bivalent case, then we can generalize smoothly to the case of
degrees of truth.
(2) Suppose that Jim is tall to degree 0.5, fat to degree 0.5, and bald
to degree 0.5, while his workmate Tim is tall to degree 1, fat to degree
0.5, and bald to degree 0.5—and we know all this. Suppose that we are
to award a prize to any man in the office who is tall, fat, and bald. On
the fuzzy account, both ‘Jim is tall, fat, and bald’ and ‘Tim is tall, fat, and
bald’ are true to degree 0.5, and we know this; so on my account, our
degrees of belief in ‘Jim is tall, fat, and bald’ and in ‘Tim is tall, fat, and
bald’ are the same—both 0.5. Yet we would sooner give a prize to Tim
than to Jim. So one’s expectation of the truth value of P is not an accurate
measure of the strength of one’s tendency to act as if P.⁴⁵ Reply: I agree
that we would sooner give a prize to Tim than to Jim. I deny that what is
driving our preference for Tim over Jim here is our degrees of belief in the
conjunctions ‘Jim is tall, fat, and bald’ and ‘Tim is tall, fat, and bald’. I think
what drives our preference is the following: with respect to fatness, Tim
and Jim are a tie; with respect to baldness, they are a tie; but with respect
to tallness, Tim is ahead. So tallness breaks the tie, and makes us prefer
Tim. But this is quite different from saying that our degree of belief in the
conjunction ‘Tim is tall, fat, and bald’ is higher than our degree of belief in
the conjunction ‘Jim is tall, fat, and bald’—something which I deny.
(3) If your degrees of belief do not conform to the probability calculus,
then you are subject to Dutch book, i.e. you are irrational. Reply: One
should not bet at all on a proposition S unless one is in a vagueness-free
situation with respect to S; if one does bet in a non-VFS, then it is for
that reason alone that one is irrational. Suppose you are not in a VFS
with respect to S. Suppose, first, that you know that S is k true, for
some k ∈ (0, 1); say k = 0.5 for the sake of argument. Then you should
not bet on S. For to bet is to agree to an arrangement whereby you get
such-and-such if S turns out to be the case. But you already know what is
the case—and you know that it is, in the nature of things, indeterminate
whether S —hence indeterminate whether you get your payoff. Knowing
all this, you should not bet in the first place. Second, suppose that you
do not know whether S is true—and you cannot rule out that S has an
intermediate degree of truth. In this case again you should not bet, because
for all you know, the bet will not—for the sort of reason just seen—be
able to be decided. Of course, if there is in place some system for deciding
⁴⁵ Based on objections from Roy Sorensen and Dorothy Edgington (in discussion).
242
⁴⁶ My comments about not betting in non-VFSs are concerned with standard bets—i.e. bets which
do not specify what is to happen (who gets what) when the proposition in question is neither true nor
false. Milne 2007 discusses a new type of betting arrangement, tailor-made for vagueness, on which
one could legitimately bet in a non-VFS. The basic idea (although this is not the way Milne expresses it)
is that if one bets on S, and S is n true, then one receives n times the stake (so in the special case where
S is 1 true, one receives all the stake—so one’s net winnings are the stake minus what one put into the
stake; and where S is 0 true, one receives none of the stake). Of course, this complements rather than
conflicts with my comments above (Milne was not suggesting otherwise). I say that one should not
accept an ordinary bet if one thinks that vagueness may be present—for when vagueness is involved,
there is no way of deciding such a bet. This does not mean that one should not accept a new kind of
bet—one designed precisely to avoid the problem faced by ordinary bets when vagueness is present,
by explicitly building in a decision procedure which works even when the proposition on which one
is betting has an intermediate degree of truth.
243
⁴⁷ Based on objections from Andy Egan and Josh Parsons (in discussion).
244
earlier. But with the leaf in plain sight, you would not accept a bet that it
is red at any price: for we can all see quite plainly that the leaf is neither
clearly red nor clearly non-red, and so we can see at the outset that the bet
will misfire.⁴⁸
(6) Another objection in the ballpark of the previous three: Barnett
(2000) claims that some attitudes cannot be beliefs if they do not satisfy the
standard laws, i.e. the probability axioms. That is, it is constitutive of beliefs
that they behave like probabilities. Reply: What is constitutive of belief
is the idea that a belief that S is a tendency to act as if S in appropriate
circumstances. It turns out —i.e. it is a fact about beliefs, but not constitutive
of them—that in certain situations—viz. VFSs—beliefs do behave like
probabilities. But in other situations they do not—and yet they still count
as beliefs, because of their connection with tendencies to act.
(7) If my degree of belief measures my tendency to act as if S, how can
it be that I might have the same degree of belief that S in two situations,
and yet behave very differently in those situations? Say my degree of belief
in S is 0.5 because I am uncertain whether S is 1 true or 0 true. Then I
might bet on S, if the price and prize are right. But suppose my degree
of belief in S is 0.5 because I am certain that S is 0.5 true. Then, for the
reasons discussed above, I will not bet on S, no matter what the price and
prize. So the same state—a degree of belief of 0.5 in S —leads to different
actions. How can this be? Reply: These different actions are the results not
of a single belief, but of complexes of beliefs, which are different in the two
situations (cf. n. 32 above). A 0.5-degree belief that S combined with the
belief that, whatever further evidence comes in, I will not alter my degree
of belief in S, leads to refusing to bet; a 0.5-degree belief that S combined
with the belief that further evidence might come in leading me to believe
to degree 1 that S, and that further evidence might come in leading me to
believe to degree 0 that S, leads to accepting certain bets. What I do want
to maintain is that a degree 0.5 belief that S equates to a certain tendency
to act as if S, no matter what the source of this degree of belief. Thus I
maintain that if, in a given situation, you have a degree 0.5 belief that Bill
⁴⁸ Those who feel strongly that where there are degrees of belief there must be betting quotients
can find comfort in the kind of betting arrangement discussed in Milne 2007 (see n. 46 above). Milne
shows that the fair betting quotient a rational agent assigns to a bet of his kind on A perfectly matches
the agent’s degree of belief in my sense that A, i.e. her expectation of A’s degree of truth.
245
is bald (because you are in a VFS with respect to this proposition, and are
uncertain whether Bill is 1 bald or 0 bald), and a degree 0.5 belief that Ben
is bald (because you are in a UFS with respect to this proposition, and are
certain that Ben is a dead-central borderline case of baldness), then your
tendency to treat Bill as bald will be exactly the same as your tendency to
treat Ben as bald. So, for example, if you need to sign up a bald man for
a door-to-door sales campaign, you will be indifferent between signing up
Bill and Ben.⁴⁹
Now we consider some objections to the idea that the agent’s epistemic
state is modelled by a probability measure over possible worlds. First two
objections to the possible worlds part; then an objection to the probability
measure part.
(8) There are agents for whom it is epistemically possible that Hesperus
is not Phosphorus, and yet there are no possible worlds in which Hesperus
is not Phosphorus. Reply: The agent assigns positive probability to worlds
in which there is one heavenly body which is visible both in the evening
and in the morning, and positive probability to worlds in which there
are two heavenly bodies, one of which is visible in the evening and
one of which is visible in the morning. This captures the phenomenon;
it simply cannot be described in terms of the agent assigning a positive
degree of belief to both ‘Hesperus is Phosphorus’ and ‘Hesperus is not
Phosphorus’.⁵⁰
(9) It might be that I know everything about the objective state of
the world—i.e. I assign measure 1 to a set containing just one possible
world—and yet there are still things I do not know, for example, what
time it is, where I am, or who I am. This is impossible on my view: once
I assign measure 1 to a set containing just one possible world, there can
be nothing left to find out. Reply: If we take seriously these Perry-type
worries, we can replace the measure over worlds with a measure over
centred worlds (Lewis 1979; Chalmers 2001).
(10) It is absurd to suppose that an agent’s epistemic state is sufficiently
determinate to be modelled by a probability measure over worlds, which
assigns a unique real number to every set of worlds. Reply: There are two
⁴⁹ If you need a degree-1 bald man (or a perfectly bald man, or a very bald man), then of course you
will not be indifferent. Cf. n. 33 and objection 1.
⁵⁰ Cf. Kripke 1980, 102 ff.
246
parts to this worry: the generality of the measure (it assigns a number to
every set of worlds) and the precision of the measure (it assigns a unique real
number to each set of worlds). Regarding generality, we have not in fact
said that the probability measure assigns a number to every set of worlds:
only to every set of worlds in F . So if we take this worry seriously, we
can simply suppose F to be rather coarse; the upshot would be that the
agent has degrees of belief in fewer propositions. Regarding precision, if
we take this worry seriously, we can replace the single measure with a
family of measures; the upshot would be that the agent does not have a
unique degree of belief in each proposition, but rather a family of degrees
of belief.
above example, the set will contain worlds in which ‘Snow is white’ is
true to degree 1 and ‘Bob is bald’ is true to degree 0.5. The sense in which
the agent presupposes these propositions is cashed out thus: we suppose that
the agent assigns measure 1 to the corresponding context set. In a non-
defective context, each agent has the same set of presuppositions—that is,
they presuppose the same propositions to the same degrees—that is, they
all assign measure 1 to the context set determined by these propositions
and degrees.⁵¹
An agent may have a degree of belief of (say) 0.5 that S because she
is uncertain as to S’s truth value or because she is certain that S’s truth
value is 0.5. In both cases she will express her belief via a 0.5-confident
utterance of S. However, only in the latter case will this utterance count as
an assertion. Recall Stalnaker’s key idea that an assertion puts forward a set
of worlds; if the assertion is accepted, the context set is adjusted by striking
out worlds not in the asserted set. Now note that a 0.5-confident utterance
born of uncertainty does not specify a set of worlds. If I express that I do not
know what S’s truth value is, but that the weighted average of the degrees
of truth which, for all I know, it might have is 0.5, then I do not rule out
any worlds: for all I have said, we still might be in a world where S has any
degree of truth whatsoever. On the other hand, a 0.5-confident utterance
born of certainty that S is 0.5 true does specify a set of worlds: the set of
worlds in which S is 0.5 true. That is why unconfident utterance of S
counts as an assertion only if the utterer is in a UFS with respect to S. But
now recall Stalnaker’s basic idea that a conversation consists of assertions,
where the purpose of an assertion is to narrow the context set—and
where all participants presuppose that this is the purpose of speaking in a
conversation. Given this, if someone makes an n-confident utterance that
S in a conversation, we should interpret her as making an assertion—i.e.
as being in a UFS with respect to S, and claiming that the actual world is
one in which S is n true.
If this assertion is not challenged, how will the context be updated?
Well, each conversationalist will conditionalize her probability measure on
⁵¹ This still leaves plenty of room for the conversationalists to be in different epistemic states, i.e. to
have different probability measures—it’s just that there is one set of worlds which all their measures
must assign 1. Note also that, just as in Stalnaker’s picture a conversationalist need not really believe
what she presupposes for the purposes of a conversation, so too in this generalization of his picture a
conversationalist may for the purposes of a conversation adopt a probability measure which does not
represent her true epistemic state.
248
the set Sn of worlds in which S is n true. That is, she will update her old
probability measure P to the new probability measure P given by:⁵²
P(T ∩ Sn )
P (T) = P(T/Sn ) =
P(Sn )
This guarantees that where C is the old context set (so P(C) = 1),
P (C ∩ Sn ) = 1—that is, the old context set intersected with the newly
asserted set of worlds becomes the new context set; that is, it gets measure 1.
To conclude §5.3: I have not (of course) shown one by one that degrees
of truth are compatible with every important development that has been
made in areas of philosophy of language, beyond the study of vagueness,
where bivalence has hitherto been taken for granted. However, given
the level of interoperability we have now demonstrated between degrees
of truth and other important parts of our conceptual repertoire—logical,
semantic, set-theoretic, and now probabilistic machinery—we can
certainly see that the degree-theoretic approach to vagueness is very far
from being a gated community in theory space.⁵³
⁵² This will not be defined if P(Sn ) = 0, but that is not a problem here, because a conversationalist
should not assert that the world is a way which it was already presupposed not to be.
⁵³ Many thanks to Agustín Rayo, for putting to me both the general challenge of showing that
degree-theoretic treatments of vagueness can mesh with developments in philosophy of language outside
the study of vagueness, and the case of Stalnakerian pragmatics as a particular instance of this challenge.
⁵⁴ We have taken the degrees of assertion to be represented by the real interval [0, 1], so that they
correspond to the degrees of truth. In practice, of course, we cannot discriminate speech acts this
finely—just as we cannot, in practice, determine the degrees of truth of utterances as finely as the
apparatus of fuzzy semantics allows.
249
⁵⁵ One might think that in the classical case, assertibility should be regarded as a graded notion. For
what proportion of conversationalists have to object, and how strongly, before an utterance counts as
not having been accepted? Perhaps there is something to this thought, although I shall not pursue it
here. The essential point remains that introducing degrees of assertion does not in itself provide a reason
for thinking of assertibility as a graded notion. If there are reasons why we should think of assertibility
in this way, they apply equally to the classical case (where we have just one act of assertion) and to the
degrees of truth case (where we have one act of assertion for each degree of truth).
250
⁵⁶ It is not in dispute that assertion is subject to many norms (although sometimes terms other than
‘norm’ are used—e.g. ‘rules’ or ‘warranted assertibility conditions’). What is in dispute in the literature
is whether one norm is primary—in the sense that it is uniquely characteristic of assertion to be subject
to this norm, and that the other norms to which assertion is subject are a product of this primary norm
and more general rules not specific to assertion—and if so, which norm it is. For example, Williamson
(2000, ch. 11) argues against the idea that the truth norm ‘assert p only if p is true’ is primary in this
sense, and in favour of the idea that the knowledge norm ‘assert p only if one knows p’ is primary.
We need not enter this debate here. In particular, in discussing the truth norm, I do not mean to be
claiming (or denying) that it is the primary norm of assertion.
⁵⁷ In §5.2, I said that a sentence is ‘assertion grade’ or ‘true enough to assert safely’ if its degree of
truth is greater than or equal to 0.5. This does not mean that if a sentence S has a degree of truth of 0.5
or greater, then an (ordinary, unhesitant) assertion of S is acceptable. Rather, the idea is that a sentence
is ‘assertion grade’ or ‘true enough to assert safely’ if the level of confidence appropriate in an utterance
of the sentence is at least as high as the level of confidence appropriate in an utterance of its negation.
251
In response, the first point to make is that, faced with a situation and
a predicate, we have many more options than simply to apply or to
withhold the predicate. We can apply the predicate with varying degrees of
confidence or hesitation—i.e. where F is the predicate and a the object, we
can assert Fa with varying degrees of confidence. Second, we ‘‘replace the
old connection between justified assertion and truth’’ as indicated above:
the appropriate (or in Wright’s terms, justified) degree of assertion is the
one corresponding to the degree of truth of the sentence in question.
And now we can see that tolerance does not threaten. For if a and b
are dissimilar only to some very small extent, it does not follow that the
degree of assertion appropriate for Fb is the same as the degree of assertion
appropriate for Fa: it follows only that these degrees are very similar.
5.5 Truth-Functionality
In §2.4 (pp. 85 ff.), I discussed the objection to recursive many-valued views
that their interpretation of the sentential connectives as truth functions does
not cohere with ordinary usage of compound sentences about borderline
cases, and/or with intuitions about the truth of such sentences. This
objection has been pushed strongly against the fuzzy view of vagueness;
Williamson (2003, 694) calls it a ‘‘dark cloud over fuzzy logic’’. The
objection proceeds by way of alleged counterexamples: a particular sentence
is given, and then it is argued that this sentence obviously has a particular
truth value (or assertibility status), whereas on the fuzzy account it has some
other truth value (or some truth value which is not compatible with its
having this assertibility status—e.g. it is unassertible, but the fuzzy theory
gives it a non-zero degree of truth).
One serious problem with all discussions of this issue in the literature
is that none of them is based on proper data about ordinary usage. Some
authors (myself included) have put informal questionnaires to undergraduate
students and non-philosophers; others have simply consulted their own
intuitions. No one has done an adequate study of a sort which would meet
minimal standards for empirical work in linguistics or psychology.⁵⁹ The
lack of reliable data will therefore need to be factored in to our discussion.
The sentences we are interested in include the following, where p is a
colour sample midway between clear red and clear orange:⁶⁰
1. p is red.
2. p is not red.
3. p is orange.
4. (a) p is red or p is not red.
(b) p is or isn’t red.
(c) p is either red, or it isn’t.
(d) Either p is red, or it is not red.
..
.
5. (a) p is red or p is orange.
(b) p is red or orange.
(c) p is either red or orange.
(d) Either p is red, or it’s orange.
..
.
6. (a) It is not the case that p is red or that p is not red.
(b) p is neither red, nor not red.
⁵⁹ Bonini et al. 1999 have undertaken psychological tests of the attitudes of speakers to certain sen-
tences involving vague predicates, but while their study may be unobjectionable from a methodological
point of view, its content is inadequate, for they did not test attitudes to crucial compound sentences,
such as ‘Fa or not Fa’ and ‘Fa and not Fa’, where a is borderline F. They simply assert, without testing
this claim, that ‘x is red and x is not red’ seems clearly false when x is a borderline case of ‘red’—an
intuition I challenged in §2.4.
⁶⁰ Of course we are not specifically interested in colours; in general we are interested in sentences
about borderline cases. Conditionals will be discussed in §5.5.1.
253
⁶¹ For more details on this kind of framework, see Smith 2006, §2.
⁶² I will often, as here, suppress talk of the intended interpretation, for the sake of increased
readability, but it should be kept in mind that a wf by itself is not a content: a wf plus an interpretation
is a content of a sentence. Which interpretation of a given wf is supposed to be the intended one will
always be obvious or immaterial, where not explicitly mentioned.
⁶³ That is, on the interpretation which is intended relative to C.
255
⁶⁴ The latter term is due to DeRose 1999; see also DeRose 2002.
256
papers cited in n. 64). We do not need to enter this debate here. The way
I shall proceed is as follows. Given a sentence S which is intuitively n true
and/or n-assertible in a context C, and where the wf p—which, it seems,
the fuzzy theory must regard as the content expressed by S in C —is m
true (m = n), I will content myself with finding a wf q which is n true, and
which can be regarded as a plausible reading of S in C —i.e. which is such
that it is plausible to say that ordinary speakers hear S as q in C. I shall leave
open the further issue of whether to spell out this idea that speakers read or
hear S as q in C via contextualism or a WAM. The contextualist approach
would be to say that in C, S actually expresses q. The WAM would be
to say that while S expresses p in C, its warranted assertibility condition
in C is such that it is assertible to the degree that q (not p) is true. I think
that in some of the cases to be discussed, a contextualist approach is more
plausible, and in others a WAM is more plausible. However, as long as at
least one approach is plausible in each case, it is unnecessary for present
purposes—i.e. defending the fuzzy view against the truth-functionality
objection—to pursue the question of which approach is more plausible in
each case.
So, to the cases. Consider sentence (8a): ‘‘p is red and p is orange.’’
According to Williamson’s intuitions, (8a) is clearly incorrect.⁶⁵ According
to my observations, speakers generally hedge over (8a) just as they do over
(1) (‘‘p is red’’) and (3) (‘‘p is orange’’), when p is borderline red/orange
(recall §2.4).⁶⁶ Given these conflicting views, I think we must accept that
proper empirical studies might show that almost all speakers find (8a) to be
0-assertible, or that almost all speakers find (8a) to be 0.5-assertible, or that
significant numbers of speakers go each way (in a given context), or that all
speakers go the same way in each context—but different ways in different
contexts. Yet, whatever happens, there is no prospect of a problem here
for the truth-functional degree theorist—for there is a plausible reading
of (8a) which is 0.5 true, and another plausible reading which is 0 true.
The former is simply the surface reading Rp ∧ Op. In order to state the
latter reading, we need to add some symbols to our formal language. Let
us use angle brackets as quotation marks—so where A is a wf, A is
⁶⁵ The actual sentence that Williamson 1994, 136 considers is ‘He is awake and he is asleep’, said
of someone drifting off to sleep. It is not clear whether Williamson means ‘unassertible’ or ‘false’ by
‘incorrect’; but, as we have seen, the challenge can be put either way.
⁶⁶ Forbes 1983, 244 has similar intuitions.
257
⁶⁷ More formally: On the syntactic side, where A is a well-formed formula, A is a term, and
where x is not a well-formed formula, x is not a symbol of the language. T1 is a one-place predicate.
On the semantic side, there are no new requirements on interpretations (i.e. the new vocabulary is not
new logical vocabulary). An interpretation must assign each term of the form A a referent: it is not
required that the referent of this term by the wf A —however, in general, interpretations which lack this
feature will not be intended. Similarly, an interpretation must assign T1 an extension: it is not required
that the extension of this predicate on an interpretation M assign 1 to those things in the domain which
are wfs that are assigned the truth value 1 on M and 0 to all other things in the domain—however, in
general, interpretations which lack this feature will not be intended. For further discussion—including
of how the semantic paradoxes may be handled within a framework of this kind—see Smith 2006.
⁶⁸ I here apply an idea of Keefe 2000, 164. The idea is taken up and developed in great detail by
Weatherson 2004 (cf. also Weatherson 2005, §5), who attributes the idea to Fine 1997 [1975]. Reading
Weatherson’s paper was a great help to my thinking about the truth-functionality objection to the fuzzy
view. Weatherson presents his discussion in terms of a one-place sentential operator ‘determinately’,
rather than in terms of a predicate ‘is definitely true’. My own discussion could also be framed in such
terms—or in terms of predicates Pn which are 1 true of objects of which P is n true, and 0 true of all
other objects; on this approach the second reading of (8a) is R1 p ∧ O1 p.
258
⁶⁹ I am not suggesting that Keefe is confused here—only that her way of putting the point could
confuse others.
259
Each time you do this, the foreman comes along and says, pointing to the
sample, ‘‘That is red or that is orange’’.⁷⁰ Eventually, we may suppose, you
get the message and classify all the samples one way or the other, rather than
setting some aside. Now the truth-functional degree theorist can explain
why we might find ‘‘That is red or that is orange’’ 1-assertible in this
context as follows. In this context—which is rather special in a number of
ways, one being that each sample must be put into a bin, another being that
there are only two bins available—it is very natural to hear the foreman’s
utterance of ‘‘That is red or that is orange’’ as a reminder of precisely these
two points. But of course the truth-functional degree theorist can happily
⁷⁰ In Tappenden’s example the foreman waits until you have a pile of samples, and then says ‘‘Every
one of these samples is either red or orange’’—but we are looking for a case in which the sentence
uttered is of the form ‘‘p is red or p is orange’’, so I adapt the example slightly.
260
accept that the claim which we would express with the sentence ‘‘Each
sample must be put into the ‘red’ bin or the ‘orange’ bin’’ (or ‘‘Each sample
must be put into a bin, and there are only two bins available’’, etc.) is
1 true.
Second, moving (slightly) from sentences of form (5a) to ones of form
(5c), Keefe (2000, 163) gives the following example:
suppose F, G and H are incompatible and a is on the borderline between being F
and being G and is definitely not H: it would then be appropriate and informative
to say ‘a is either F or G’. (E.g. to the question ‘is it red?’ asked of a borderline
blue–green patch, the reply ‘no, it’s blue or green’ is appropriate.)
Now the truth-functional degree theorist can explain why we might find
‘‘It’s blue or green’’ 1-assertible in this context thus: in this context, it is
very natural to hear this as saying that the patch is (definitely) not any
colour other than blue or green. Indeed, Keefe goes on to say precisely
that ‘a is either F or G’ ‘‘can be highly informative by saying something
about the properties of a, in particular, implying that a does not have those
properties that are incompatible with both F and G (e.g. that it is blue
or green and not red)’’ (p. 164).⁷¹ Once again, then, the truth-functional
degree theorist can provide a reading of the sentence at issue which is
both plausible in the context and has a degree of truth which matches the
claimed assertibility status of that sentence. The fact that ‘a is either F or G’
is 1-assertible in this kind of context does not show that the fuzzy theorist
is wrong to treat disjunction as a truth function; it shows that ‘a is either
F or G’ is not always heard as Fa ∨ Ga (relative to the obvious intended
interpretation).
Let us turn to a third alleged problem case for the fuzzy view, sentence
(9a): ‘‘It is not the case that p is red and p is not red.’’ Just as I tend
to hear ‘‘p is red and p is not red’’ as 0.5-assertible when p is borderline
red/orange,⁷² so I tend to hear its negation, that is, (9a), as 0.5-assertible
also. But Weatherson (2004, 1) and others claim that many competent
A way to respect the thought that ‘Jack is bald and Mack isn’t’ is false would
be to say—along now-familiar lines—that this sentence can be read as
T1 Bj ∧ T1 ¬Bm. Now consider the negation, ‘It is not the case that Jack
is bald and Mack isn’t’. We could explain someone finding this 1-assertible
by saying that they hear it as ¬(T1 Bj ∧ T1 ¬Bm). But I think that there is
also another way in which one might hear this sentence, which would also
lead one to regard it as 1-assertible. Here is a context in which this other
reading seems natural. Suppose that for some reason we need to classify
each person in a group as either bald or not bald. Jack and Mack are hard
cases: we find it difficult to decide which way to classify them. While we
are pondering the problem, someone says ‘‘(Well, one’s thing for sure:) It
is not the case that Jack is bald and Mack isn’t’’. This sounds fully assertible.
The fuzzy theorist can happily accept this, explaining it as follows: in this
context we hear this sentence as saying that however we end up classifying
them, Jack and Mack must be classified in the same way (because they are
identical in respect of baldness)—a claim which, we may suppose, is 1 true.
A variation on the previous case is where we have adjacent objects a and
b in the middle of a Sorites series—for example, a is a borderline heap,
and b has one less grain than a—and we say ‘‘It is not the case that a is a
empirical work might reveal that ordinary speakers do in fact consider statements such as ‘‘p is red and
p is not red’’ to be 0-assertible.
262
heap and b is not a heap’’. This is a Sorites premiss, in the sense that one
can present the Sorites paradox using negated conjunctions of this form in
place of conditionals ‘‘If a is a heap, then b is a heap’’ without affecting
how compelling the argument is.⁷³ (The Sorites reasoning then proceeds as
follows: a is a heap—as established at the previous stage of the reasoning;
so it is not the case that b is not a heap; so b is a heap too.) So in this sort of
context, ‘‘It is not the case that a is a heap and b is not a heap’’ compels our
assent—i.e. is highly assertible. The truth-functional degree theorist can
explain this as follows. We have just seen that ‘‘It is not the case that Jack
is bald and Mack isn’t’’ can be heard as saying that, however we end up
classifying them, Jack and Mack must be classified in the same way (because
they are identical in respect of baldness). Similarly, ‘‘It is not the case that a
is a heap and b is not a heap’’ can be heard as saying that a and b must be
classified in the same way (because they are so similar in respects relevant
to being a heap). That is, the sentence in question can very naturally be
heard as an expression of the thought that ‘heap’ conforms to Tolerance.
But then my account in §3.5.4 of why the Sorites is compelling can be
applied here: we explain why someone is inclined to regard ‘‘It is not the
case that a is a heap and b is not a heap’’ as highly assertible in the context
of Sorites reasoning by saying that she hears this sentence as saying that a
and b must be classified in the same way—and so she is inclined to agree,
because her acceptance of Closeness (i.e. her feeling that ‘heap’ is vague)
licenses her to accept Tolerance, and hence this application of it.⁷⁴
Let’s consider a fourth kind of case: (10a). Weatherson (2004, 23) says
that the sentence ‘‘It is not the case that Louis is bald, but nor is it the
case that he is not bald’’ is ‘‘a legitimate, if slightly long-winded, way to
communicate that Louis is a penumbral case of baldness’’. I quite agree that
the sentence sounds 1-assertible when heard in this way, but this poses no
problem for the fuzzy view: for if we are indeed to regard the sentence as
saying that Louis is a borderline case of baldness, we will have to regard it
as saying ¬T1 Bl ∧ ¬T1 ¬Bl (which is 1 true when Bl is 0.5 true), not
¬Bl ∧ ¬¬Bl (which is 0.5 true when Bl is 0.5 true).
The style of my response to the truth-functionality objection has now
been well illustrated. Given that a certain sentence is n-assertible, we need
⁷³ Indeed, Weatherson 2004, §5.3 thinks the paradox is more compelling in the former form.
⁷⁴ For further discussion, see §5.5.1 below.
263
⁷⁵ Note here that the classical theorist who thinks that ‘‘p is red or orange’’ says Rp ∨ Op (relative
to the obvious classical interpretation) and the fuzzy theorist who thinks that ‘‘p is red or orange’’ says
Rp ∨ Op (relative to the obvious fuzzy interpretation) disagree over the content (truth conditions) of
this sentence. The fuzzy theorist who wants to agree with the above classical theorist about the content
of this sentence will have to claim that it says, say, T1 Rp ∨ T1 Op (relative to the obvious fuzzy
interpretation). The point to remember is that we are taking contents to be interpreted wfs, not just wfs;
so it should not be a surprise that when we move from classical to fuzzy interpretations, we might need
to change the wf in order to preserve the content.
264
in the sense of how far its actual degree of truth is from the maximum
truth value. A sentence might be very close to the truth in the first sense,
and yet very far from the truth in the second sense.⁷⁶ For example, if Bob
1
is 6 ft 64 in., then although we would not have to change Bob much to
render the sentence ‘Bob is six feet tall’ true, as things actually stand, the
sentence is simply false. Likewise in Edgington’s second case, b is indeed a
better choice than c, but this is simply because one would have to change
less about b to make Rb ∨ Sb true than one would have to change about
c to make Rc ∨ Sc true—and this has no direct implications for the actual
truth values of Rb ∨ Sb and Rc ∨ Sc.
⁷⁷ See e.g. Forbes 1983, 243–4; 1985, 171–2; and Williamson 1994, 123–4. There is another
approach which says that an argument is valid in fuzzy logic just in case there is no interpretation on
which the conclusion is less true than the least-true premiss (cf. n. 16 above). On this approach, modus
ponens, and the above Sorites argument, come out invalid. See e.g. Machina 1976, 69–75; Williamson
1994, 123–4; and Priest 1998, 332. I side with Williamson, against Machina, in thinking that this
approach does not yield a satisfying explanation of why the Sorites argument is compelling (given that,
on this approach, it is not valid). There are also other variations, such as the combination in Lakoff
1973 of a Machina-type definition of validity with a non-Łukasiewicz conditional, on which modus
ponens comes out valid.
267
That’s a neat story, but it cannot be the correct account of why the
Sorites is both compelling and mistaken. For, as Wright points out, on this
approach the fuzzy theorist is left without anything to say about the Sorites
paradox as formulated in other ways:
Can a degree-theoretic account explain the plausibility of the major premisses?
There is no difficulty, of course, with the usual, quantified conditional form of
premise. The explanation will claim that each instance, Fa → Fa , of (∀x)(Fx →
Fx ) is almost true: that its consequent enjoys a degree of truth ever so nearly but not
quite as great as that of its antecedent. And this claim will then be followed . . . by
a stipulation that the degree of truth of any universally quantified statement is the
minimum of the degrees of truth enjoyed by its instances. . . . But . . . the major
premise doesn’t need to be conditional at all. In the case of the Sorites-series of
indiscriminable color patches for instance, we could just as well take it in the form
(∀x) − [red(x) & − red(x )].
All the ways of making the conditional form of major premise seem intuitively
plausible would be applicable to this conjunctive form. . . . [the degree theorist]
needs to explain . . . with what right such a conjunctive major premise may
be regarded as almost true; otherwise he cannot explain its plausibility, or duly
acknowledge the force of the arguments which seem to sustain it.
(1987, 251–2)
As Wright then goes on to point out, one cannot see how the degree
theorist could give an account on which such a conjunctive major premiss is
almost true: and in any case, on the standard fuzzy account, such premisses
are not almost true. Thus, the standard fuzzy explanation of the plausibility
of the conditional formulation of the Sorites paradox does not extend to
other formulations—yet clearly the various formulations are just stylistic
variants of one fundamental problem. If a solution addresses this problem
only when it is formulated in one specific way, then the solution is not
really addressing the fundamental problem at all. The apparent merit of the
standard fuzzy solution of the Sorites paradox is thus an illusion.
Given that I do not want to adopt the standard fuzzy resolution of the
Sorites, I have no need for a truth definition for the conditional which
renders each premiss ‘If pile i is a heap, then pile i − 1 is a heap’ either
totally true or almost totally true. Furthermore, I have a positive reason for
rejecting this truth definition. In the spirit of wanting a classical consequence
relation, I would also like to retain the equivalence of A → B , ¬A ∨ B ,
268
and ¬(A ∧ ¬B ), and the usual connection between consequence and the
conditional: B is a consequence of A just in case A → B is a tautology.
The Łukasiewicz conditional lacks both these properties (relative to the
definition of consequence that I proposed above). I therefore propose
simply to define the truth conditions for the conditional so that A → B
is equivalent to ¬A ∨ B and ¬(A ∧ ¬B ) (the latter two are already
equivalent).⁷⁸ It is then easy to show that B is a consequence of A just in
case A → B is a tautology.⁷⁹
At this point two tasks remain: (1) to fend off the charge that
the Łukasiewicz conditional provides a better reading of the English
‘if . . . then . . . ’ than does the fuzzy material conditional just defined; (2) to
explain how we are going to resolve the Sorites paradox, given that we
have rejected the standard fuzzy resolution.
(1) It might be thought: doesn’t the standard fuzzy semantics for the
conditional provide a better formal rendition of the English ‘if . . . then . . . ’
than the account proposed above, according to which A → B , ¬A ∨ B ,
and ¬(A ∧ ¬B ) always have the same truth value? For example, consider
Bob, a borderline case of ‘bald’, and Bill, who has one less hair than Bob.
Let us suppose ‘Bob is bald’ is 0.5 true and ‘Bill is bald’ is 0.51 true. Then
‘If Bob is bald, then Bill is bald’ is 0.51 true, according to my semantics,
whereas on the standard fuzzy semantics, it would be 1 true—and isn’t
the latter the more intuitive assignment? In fact, this is not at all clear. In
saying ‘If Bob is bald, then Bill is bald’ one might mean that if one were
to stipulate a sharp boundary for ‘bald’, and it enclosed Bob, then it must
enclose Bill also. That is, one might be saying that if Bob is to count as
bald (under some precisification of ‘bald’), then Bill is to count as bald
too. This claim is certainly something I want to accept—and I can easily
accept it. For when one hears ‘If Bob is bald, then Bill is bald’ as making
this claim about boundary stipulation, one is not hearing it in accordance
with its surface reading; i.e. one is not hearing it as saying Bb → Bl (l for
‘Bill’) relative to the obvious intended interpretation. My account assigns
this wf the value 0.51 on that interpretation; it assigns the sentence ‘If Bob
is bald, then Bill is bald’ the value 0.51 only in so far as this sentence is
given its surface reading. So, suppose that I say ‘If Bob is bald, then Bill is
bald’, and, as we might put it, I mean just that (I simply mean that if Bob
is bald, then Bill is bald). In this case it does not seem that the sentence should
definitely be true! Say I am unwrapping my Christmas presents; I get to a
longish object and say, ‘If this is a spade, I will use it to dig a vegetable
garden’. It turns out to be a two-piece fishing rod, and looking at it, I say
again, ‘If this is a spade, I will use it to dig a vegetable garden’. This is an
odd thing to say—for we can all clearly see that it is not a spade.⁸⁰ Now
suppose you inform me that Bill has one less hair than Bob, and then Bob
and Bill are brought in, and I see that Bob is a borderline case for ‘bald’.
If I now say ‘If Bob is bald, then Bill is bald’ (and I mean just that), then
far from being clearly true, my statement is odd: perhaps not quite as odd
as the spade statement—because it is not clearly false that Bob is bald—but
quite odd nevertheless. Thus, on this occasion—on which it is plausible to
give my sentence its surface reading Bb → Bl —there is no clash between
the intuitive assertibility status of the sentence and the truth value assigned
to the latter wf by my semantics.
I do not think, then, that the Łukasiewicz conditional is a better rendition
of ‘if . . . then . . . ’ than the fuzzy material conditional defined above. The
sort of example considered above, which is often supposed to show that
it is better, does not show this—and other examples show that it is
worse. Suppose that ‘Ben is tall’ is 0.51 true. Then, using the Łukasiewicz
conditional, both ‘If Bob is bald, then Bill is bald’ and ‘If Bob is bald,
then Ben is tall’ are true to degree 1. Presumably, however, someone who
believes that the former is true to a high degree will not also believe that
the latter is true to a high degree. This would indicate that the real source
of the intuition that the former is true to a high degree is understanding
‘If Bob is bald, then Bill is bald’ as meaning that if one were to stipulate
a sharp boundary for ‘bald’, and it enclosed Bob, then it must enclose Bill
also—and this, as noted above, is something I can accommodate.
⁸⁰ Cf. Stalnaker 1999, 94 n. 18: ‘‘an indicative conditional is appropriate only in a context where it
is an open question whether the antecedent is true.’’
270
(2) What are we going to say about the Sorites paradox? Recall that I
have already laid out a recipe for approaching the Sorites paradox, given
a semantics which accommodates predicates satisfying Closeness. I argued
in §3.5.4 that if we suppose that a predicate conforms to Closeness, we
can see both why a Sorites paradox for this predicate is compelling and
also how the paradox is mistaken. The Sorites argument is compelling
because when the major premiss (or premisses) expresses Tolerance, the
argument is valid, and also the major premiss is plausible (to one who
accepts Closeness but not Tolerance) because accepting Closeness licenses
one to accept Tolerance as a useful approximation of what one believes,
in ordinary situations. The argument is flawed, however, because it leads
us to see that situations involving Sorites series are ones in which we
cannot happily use Tolerance as an approximation of Closeness, and
must work with Closeness itself—and when reformulated so that the
major premiss expresses Closeness (but not Tolerance), the argument
is invalid.
Applying this strategy to the Sorites argument under consideration in
the present section yields the following account. We find the paradox
compelling because, first, we naturally hear the conditional premisses ‘If
pile i is a heap, then pile i − 1 is a heap’ as expressing (consequences of)
the claim that ‘heap’ is Tolerant, and, second, the argument is valid on this
reading. The argument is ultimately unconvincing because the argument
itself shows us that we are in one of those special situations where we cannot
safely employ Tolerance as an approximation of Closeness, but must work
with Closeness itself—and when we read the conditional premisses ‘If pile
i is a heap, then pile i − 1 is a heap’ as expressing (consequences of) the
claim that ‘heap’ conforms to Closeness (but not Tolerance), the argument
is not valid.
It will be helpful to state precisely the Tolerance and Closeness readings
of the conditional premisses. In order to do so, we need to add some
symbols to our formal language. Let us use [A ] as a name of the degree
of truth of A .⁸¹ Let us also add an identity predicate = and treat it in a
⁸¹ On the syntactic side, where A is a closed wf, [A ] is a term, and where x is not a closed wf, [x]
is not a symbol of the language. On the semantic side, an interpretation must assign each term of the
form [A ] a referent: it is not required that the referent of this term on an interpretation M be the degree
of truth assigned to A on M—however, in general, interpretations which lack this feature will not be
intended. Cf. n. 67 above.
271
purely classical manner.⁸² Finally, let ≈T denote the relation which holds
between truth values that are very close in respect of truth.⁸³ Now, when I
say that we naturally hear the conditional premiss ‘If pile i is a heap, then
pile i − 1 is a heap’ as expressing (a consequence of) the claim that ‘heap’
is Tolerant, I mean that we hear it as [Hpi ] = [Hpi−1 ]. On this reading, the
full argument is:
1. [Hp10,000 ] = [Hp9,999 ]
2. [Hp9,999 ] = [Hp9,998 ]
..
.
9,999. [Hp2 ] = [Hp1 ]
10,000. [Hp10,000 ] = 1
∴ 10, 001. [Hp1 ] = 1
This argument is classically valid, and hence valid in the version of fuzzy
logic presented in this chapter. When I say that we retreat to reading the
conditional premiss ‘If pile i is a heap, then pile i − 1 is a heap’ as expressing
(a consequence of) the claim that ‘heap’ conforms to Closeness (but not
Tolerance), I mean that we read it as [Hpi ] ≈T [Hpi−1 ]. On this reading,
the full argument is:
1. [Hp10,000 ] ≈T [Hp9,999 ]
2. [Hp9,999 ] ≈T [Hp9,998 ]
..
.
9,999. [Hp2 ] ≈T [Hp1 ]
10,000. [Hp10,000 ] = 1
∴ 10, 001. [Hp1 ] = 1
This argument is not classically valid, and hence not valid in the version of
fuzzy logic presented in this chapter. (We could make a valid argument by
adding a premiss saying that ≈T is transitive—but this premiss would be
clearly false on interpretations on which ≈T has its intended extension.)
⁸² On the syntactic side, = is a binary predicate. On the semantic side, a = b is 1 true if the
referent of a is identical to the referent of b, and 0 true otherwise. This is a new requirement on
interpretations—i.e. we treat identity as an item of logical vocabulary. However, the result about our
fuzzy consequence relation being identical to the classical consequence relation carries over when we
enrich our first-order language with identity in this way. Note that there do exist fuzzy treatments
of identity which (try to) render it a vague relation; for my reasons for thinking that identity must
always be treated as precise—even in the context of a fuzzy semantics which allows other relations to
be vague—see Smith 2008.
⁸³ Recall §3.4.2. ≈T is treated as a non-logical predicate.
272
⁸⁵ Williamson 1997 [1992], 265–6; cf. Williamson 1994, §7.2. Horwich 1998, 76–7 presents a closely
related argument. I use angle brackets here in the way introduced in §5.5.
275
The truth value set of FL is a countable set of the form {true, false, not true,
very true, not very true, more or less true, rather true, not very true and
⁸⁶ Given my account of the conditional, T1 A ↔ A has the same degree of truth as (¬T1 A ∨
A ) ∧ (¬A ∨ T1 A ), which is not a tautology—indeed, it can have an arbitrarily low positive degree
of truth (e.g. it is 0.01 true when A is 0.99 true).
⁸⁷ Given my account of the conditional, TA ↔ A has the same degree of truth as (¬TA ∨
A ) ∧ (¬A ∨ TA ). For this to be less than 0.5 true, one of its conjuncts must be less than 0.5 true,
hence both disjuncts of this conjunct must be less than 0.5 true—but that cannot happen if A and
TA have the same degree of truth, for then one disjunct is the negation of something which has the
same value as the other disjunct.
276
not very false, . . . }, where each element of this set represents a fuzzy subset
of the truth value set of L1 , that is, of [0, 1].⁸⁸ Now consider the following
passage from Haack:
Zadeh offers us not only a radically non-standard logic, but also a radically non-
standard conception of the nature of logic. It would scarcely be an exaggeration to
say that fuzzy logic lacks every feature that the pioneers of modern logic wanted
logic for . . . it is not just a logic of vagueness, it is—what from Frege’s point of
view would have been a contradiction in terms—a vague logic.
(Haack 1979, 441)⁸⁹
Haack notes explicitly that she is concerned with fuzzy logic in the elaborate
sense—not with the view discussed in this book. I think that what Haack
says in these passages is right—but none of it applies to the fuzzy view as
discussed here.
⁸⁸ Here we see that Zadeh’s terminology is somewhat unfortunate: he means by ‘fuzzy subset of
[0, 1]’ a function from [0, 1] to [0, 1], and yet he withholds the term ‘fuzzy’ from [0, 1]-valued logic.
For further details of Zadeh’s more elaborate view, see Smith 2004, §14. On the terminological point:
many writers on fuzzy logic employ a distinction between narrow and broad senses of ‘fuzzy logic’, but
usage of this terminology is not consistent throughout the literature (cf. e.g. Hájek 1998, 2, and Novák
1998, 75).
⁸⁹ Cf. Haack 1978, 167: ‘‘Fuzzy logic, in brief, is not just a logic for handling arguments in which
vague terms occur essentially; it is itself imprecise. It is for this reason that I said that Zadeh’s proposal is
much more radical than anything previously discussed; for it challenges deeply entrenched ideas about
the characteristic objectives and methods of logic. For the pioneers of formal logic a large part of the
point of formalisation was that only thus could one hope to have precise canons of valid reasoning.
Zadeh proposes that logic compromise with vagueness.’’
6
Worldly Vagueness and Semantic
Indeterminacy
determinacy over the degree to which a is F. All predications of ‘‘is red’’ will receive
a unique, exact value, but it seems inappropriate to associate our vague predicate
‘‘red’’ with any particular exact function from objects to degrees of truth. For a
start, what could determine which is the correct function, settling that my coat is
red to degree 0.322 rather than 0.321?
(Keefe 1998b, 571)
One immediate objection which presents itself to [the fuzzy] line of approach is the
extremely artificial nature of the attaching of precise numerical values to sentences
like ‘73 is a large number’ or ‘Picasso’s Guernica is beautiful’. In fact, it seems
plausible to say that the nature of vague predicates precludes attaching precise
numerical values just as much as it precludes attaching precise classical truth values.
(Urquhart 1986, 108)
All three writers voice the same basic objection, which is that the fuzzy
account involves artificial precision: it is highly implausible to suppose that
any given vague sentence is assigned some particular unique real number
between 0 and 1 as its degree of truth, rather than some other nearby
number. Haack simply makes the claim that the assignment of unique fuzzy
values is artificial and implausible, whereas Keefe and Urquhart both back
up this claim with reasons. Keefe’s reason concerns the question of what
could determine the unique correct assignment, while Urquhart’s reason has
to do with the nature of vague predicates.
I endorse the basic claim, and I endorse Keefe’s reason for making it,
but I wholeheartedly reject Urquhart’s reason. Consider the latter point
first. Urquhart claims that ‘‘it seems plausible to say that the nature of vague
predicates precludes attaching precise numerical values just as much as it precludes
attaching precise classical truth values’’ (my emphasis). This is simply false.
As we have seen in Chapters 3 and 4, the nature of vague predicates—as
captured by the Closeness definition—precludes classical semantics, because
in order to accommodate predicates which satisfy Closeness, we need to
countenance degrees of truth. The nature of vague predicates most certainly
does not preclude the attaching of precise numerical truth values, however:
the fuzzy semantic framework is a prime example of a framework which
can accommodate predicates which satisfy Closeness. This illustrates the
importance of including in our investigation a careful analysis of what
vagueness is. If we do not—if we jump straight into the question of the
correct theory of vagueness—then when we find an unattractive feature in
279
¹ Apart from the three passages quoted above, other sources for this objection to the fuzzy view
include Copeland 1997, 521–2; Goguen 1968–9, 332; 1979, 54; Lakoff 1973, 462, 481; Machina 1976,
61; Rolf 1984, 223–4; Schwartz 1990, 46; Tye 1995, 11; and Williamson 1994, 127–8.
280
² Recall that the extension of a predicate is a fuzzy subset of the domain, which is a function from
members of the domain to the set [0, 1] of fuzzy truth values.
³ Telly Savalis has not a hair on his head; Fabio has a thick, flowing mane.
⁴ See Quine 1960, ch. 2; 1969; 1970; Kripke 1982; and Putnam 1983 [1977]; 1978; 1981, ch. 2.
281
⁵ Quine in fact argues that it is, furthermore, underdetermined whether ‘gavagai’ should be translated
as a predicate at all, as opposed to a singular term denoting the fusion of all rabbits or the universal
rabbithood.
282
⁶ Quine does not merely assert this: he has at least three arguments for the publicity premiss—but
they need not concern us here.
⁷ For a famous statement against primitive semantic facts see Fodor 1987, 97. Cf. also Kripke 1982,
51–3. Not everyone accepts step 4: Boghossian 1989, §VI seriously considers primitive meaning facts
(although concerning mental content, rather than meanings of public language expressions) in response
to Kripkenstein’s sceptical problem. Van Cleve 1992, 358 refers to primitivism about the reference of
words as the position of Plato’s Cratylus, and attributes primitivism about the reference of thoughts to
Chisholm and Brentano. See also Hohwy 2001, 2 n. 2 for references to some other possible examples
of endorsement of primitivism about meaning.
283
⁸ Quine, of course, argues for the indeterminacy of translation. However, if one looks closely at
his argument, one sees that ultimately the characterization that I have just given does apply to Quine.
See in particular Quine 1960, 72, where the argument climaxes not in a demonstration that different
sets of analytical hypotheses—or in our terminology, different interpretations—can conform equally
well with the data, but in claims to the effect that this cannot be doubted: ‘‘Both analytical hypotheses
may be presumed possible. Both could doubtless be accommodated by compensatory variations in
analytical hypotheses concerning other locutions, so as to conform equally to . . . all speech dispositions
of all speakers concerned. . . . There can be no doubt that rival systems of analytical hypotheses can
fit the totality of speech behavior to perfection, and can fit the totality of dispositions to speech
behavior as well, and still specify mutually incompatible translations of countless sentences insusceptible
of independent control.’’
⁹ Soames 1997 in effect criticizes the procedure. Recall Ch. 2, n. 10. Soames argues that even
though we cannot see how to derive meaning facts from facts about usage and the environment, it does
not follow that meaning facts do not supervene on such facts—and indeed this would not follow even if
we had assembled all the type-T facts, and could not see how to derive the meaning facts from them.
So, when ‘determines’ is understood in terms of supervenience, or metaphysical determination, Soames
rejects 1 (or rather, the instance of 1 involved in Kripkenstein’s sceptical argument—see below).
¹⁰ Quaddition, symbolised ⊕, is defined as follows (where + is the ordinary addition function, and
k is some number greater than any of the numbers which I have previously added):
x + y if x, y < k
x⊕y=
5 otherwise
284
each pair of natural numbers the sum of those two numbers, rather than
the binary function on the domain which assigns to each pair of natural
numbers the quum of those two numbers. Unlike Quine, Kripkenstein
does not restrict us to publicly observable facts. Thus the argument strikes
at someone who rejects premiss 2 of Quine’s argument—for example,
someone who thinks that mental facts accessible only to the speaker herself
are relevant to determining the intended interpretation of a speaker’s words.
In other words, class T in Kripkenstein’s argument is wider than it is in
Quine’s. Nevertheless, the argument is that facts in the new, wider class
T still do not suffice to determine a unique intended interpretation of
our discourse, and thus the stronger conclusion of indeterminacy—not
underdetermination—of meaning is reached.¹¹
Let us now return to the problem for the fuzzy view of vagueness with
which we began. Our problem can readily be seen to be a specific instance
of the abstract argument isolated above—and furthermore, an instance that
is virtually irresistible. This will lead to the conclusion that there is no
fact of the matter concerning which fuzzy interpretation of a given vague
discourse is the unique intended one—that is, to the conclusion that the
discourse does not in fact have a unique intended fuzzy interpretation.
Hence, contra the fuzzy picture as presented above, there will not be a
fact as to the precise degree of truth of a statement such as ‘Bob is bald’
(i.e. its degree of truth simpliciter, as opposed to its degree of truth on this
or that interpretation). Let discourse D be any discourse involving vague
predicates. As for type T, there is widespread agreement concerning the
sorts of facts it should contain:
• All the facts as to what speakers of D actually say and write, including
the circumstances in which these things are said and written, and any
causal relations obtaining between speakers and their environment.
• All the facts as to what speakers of D are disposed to say and write in
all kinds of possible circumstances.
• All the facts concerning the eligibility as referents of objects and sets.¹²
¹¹ More subtly, the conclusion is that we must accept indeterminacy if we retain the truth-conditional
picture of meaning embodied in the project of assigning interpretations—of the model-theoretic sort
we have been considering—to languages. Kripkenstein thinks that (sufficient) determinacy can be
regained if we reject this picture altogether and adopt a conception of meaning based on assertion
conditions rather than truth conditions.
¹² See Merrill 1980 and Lewis 1999 [1983].
285
¹³ Elsewhere I argue for a solution to the problem of reference according to which the intended
interpretation of a discourse is the simplest one which satisfies the constraints imposed by usage—where
‘simplest’ is cashed out precisely, in terms of Kolmogorov complexity. I argue that this approach
promises a solution to the sceptical arguments of Kripkenstein and others. I am not convinced that
extending this approach to the current situation would lead to the conclusion that there is a unique
intended fuzzy interpretation of vague discourse—hence the need to explore what happens if we deny
this conclusion (see below).
286
value near (or equal to) 0. Any candidate correct interpretation that
assigns the predicate ‘is tall’ a function which maps a to a value near 1
must assign the predicate ‘is short’ a function which maps a to a value
near 0.
This hardly scratches the surface: there is a vast number of constraints of
this sort. Some of them are general: for example, if F is (intuitively) a
vague predicate, then any candidate correct interpretation must be such
that if a and b (in the domain of that interpretation) are very close in
F-relevant respects, then Fa and Fb are very close in respect of truth (on
that interpretation). Some are specific to a single predicate: for example,
any candidate correct interpretation must assign to ‘is bald’ a function
which maps Telly Savalis to 1. Consider all the constraints of this sort,
generated by all the type-T facts. If an interpretation is not ruled out as
incorrect by these constraints, then it will be acceptable. The upshot of our
earlier discussion is that many fuzzy interpretations of our ordinary vague
discourse are acceptable.
This gives us fuzzy plurivaluationism—a view which stands to fuzzy
models in precisely the way plurivaluationism (discussed in §2.5) stands
to classical models. (It does not give us fuzzy supervaluationism, which
was discussed in §2.4.1.) Fuzzy plurivaluationism has no further semantic
machinery, besides the fuzzy interpretations. There is no further semantic
story, beyond that told by the many acceptable interpretations. There
are many candidate correct fuzzy interpretations, any of which would
do perfectly well (i.e. meet all the constraints imposed by all the type-
T facts), and there is nothing to decide between them. Our vague
language has many acceptable fuzzy interpretations simultaneously, and as
far as semantics is concerned, that is the end of the story. So we have
indeterminacy, or plurality, of meaning. Unlike in the supervaluationist
picture, the language is not in a unique (higher-order) semantic state.
Semantic states are individuated by interpretations, and there are many
of them. This multiplicity of fuzzy interpretations is the full and final
semantic story.
That said, if a sentence has a certain degree of truth, say 0.3, on every
acceptable interpretation, then we can talk as if there is just one intended
interpretation, on which the sentence is 0.3 true. However this is just talk.
We are not doing something analogous to the supervaluationist: distilling a
288
¹⁴ See e.g. Williamson 1994, 127; Copeland 1997, 522; and Priest 1998, 331–2.
291
several different contents, and no one of these is more natural than all the rest.
In these cases there will be many unnatural contents not eliminated by our
dispositions which naturalness does manage to eliminate, but there will still be
many contents left uneliminated. For example, as far as our dispositions to usage
go, ‘tall woman’ might denote any one of the following properties: woman taller
than 1680 mm, woman taller than 1681 mm, woman taller than 1680.719 mm,
etc. And it does not seem that any of these properties is more natural than any
other. Hence there is no precise fact about what the phrase denotes. Hence it
is vague. In sum, our dispositions are never enough to settle the content of a
term. In some cases, such as ‘water’, ‘rabbit’, ‘plus’, ‘brain’ and ‘vat’, nature is kind
enough to finish the job more or less. In others it is not, and vagueness is the
result.
drawn (in my terms, the jolt problem), and the indeterminacy which we
posit in response to the boundary-determination question is separate from
the fuzziness which we posit in response to the nature-of-the-boundary
question.
Classical plurivaluationism posits only semantic indeterminacy. As we
have seen in Chapters 2 and 4, it solves the location problem faced by
epistemicists—but it runs headlong into the jolt problem. (In the terms
of the previous paragraph, it solves the boundary-determination prob-
lem, but not the nature-of-the-boundary problem.) No classical model is
an acceptable interpretation of vague language (although it might be an
admissible precisification of an intended non-classical interpretation). Adding
more classical interpretations—that is, positing many acceptable ones, rather
than one intended one—solves the problem of how our practice could
associate our language with a unique interpretation (by denying that
there is a unique intended interpretation), but it goes no way at all to
solving the problem that the account cannot accommodate vagueness.
The fuzzy account (with a unique intended fuzzy interpretation), on the
other hand, solves the jolt or nature-of-the-boundary problem—it accom-
modates predicates with blurred boundaries, it accommodates vagueness
(understood in terms of Closeness)—but it runs into the location or
boundary-determination problem. Fuzzy models are acceptable as interpret-
ations of vague language. However, there is still the problem of how our
practice could associate our language with a unique (fuzzy) interpretation.
By combining both semantic indeterminacy (in the absence of a unique
intended interpretation) and worldly vagueness (in the fuzzy interpretations
themselves), fuzzy plurivaluationism solves both the location problem and
the jolt problem.
Let us take stock. I have described a modification of the fuzzy position,
in response to the problem of artificial precision. The problem was
that it is highly implausible to suppose that any given vague sentence
is assigned some particular unique real number between 0 and 1 as its
degree of truth, rather than some other nearby number. It just seems
ridiculous to suppose that ‘This cup is red’ is true to degree 0.678,
rather than 0.677 or 0.679, or even perhaps 0.6 or 0.7. I have done two
things. First, I have discussed the source of this intuition. The problem
is not that assigning a unique fuzzy truth value violates the nature of
vagueness. Rather, the problem is that we cannot see what could possibly
293
crops up amongst both friends and foes of the fuzzy approach. Keefe (1998b,
570) writes:
The comparatives corresponding to multi-dimensional predicates typically have
indeterminate instances. For example, there are pairs of people about whom there
is no fact of the matter who is nicer (more intelligent), nor that they are equally
nice (intelligent): the vagueness of ‘‘nice’’ (‘‘intelligent’’) is such as to leave the
question unsettled.
The point does not apply only to multi-dimensional predicates: Keefe also
discusses the example of ‘‘a is tall’’ vis-à-vis ‘‘b is red’’. Her conclusion is that
‘‘we cannot assume that there is always a fact of the matter about which
of two borderline sentences is more true’’ (Keefe 1998b, 570). Williamson
(1994, 128) makes the same objection:
Consider . . . a purely comparative statement:
(#2 ) ‘It is wet’ is truer than ‘It is cold’.
(#2 ) is vague. . . . In many contexts it is neither clearly true nor clearly false,
attempts to decide it can founder in just the way characteristic of attempts to
decide ordinary vague statements in borderline cases, and so on. What needs to be
acknowledged is the vagueness of . . . (#2 ).
Suppose that our practice does not determine that the sentence ‘Bob is
bald’ should be 0.3 true rather than 0.4 true; nor does it determine that
the sentence ‘Harry is nice’ should be 0.3 true rather than 0.4 true. There
is still an open question: what, if anything, does our practice determine
about the relative degrees of truth of ‘Bob is bald’ and ‘Harry is nice’?
Consider, first, some uncontroversial cases. Bill has some hair, but not
much, and is a little over six feet in height. Ben’s hair is just starting to
thin a bit, and he is a little under six feet in height. Our practice does
not determine that ‘Bill is tall’ is 0.95 true rather than 0.97 true, or that
‘Bill is bald’ is 0.4 true rather than 0.45 true, or that ‘Ben is tall’ is 0.93
true rather than 0.95 true, or that ‘Ben is bald’ is 0.05 true rather than 0.1
true. It does, however, determine that ‘Bill is tall’ is truer than ‘Ben is tall’,
that ‘Bill is bald’ is truer than ‘Ben is bald’, and that ‘Bill is tall’ is truer
than ‘Ben is bald’. The last case is interesting. Bill’s tallness has nothing
whatsoever to do with Ben’s baldness; yet it is intuitively obvious that Bill
is closer to being definitely tall than Ben is to being definitely bald. So our
practice determines not only (some) ordering facts between sentences about
the same subject matter—as in the first two cases just mentioned—but
also (some) ordering facts between sentences about unrelated subjects (e.g.
Bill’s tallness and Ben’s baldness). Now the question is, does our practice
determine a relative ordering (in respect of truth) of all sentences—that is,
does it determine, for any two sentences whatsoever, either that they are
exactly the same in respect of truth, or that one is strictly more true than
the other? The objection currently under consideration says No. Suppose
this is correct. In particular, suppose that our practice does not determine
anything about the relative truth of ‘Bob is bald’ and ‘Harry is nice’. In that
case, some acceptable interpretations will assign these sentences the same
degree of truth, some will assign the former a higher degree of truth than
the latter, and some will assign the latter a higher degree of truth than the
former. As none of these interpretations is any more or less correct than the
others, it will then turn out that we cannot say that ‘Bob is bald’ and ‘Harry
is nice’ are exactly as true as one another, or that one is truer than the other.
In the circumstances, this is the desired result. On the other hand, suppose
that in fact our practice does determine something about the relative truth
of ‘Bob is bald’ and ‘Harry is nice’—say, that the former is truer than the
latter. Then, as a matter of fact, on every acceptable interpretation, ‘Bob
is bald’ is truer than ‘Harry is nice’. (On some interpretations they are
296
0.31 and 0.3 true, respectively, on others 0.33 and 0.31 true, on others
0.4 and 0.3 true, etc.) Then, while neither sentence has a unique degree
of truth, we can say that one sentence is truer than the other (because
this holds on every acceptable interpretation). If this is the case, then the
linear ordering objection to the original fuzzy account is mistaken—but
in this case, my fuzzy plurivaluationist account agrees with the original
fuzzy account on precisely the point which the objector has (ex hypothesi)
got wrong.
In fact, however, while there are some analogies between what we are
doing and the process of measurement, the overall idea is quite different.
When we measure the length of a piece of wood, say, we can assign
this length the number 30 (centimetres) or the number 12 (inches) or the
number 3 (hands), but there is nothing indeterminate about the length
of the wood in itself. The piece of wood has one, unique length, and
we are simply naming this length in different ways, using different systems
of length-names.¹⁹ In the present case, on the other hand, where we are
¹⁷ But see n. 22 below. ¹⁸ Cf. Sanford 1975, 29, and Machina 1976, 61.
¹⁹ This is why I stressed in §5.1 that we get a clearer picture of measurement if we countenance
lengths and so on, rather than just considering mappings directly from objects to real numbers. NB:
298
assigning vague sentences degrees of truth, and these assignments are not
unique, the whole point is precisely that there is indeterminacy as to how
true the sentence is, not simply as to how to name its amount of truth,
this being an entirely determinate matter in itself. If we were to apply the
measurement model to the present case, we would end up with a quite
different picture from the one I have presented. The idea would be that
the different degrees of truth which we assign to sentences on the different
acceptable interpretations are different names for the one unique, unvarying
amount of truth possessed by each sentence. This would suggest a picture
in which our truth values are not the elements of [0, 1] at all. Rather, the
real truth values are these unvarying amounts of truth—and they are simply
named by real numbers. But this is not at all the picture we want. Our
different acceptable interpretations are not different acceptable descriptions of
one unique, underlying semantic reality—they are, to put it dramatically,
different semantic realities, each equally real. We have genuine semantic
indeterminacy here, not a choice as to how to describe one determinate
situation.
Conversely, if we were to apply our own picture to the case of
measurement, we would end up with something quite different from the
standard story about measurement. Instead of a stick which can be assigned
30 (centimetres) or 12 (inches), we would have a stick whose length was
inherently indeterminate, which could be assigned 29 or 30 or 31 centimetres.
That is, the different assignments would not represent different systems for
assigning lengths, but different lengths within the same system. Now
suppose we were trying to measure such inherently indeterminate sticks.
We would need some additional machinery to handle the non-uniqueness
of assignments within one measurement system, as well as the machinery
(from standard measurement theory) for handling the non-uniqueness of
acceptable measurement systems. Now there is no reason to think that
this extra machinery would simply be the same as the original machinery,
applied again. Indeed there are reasons, which will emerge below, why this
will not work. But this means that our case—where the indeterminacy of
assignments of truth value is analogous not to a multiplicity of acceptable
measuring systems, but to indeterminacy of the correct length assignment
All the systems of length-names use the same names—the real numbers—but they assign these names
according to different rules.
299
within one such system—is not best handled by the machinery of standard
measurement theory.
This point will come out more clearly if we examine the issue of
different sorts of measurement scale. Consider transformations of the real
numbers (functions from the reals to the reals). These fall into various
types, for example:
• A function φ is a similarity transformation if there exists a positive real
number α such that for every x, φ(x) = αx.
• A function φ is the identity transformation if for every x, φ(x) = x.
• A function φ is a monotone transformation if whenever x < y, φ(x) <
φ(y).
• A function φ is a linear transformation if there exists a positive real
number α and a real number β such that for every x, φ(x) = αx + β.
Now, suppose that we are measuring a particular attribute A. We have a
mapping f from objects to the real numbers, with the number assigned
to an object representing its quantity of A. Suppose that there are other
mappings f1 , f2 , f3 , . . . from objects to the real numbers which would serve
just as well to measure the attribute A. We now determine the relationships
between these mappings. There will be a type of transformation T such
that if fi is an admissible mapping, then there is a T transformation φ with
fi = φ ◦ f . According to the type of transformation T involved, we say that
the attribute is measured on this or that type of scale. For example:
• Where T is a similarity transformation, we say that A is measured on
a ratio scale.
• Where T is the identity transformation, we say that A is measured on
an absolute scale.
• Where T is a monotone transformation, we say that A is measured on
an ordinal scale.
• Where T is a linear transformation, we say that A is measured on an
interval scale.
Thus mass is measured on a ratio scale: mass as measured in kilograms (say)
may be converted to mass as measured in pounds (say) by multiplying by
a constant. Counting is an example of an absolute scale: when we want
to measure how many objects there are in a given collection—i.e. the
cardinality of the collection—there is only one admissible way of assigning
300
²⁰ For more details on measurement theory see Coombs et al. 1954; Suppes and Zinnes 1963; and
Krantz et al. 1971.
²¹ Recall the discussion in §3.4.1.
301
in the way described above, no matter what the conditions on T. This brings
us back to the point made earlier, that at a fundamental conceptual level,
what we are doing is different from measurement. We are not (to repeat)
assigning different numbers to a sentence as different ways of naming its
fixed amount of truth: we are assigning it different amounts of truth. The
analogous case in the field of (say) length measurement is not (again to
repeat) where we assign a piece of wood both the numbers 12 (inches) and
30 (centimetres), but where we have a piece of wood whose actual length
is indeterminate, and where we might, say, assign the wood any number
between 29 and 31 centimetres. Now the fact that we may admissibly assign
this piece of wood any such number of centimetres does not mean that
if this other piece of metal—also inherently indeterminate in length—is
admissibly assigned the length 30 centimetres, then automatically it is
also admissibly assigned any length between 29 and 31 centimetres. The
length of the metal might be less indeterminate than that of the wood: it
might ‘fluctuate’ only between 29.99 and 30.01 centimetres! This is the sort
of thing we have in the case where we assign acceptable degrees of truth to
sentences. Perhaps two sentences can both acceptably be assigned the value
0.5. This does not mean that if one is also acceptably assigned 0.4, so is the
other. That does not seem to me to be something that we want to build
into the very fabric of our theory. Thus I reject the idea that the acceptable
interpretations must be related by some sort of T transformation on [0, 1]
in the way indicated above.²²
²² It might still be the case that for a particular predicate, the different acceptable extensions of this
predicate are related by some sort of transformation of [0, 1]—with different transformations being
appropriate for different predicates (e.g. ‘short’ is measured on one kind of scale, while ‘beautiful’ is
measured on another kind of scale). This may have been what Goguen had in mind in the passage
quoted above; cf. also Norwich and Turksen 1982.
302
the reals are a suitable structure for measuring the quantity in question
in the first place. This is not automatic. For example, certain preference
orderings amongst options cannot be represented by assignments of real-
number utility values to options.²³ Thus we have a quantity—how much one
likes or values something —which cannot (in certain situations) be measured
using real numbers at all (as opposed to being able to be measured by them,
but not in just one unique way—as is the case with the quantities of heat,
mass, and so on).
In the present case, I have argued that there is not (in general) just one
correct assignment of truth values to sentences—in other words, there
is not just one intended interpretation of a discourse. This corresponds
to the uniqueness part of the measurement project. What about the
representation part? The worry might be that no fuzzy interpretation is
acceptable. Consider, for example, the linear ordering worry. Someone
gripped by this worry might think that my answer was not good enough.
Recall my answer: on each acceptable interpretation, all sentences are
linearly ordered in respect of truth (each acceptable interpretation is a
standard fuzzy interpretation, and the linear ordering worry arises precisely
because such interpretations linearly order sentences in respect of truth);
however, it might be that different acceptable interpretations order two
given sentences differently—one makes p truer than q, one makes them
exactly the same in respect of truth, and one makes q truer than p—so
that, overall, no one ordering can be said to be correct. The objector might
think, however, that any interpretation which linearly orders all sentences
in respect of truth is thereby unacceptable. For nothing about our practice
mandates that p should be truer than q, or the same as q in respect of
truth, or less true—so are not all three sorts of interpretations mentioned
above (ones which make p truer than q, the same in respect of truth, or
less true than q) therefore unacceptable? If this were correct, there would
be no acceptable fuzzy interpretations of any vague discourse, which would
mean, on my approach, that all vague statements mean nothing.²⁴
As I stressed earlier, my view is that the true semantic story regarding
vague sentences is that (in general) they have many acceptable fuzzy inter-
pretations. This plurivaluationist view contrasts with the supervaluationist
view that the true semantic story regarding vague sentences is distilled from
their many admissible precisifications, where none of these precisifications
is a correct interpretation of a vague sentence as it stands. So in response
to the present worry, I would not want to concede that my fuzzy models
are not acceptable interpretations of vague sentences as they stand. But I
do not have to: the solution lies elsewhere. Distinguish what is acceptable
in an interpretation from what is mandatory. Our practice (together with
the facts about eligibility, simplicity, and so on) imposes certain constraints
on correct interpretations of our discourse—for example, that they must
not assign to ‘is bald’ a function which maps Telly Savalis to a value other
than 1; that they must not assign functions to ‘is red’ and ‘is orange’ which
both map some object in the domain to 1; and so on. Anything directly
required by these constraints is mandatory; anything not ruled out by the
constraints is acceptable. Now what I want to say is that nothing about
our practice requires a particular ordering of ‘Bob is bald’ and ‘Harry is
nice’ in respect of truth (it does not require that the former is truer than
the latter, or that the latter is truer than the former, or that the two are
equally true), but that any particular ordering is acceptable—that is, nor
does our practice require that they should not be ordered in any particular
way (it does not require that it is not the case that the former is truer
than the latter, it does not require that it is not the case that the latter
is truer than the former, and it does not require that it is not the case
that the two are equally true). Intuitively, this seems right: we have not
mandated that these sentences should be ordered in any particular way, but
neither have we mandated that they must be incomparable. When we first
consider the linear ordering worry for the standard fuzzy view—where
that view tells us that on the unique intended interpretation, ‘Bob is bald’
is (say) strictly truer than ‘Harry is nice’—we think: we did not mandate
this ordering; what we have fixed leaves it open which sentence is truer;
what is wrong with this other interpretation, on which ‘Harry is nice’
is strictly truer than ‘Bob is bald’? This other interpretation seems just
as good as the one which the fuzzy theory deems the unique intended
interpretation—it does not seem just as bad (because it linearly orders
sentences in respect of truth). On the contrary, the whole point was that all
these interpretations are equally compatible with our constraints on correct
interpretations, so the fuzzy theory has no business saying that only one of
them is correct.
304
²⁵ See e.g. Wright 1987, 254–6; Tye 1996, 219; and Keefe 2000, 113. Cf. also Wright 1975, 349–50;
Sainsbury 1991, 169; and Williamson 2003, 695, who refers to what I am calling the ‘sharp boundaries
problem’ as the ‘problem of higher-order vagueness’.
305
interpretation. So, while there is no man such that we can talk as though
he is the last bald man, there is a last man such that we can talk as
though he is bald. So in this sense, there is still a last bald man, according
to fuzzy plurivaluationism.
One obvious way of trying to avoid this outcome is by saying that the
notion of an acceptable interpretation is vague. In §6.2.1, I discuss and
reject this approach. I then argue, in §6.2.2, that the ‘problem’ of the last
bald man is not a problem at all. There are two sorts of reason why one
might object to the existence of a last bald man: reasons based on the
nature of vagueness and reasons based on the nature of the facts which
determine meaning. Neither counts against the existence of a last bald
man. So we may happily accept that there is a last bald man in our Sorites
series.
The way in which a story along these lines would help with the last bald
man problem is as follows. If the notion of an acceptable interpretation is a
vague one, then there is some vagueness as to who is the last man who is
1-bald on every acceptable interpretation. This means that we do not have
to accept a precise cut-off between the last man who is 1-bald on every
acceptable interpretation and the rest of the men in the series.
²⁶ For discussions or mentions of similar or related views, see Cook 2002; Edgington 1997, 297, 310;
Field 1974, 227; Horgan 1994, 160; Keefe 2000, 117–21; McGee and McLaughlin 1995, 238; Rolf
1984, 222; Sainsbury 1997 [1990], 260; Tye 1990, 551; 1997 [1994], 287; 1995, 16; 1996, 219–20; Varzi
2001, 58–9; and Williamson 1994, 130; 2003, 695.
306
I agree entirely. I understood fuzzy model theory in the first place because
I took it to be a piece of standard mathematics. I understood the definition
of a fuzzy set as a function from a background set to [0, 1], the definition
of a model as a set of objects together with a function from items of our
standard first-order language to elements of the domain (etc.), the recursive
definition of the truth value of a closed formula on an interpretation,
and so on, to have their place alongside the definitions and constructions
that one finds in standard mathematics textbooks. These definitions and
constructions are all presented in a precise language, governed by classical
logic and semantics, of which every competent mathematician has a
working understanding. Now if you turn around at the end of your
presentation of fuzzy model theory and tell me that the language in
which you made your presentation was governed by the very semantics
that you just presented, then I have to say that I did not understand
your presentation at all. I am back at square one. I thought you were
presenting a piece of normal mathematics, and I know how to understand
that sort of thing. If you were not, then I do not know what you
were doing, and I do not know how to understand it. Of course, what
you say does conjure up some sort of picture for me—it is not as if
you uttered complete nonsense—but it is not the sort of perfectly clear
picture I get when I work through and understand a piece of normal
mathematics. And this is not because I have not worked hard enough:
it is because—by your lights as well as mine—what you are doing is
not presenting a piece of normal mathematics, in normal mathematical
language.²⁷
one can only make precise remarks about those vague utterances. Since the
expressive limitations of such a meta-language render it incapable of giving the
meanings of object-language utterances, it can hardly be regarded as adequate
for a genuine semantic treatment of the object-language. . . . the formality of the
semantics [comes] at the cost of giving up the central task of genuine semantics:
saying what utterances of the object-language mean.
‘‘Saying what ‘Bob is bald’ means’’ might mean making a (different) claim
which has the same content as ‘Bob is bald’, or it might mean giving an
account of the semantic relations between parts of the sentence ‘Bob is
bald’ and parts of the world, and of how these combine to determine the
truth status of the whole sentence—i.e. it might mean (as I put it earlier)
saying what is going on between a person and the world when she says
‘Bob is bald’. It is only in the latter sense of ‘saying what utterances mean’
that it is the central task of genuine semantics to say what utterances of the
object language mean. But fulfilling this task involves, precisely, making
remarks about utterances of the object language (as Williamson puts it),
rather than making remarks which have the same content as utterances of the
object language. Semantics is not translation. So even if vagueness is not
eliminable—if no claim in a precise language has the same content as some
vague claim—this does not threaten the capacity of a precise metalanguage
to say what vague utterances mean (in the second sense indicated above),
and hence to carry out the central task of genuine semantics. Indeed, quite
the contrary, as I have argued. We do not understand the relationship
between vague language and the world, and we want an explanation of it.
The proponent of fuzzy semantics presented in a vague metalanguage appears
to give us such an explanation—that is, fuzzy model theory—but then
she adds at the end that we must understand that the relationship between
what she has said and the relationship between vague language and the world is the
same as the relationship between vague language and the world. But what
is this relationship? Our problem was precisely that we do not have a clear
theoretical understanding of this relationship, and the fuzzy metalinguist
has not provided such an understanding.
Fb, even though a and b are very close in F-relevant respects. In this sense,
both the classical and the fuzzy views clearly involve sharp boundaries. But
that’s no problem: any non-vague theory of vagueness will involve sharp
boundaries in this sense (unless it assigns the same semantic status to Fx for
every object x, in which case it is a hopeless theory), and, as discussed in
the previous section, we want a non-vague theory of vagueness.
Letting B(n) be ‘A man with exactly n hairs on his head is bald’, and
Val(S) be the degree of truth of the sentence S, Schwartz (1990, 46) writes:
Consider the set {B(108 ), B(108 − 1), . . . , B(0)} . . . since the set . . . is finite and
the values are well-ordered, there must be a first n such that Val(B(n)) > 0. In
other words there will be a precise, to the hair, dividing line between definitely
non-bald and borderline non-bald. This means that there is an absolute precision
implied by the degrees of truth approach that is inconsistent with the vagueness
of ‘bald’. A precise and unknown dividing line between definitely non-bald and
borderline is just as contrary to the vagueness of ‘bald’, and just as unbelievable,
as would be a precise dividing line between non-bald and bald. The same kind of
argument could presumably be repeated for any vague term.
This is precisely what I deny. The dividing line in the fuzzy picture between
definitely non-bald and borderline is not just as contrary to the vagueness of
‘bald’ as would be a precise dividing line between non-bald and bald—for,
unlike the latter sort of dividing line, the former sort does not involve a
violation of Closeness.
A variant of the last bald man objection says that the fuzzy logician is
committed to a threefold division between the sentences true to degree 1,
those true to degree 0, and the remainder—and by thus positing sharp lines
between the clear cases and the borderline cases, and the clear countercases
and the borderline cases, falls foul of the original problem of higher-order
vagueness (as described in §3.5.5).²⁸ My response to this objection is the
same as before. The force of the original problem of higher-order vagueness
against theories which approach vagueness in terms of three categories (e.g.
true, false, and undefined) stems from the fact that if we have only three
categories, we cannot accommodate Closeness: these three categories bring
with them boundaries which are sharp in the first sense outlined above.
The fuzzy theory does not face this problem: for, while we may distinguish
borderline cases of ‘bald’ (say) from objects assigned 0 or 1 by the fuzzy
set of bald things, we do not get boundaries between the clear cases and
the borderline cases, and the clear countercases and the borderline cases,
which are sharp in the objectionable, Closeness-violating sense. Thus we
do not face the problem of having to blur these boundaries by moving to
a higher level of borderline cases, i.e. the original problem of higher-order
vagueness.
Considerations to do with the nature of vagueness do not, then, count
against the existence of a last bald man in the original fuzzy view. Nor do
such considerations count against the existence, in my fuzzy plurivaluationist
view, of a last man who is 1-bald on every acceptable interpretation. On
my view, ‘is bald’ satisfies Closeness on every acceptable interpretation, and
hence we can say overall that it satisfies Closeness. Thus, the existence of a
last bald man does not threaten the vagueness of ‘is bald’ on my view.
What about considerations to do with the determination of meaning?
Someone might respond to my fuzzy plurivaluationist view that not only
does our practice not suffice to fix a unique intended interpretation of
our discourse: it does not even suffice to fix a unique set of acceptable
interpretations. For after all, what is it about our practice that could make
an interpretation which assigns ‘is bald’ a function which maps Bill to 0.99
acceptable, but an interpretation which assigns ‘is bald’ a function which
maps Ben—who differs from Bill by just a hair—to 0.99 unacceptable?²⁹
Let’s take stock of the situation. On the one hand, we have a vague
discourse. On the other hand, we have all the fuzzy interpretations of
it. The question is, what does the practice of speakers of the discourse
determine concerning which of those interpretations are acceptable and
which incorrect? First, an undeniable fact is that our practice does not
determine nothing. It is not in doubt that some interpretations are incorrect:
for example, any interpretation which assigns ‘is bald’ a function which
maps Telly Savalis to 0. Second, the upshot of §6.1 was that our practice
does not determine a unique intended interpretation. From here, we have
several options.
One option is to accept that there are no acceptable fuzzy interpretations.
This means that either vague language is meaningless, or its correct
interpretation is not a fuzzy interpretation at all. The former option
is a non-starter. The latter option does not help: if we replace fuzzy
interpretations with some other sort of interpretations, we will still face the
problem of how our practice determines a unique one, or set, of these as
correct.
A second option is to posit a fuzzy set of acceptable interpretations.
But this does not help—indeed, it makes the problem worse. Suppose
we admit degrees of correctness for interpretations. There would still be a
last bald man: the last man who is 1-bald on every interpretation which
is 1-correct. All we have done is make the determination problem harder:
now our practice has to determine, for each interpretation, not simply
whether it is acceptable, but to what degree it is acceptable! The same
comments would apply if we opted for some sort of ramified or iterated
fuzzy set of acceptable interpretations: this move would, if anything, make
the determination problem harder.
A third option would be to say that there is something vague about
the set of acceptable interpretations, but refuse to specify exactly what one
means by this (e.g., one does not cash this out as meaning that the set of
acceptable interpretations is a fuzzy set). This proposal is to be avoided, for
the sorts of reason given in §6.2.1: it sacrifices the clarity and perspicuity
that we get with a non-vague account of the semantics of vagueness.
The fourth and final option is to say that our practice determines a
classical/crisp/precise set of acceptable interpretations. Now in opposition
to this view, we have the following instance of the argument structure
from §6.1:
1. Facts of type T do not determine a unique set of acceptable interpreta-
tions of discourse D.
2. No facts of any type other than T are relevant to determining the
acceptable interpretations of D.
3. From (1) and (2): All the facts together do not determine a unique set
of acceptable interpretations of D.
4. It cannot be a primitive fact—i.e. a fact not determined by other
facts—that some interpretation M is an acceptable interpretation of D.
5. From (3) and (4): It is not a fact at all that D has a unique set of
acceptable interpretations.
I do not accept the conclusion, and I think that step 1 is the false premiss.
Suppose we conduct a survey in which we ask speakers to rate the degree
of baldness of certain men. The survey runs through a Sorites series of
313
One might think that a problem remains. Given a set of speakers, a crisp
set of acceptable interpretations emerges, in the way just discussed. But
if we were to survey different speakers, or the same speakers tomorrow,
we would probably get a different set of acceptable interpretations. But
then it seems that we cannot, once and for all, fix on a unique set of
acceptable interpretations—and hence cannot, once and for all, fix on a
last man in a given Sorites series who is bald according to every acceptable
interpretation. My response to this is that it is a true observation, but by no
means an objection: to think so would be to misunderstand our goal. The
challenge was to show how our practice could fix a unique set of acceptable
interpretations of our words. There is a parameter here: us (‘our practice’),
that is, a group of speakers. My goal was not to eliminate this parameter,
but to show how—given a set of speakers—it is plausible to think that
their practice could determine a unique set of acceptable interpretations
of their words. The goal was not to show that a unique set of acceptable
interpretations can be fixed on ‘once and for all’—that is, even in the
absence of a designated group of speakers—and so the observation that this
has not been achieved is no objection. Note that it is not simply that my
goal happened not to be to eliminate the ‘group of speakers’ parameter, and so
this is not an objection to me—but furthermore, that would in fact be a bad
choice of goal. The guiding idea behind all our discussions of problems of
the determination of meaning (beginning in §2.1.1) has been that language
is a human artefact, and that the meanings of terms depend essentially on
the ways in which speakers use them. From this perspective, the idea of
some words having some meanings in abstraction from a set of speakers whose
practice confers upon them those meanings is a non-starter. Rather, talk of the
meanings of some terms must always be relative to a group of speakers,
whose dispositions regarding the usage of those terms plays an essential part
in fixing those meanings.³⁰
³⁰ In the case of vague terms such as ‘tall’ and ‘bald’, I have assumed that the extensions of our
terms are fixed by our own dispositions regarding their usage. However, it need not always be the
case that when a speaker uses a word (at some time t), her dispositions at t regarding the use of that
word are among the determinants of its meaning—i.e. are among the determinants of the intended
interpretation(s) of her speech at t. We should not want this to be the case, because it would rule out
the possibility that what one said might be false—i.e. false on (all) the intended interpretation(s). It
is, then, part and parcel of my view that when we speak, and mean something by our words—i.e.
speak relative to an intended interpretation(s)—there must be a ‘reference class’ of speakers whose
dispositions regarding the use of the words one uses serve to determine (at least partly—there may be
other factors involved) which interpretation(s) are intended, i.e. which interpretation(s) one is speaking
315
relative to, i.e. what one means by what one is saying. However it is not part and parcel of the view
that one is oneself a member of this reference class. For example, one might speak relative to an
interpretation picked out by one’s teachers usage dispositions, or by some experts’ usage dispositions, or
by the usage dispositions one had at some earlier time (e.g. when one was less tired).
This page intentionally left blank
Conclusion
We have covered a lot of ground, and I will not recapitulate every step. It
might, however, be useful to conclude by retracing one key line of thought,
and showing how it leads to the positive view that I have presented and
defended: fuzzy plurivaluationism.
The line of thought begins with the observation that there are two
problems with the epistemicist view of vagueness—specifically, with its
positing of a particular change point in a Sorites series for P at which
the claim ‘this object is P’ goes from being true to being false. The first
problem is what I call the jolt problem. The problem is the nature of the
semantic shift posited by the epistemicist, and in particular the dramatic
difference between the semantic statuses of the claims ‘this object is P’
(which is true) and ‘that object is P’ (which is false), where these claims
concern the objects on either side of the change point—objects which
are very similar in all respects relevant to the application of the predicate
P. In short, the epistemicist thinks that as we go down the Sorites series
saying ‘this is P’ of each object in turn, there is at some point a sudden
jolt, as our claims crash all the way from true to false in one step—and
this has been thought to be implausible. The second problem is what I call
the location problem. The problem concerns the fixing of the location of the
change point, and in particular the fact that we cannot see how our own
usage of language (together with facts about our causal connections to our
environment, referential eligibility, simplicity, and so on) could fix it to be
at any particular point in the series. The epistemicist thinks that there is a
number n such that the nth object in the series is the last one to which P
truly applies—but why, it has been objected, not n − 1 or n + 1? What is
it about our usage of P that gives it a meaning which singles out a unique
n in this way?
In the literature, these two sorts of problem have tended to be run
together under the heading ‘higher-order vagueness’. For example, when
someone wants to object to (say) a truth gap view of vagueness, on the
grounds that, while it avoids a point in the Sorites series at which the claim
‘this is P’ crashes from true to false, it does not avoid dramatic semantic shifts
318
Abbott, Barbara (1997). Models, truth and semantics. Linguistics and Philosophy, 20:
117–38.
Akiba, Ken (1999). On super- and subvaluationism: A classicist’s reply to Hyde.
Mind, 108: 727–32.
(2004). Vagueness in the world. Noûs, 38: 407–29.
Barnett, David (2000). Vagueness-related attitudes. Philosophical Issues, 10: 302–20.
Beall, JC, and Colyvan, Mark (2001). Heaps of gluts and Hyde-ing the sorites.
Mind, 110: 401–8.
Beaney, Michael (ed.) (1997). The Frege Reader. Blackwell, Oxford.
Benacerraf, Paul (1965). What numbers could not be. Philosophical Review, 74:
47–73.
Bennett, Brandon (1998). Modal semantics for knowledge bases dealing with
vague concepts. In A. G. Cohn, L. Schubert, and S. Shapiro (eds.), Principles
of Knowledge Representation and Reasoning: Proceedings of the 6th International
Conference (KR-98), 234–44. Morgan Kaufmann, San Mateo, CA.
Black, Max (1997 [1937]). Vagueness: An exercise in logical analysis. In Keefe and
Smith (1997b), 69–81.
Boghossian, Paul (1989). The rule-following considerations. Mind, 98: 507–49.
Bonini, Nicolao, Osherson, Daniel, Viale, Riccardo, and Williamson, Timothy
(1999). On the psychology of vague predicates. Mind and Language, 14: 377–93.
Brady, Ross T. (1988). A content semantics for quantified relevant logics I. Studia
Logica, 47: 111–27.
Braun, David, and Sider, Theodore (2007). Vague, so untrue. Noûs, 41: 133–56.
Bueno, Otávio, and Colyvan, Mark (2006). Just what is vagueness? Typescript
(available at <http://homepage.mac.com/mcolyvan/papers/jwiv.pdf>).
Burgess, John A. (1990). The sorites paradox and higher-order vagueness. Synthese,
85: 417–74.
(1998). In defence of an indeterminist theory of vagueness. Monist, 81:
233–52.
(2001). Vagueness, epistemicism and response-dependence. Australasian Journ-
al of Philosophy, 79: 507–24.
and Humberstone, I. L. (1987). Natural deduction rules for a logic of
vagueness. Erkenntnis, 27: 197–229.
Burns, Linda (1991). Vagueness: An Investigation into Natural Languages and the Sorites
Paradox. Kluwer Academic Publishers, Dordrecht.
322
Hájek, Petr (1999). Ten questions and one problem on fuzzy logic. Annals of Pure
and Applied Logic, 96: 157–65.
Halmos, Paul, and Givant, Steven (1998). Logic as Algebra. The Dolciani Mathem-
atical Expositions, 21. The Mathematical Association of America.
Hansson, Bengt (1968). Fundamental axioms for preference relations. Synthese, 18:
423–42.
Hardin, C. L. (1988). Color for Philosophers: Unweaving the Rainbow. Hackett,
Indianapolis.
Hart, W. D. (1992). Hat-tricks and heaps. Philosophical Studies, 33: 1–24.
Heck, Richard G. (1993). A note on the logic of (higher-order) vagueness. Analysis,
53: 201–8.
Hohwy, Jakob (2001). Semantic primitivism and normativity. Ratio, 14: 1–17.
Horgan, Terence (1994). Robust vagueness and the forced-march sorites paradox.
Philosophical Perspectives, 8: 159–88.
Horwich, Paul (1998). Truth, 2nd rev. edn. Clarendon Press, Oxford.
Hyde, Dominic (1994). Why higher-order vagueness is a pseudo-problem. Mind,
103: 35–41.
(1997). From heaps and gaps to heaps of gluts. Mind, 108: 641–60.
(1999). Pleading classicism. Mind, 108: 733–5.
(2001). A reply to Beall and Colyvan. Mind, 110: 409–11.
(2003). Higher-orders of vagueness reinstated. Mind, 112: 301–5.
Islam, Amitavo (1996). The abstract and the concrete. Presented at the Annu-
al Conference of the Australasian Association of Philosophy, University of
Queensland.
Jaśkowski, Stanisław (1969). Propositional calculus for contradictory deductive
systems. Studia Logica, 24: 143–57. Translation by O. Wojtasiewicz from the
Polish: Rachunek zdań dla systemów dedukcyjnych sprzecznych. Studia Societatis
Scientarum Torunensis, sectio A, vol. I, no. 5, Toruń, 1948.
Johnston, Mark (1989). Dispositional theories of value. Proceedings of the Aristotelian
Society, supp. vol. 63: 139–74.
(1993). Objectivity refigured: Pragmatism without verificationism. In
J. Haldane and C. Wright (eds.), Reality, Representation and Projection, 85–130.
Oxford University Press, New York.
Kamp, Hans (1975). Two theories about adjectives. In Edward L. Keenan (ed.),
Formal Semantics of Natural Language, 123–55. Cambridge University Press,
Cambridge.
(1981). The paradox of the heap. In Uwe Mönnich (ed.), Aspects of
Philosophical Logic, 225–77. D. Reidel, Dordrecht.
325
Lewis, David (1993). Many, but almost one. In John Bacon, Keith Campbell,
and Lloyd Reinhardt (eds.), Ontology, Causality and Mind: Essays in Honour of
D. M. Armstrong, 23–38. Cambridge University Press, Cambridge.
(1999). Papers in Metaphysics and Epistemology. Cambridge University Press,
Cambridge.
Łukasiewicz, Jan (1970). Selected Works, ed. L. Borkowski. North-Holland,
Amsterdam.
and Tarski, Alfred (1970 [1930]). Investigations into the sentential calculus.
In Łukasiewicz (1970), 131–52.
MacFarlane, John (2006). The things we (sorta kinda) believe. Philosophy and
Phenomenological Research, 73: 218–24.
Machina, Kenton F. (1976). Truth, belief, and vagueness. Journal of Philosophical
Logic, 5: 47–78.
Malinowski, Grzegorz (1993). Many-Valued Logics. Clarendon Press, Oxford.
McGee, Vann (1997). ‘Kilimanjaro’. Canadian Journal of Philosophy, supp. vol. 23:
141–63.
and McLaughlin, Brian (1995). Distinctions without a difference. Southern
Journal of Philosophy, 33 (supp.): 203–51.
Mehlberg, Henryk (1997 [1958]). The Reach of Science. University of Toronto
Press, Toronto. Extract from §29, pp. 256–9, repr. under the title ‘Truth and
Vagueness’, in Keefe and Smith (1997b), 85–8.
Menger, Karl (1942). Statistical metrics. Proceedings of the National Academy of
Sciences of the United States of America, 28: 535–7.
Menzel, Christopher (1990). Actualism, ontological commitment, and possible
world semantics. Synthese, 85: 355–89.
Merrill, G. H. (1980). The model-theoretic argument against realism. Philosophy
of Science, 47: 69–81.
Milne, Peter (2007). Bets and fuzzy propositions: Comments on Nicholas J.
J. Smith’s ‘Degrees of truth, degrees of belief, and pragmatics’. Presented at the
Arché Vagueness Conference, St Andrews, 8 June.
Nguyen, Hung T., and Walker, Elbert A. (2000). A First Course in Fuzzy Logic,
2nd edn. Chapman & Hall/CRC, Boca Raton.
Norwich, A. M., and Turksen, I. B. (1982). The fundamental measurement of
fuzziness. In Ronald R. Yager (ed.), Fuzzy Sets and Possibility Theory: Recent
Developments, 49–60. Pergamon, New York.
Novák, Vilém (1998). Fuzzy logic. In Dov M. Gabbay and Philippe Smets (eds.),
Handbook of Defeasible Reasoning and Uncertainty Management Systems, i: Quantified
Representation of Uncertainty and Imprecision, 75–109. Kluwer, Dordrecht.
327
Raffman, Diana (1996). Vagueness and context relativity. Philosophical Studies, 81:
175–92.
(2000). Is perceptual indiscriminability nontransitive? Philosophical Topics, 28:
153–75.
Ramsey, F. P. (1990 [1926]). Truth and probability. In Ramsey (1990), 52–94.
(1990). Philosophical Papers, ed. D. H. Mellor. Cambridge University Press,
Cambridge.
Rea, George (1989). Degrees of truth versus intuitionism. Analysis, 49: 31–2.
Read, Stephen, and Wright, Crispin (1985). Hairier than Putnam thought.
Analysis, 45: 56–8.
Rescher, Nicholas (1969). Many-Valued Logic. McGraw-Hill, New York.
Restall, Greg (1994). On Logics without Contraction (PhD thesis, Department of
Philosophy, University of Queensland).
Robertson, Teresa (2000). On Soames’s solution to the sorites paradox. Analysis,
60: 328–34.
Rolf, Bertil (1984). Sorites. Synthese, 58: 219–50.
Russell, Bertrand (1997 [1923]). Vagueness. In Keefe and Smith (1997b), 61–8.
Sainsbury, Mark (1986). Degrees of belief and degrees of truth. Philosophical Papers,
15: 97–106.
(1997 [1990]). Concepts without boundaries. In Keefe and Smith (1997b),
251–64.
(1991). Is there higher-order vagueness? Philosophical Quarterly, 41: 167–82.
(1995). Paradoxes, 2nd edn. Cambridge University Press, Cambridge.
Sanford, David H. (1975). Borderline logic. American Philosophical Quarterly, 12:
29–39.
(1976). Competing semantics of vagueness: Many values versus super-truth.
Synthese, 33: 195–210.
(1993). The problem of the many, many composition questions, and naive
mereology. Noûs, 27: 219–28.
Schiffer, Stephen (1999). The epistemic theory of vagueness. Philosophical
Perspectives, 13: 481–503.
(2000). Vagueness and partial belief. Philosophical Issues, 10: 220–57.
Schwartz, Stephen P. (1987). Intuitionism and sorites. Analysis, 47: 179–83.
(1990). Intuitionism versus degrees of truth. Analysis, 50: 43–7.
and Throop, William (1991). Intuitionism and vagueness. Erkenntnis, 34:
347–56.
Shapiro, Stewart (2006). Vagueness in Context. Clarendon Press, Oxford.
Skyrms, Brian (1984). Pragmatics and Empiricism. Yale University Press, New Haven.
Smith, Nicholas J. J. (2003). Vagueness by numbers? No worries. Mind, 112:
283–90.
329
(2004). Vagueness and blurry sets. Journal of Philosophical Logic, 33: 165–235.
(2005a). A plea for things that are not quite all there: Or, Is there a problem
about vague composition and vague existence? Journal of Philosophy, 102:
381–421.
(2005b). Vagueness as closeness. Australasian Journal of Philosophy, 83: 157–83.
(2006). Semantic regularity and the liar paradox. The Monist (special issue on
Truth), 89: 178–202.
(2008). Why sense cannot be made of vague identity. Noûs, 42: 1–16.
Soames, Scott (1997). Skepticism about meaning: Indeterminacy, normativity, and
the rule-following paradox. Canadian Journal of Philosophy, supp. vol. 23: 211–49.
(1999). Understanding Truth. Oxford University Press, New York.
(2002). Replies. Philosophy and Phenomenological Research, 65: 429–52.
Sorensen, Roy (1988). Blindspots. Clarendon Press, Oxford.
(2001). Vagueness and Contradiction. Clarendon Press, Oxford.
Stalnaker, Robert C. (1999). Context and Content: Essays on Intentionality in Speech
and Thought. Oxford University Press, Oxford.
Suppes, Patrick (1957). Introduction to Logic. D. Van Nostrand, Princeton.
and Zinnes, Joseph L. (1963). Basic measurement theory. In R. Duncan
Luce, Robert R. Bush, and Eugene Galanter (eds.), Handbook of Mathematical
Psychology, i. 1–76. John Wiley and Sons, New York.
Tappenden, Jamie (1993). The liar and sorites paradoxes: Toward a unified
treatment. Journal of Philosophy, 90: 551–77.
Tye, Michael (1989). Supervaluationism and the law of excluded middle. Analysis,
49: 141–3.
(1990). Vague objects. Mind, 99: 535–57.
(1997 [1994]). Sorites paradoxes and the semantics of vagueness. In Keefe
and Smith (1997b), 281–93.
(1994). Why the vague need not be higher-order vague. Mind, 103: 43–5.
(1995). Vagueness: Welcome to the quicksand. Southern Journal of Philosophy,
33 (supp.): 1–22.
(1996). Fuzzy realism and the problem of the many. Philosophical Studies, 81:
215–25.
Unger, Peter (1979a). There are no ordinary things. Synthese, 41: 117–54.
(1979b). Why there are no people. Midwest Studies in Philosophy, 4: 177–222.
Urquhart, Alasdair (1986). Many-valued logic. In D. Gabbay and F. Guenthner
(eds.), Handbook of Philosophical Logic, iii. 71–116. D. Reidel Publishing
Company, Dordrecht.
Van Cleve, James (1992). Semantic supervenience and referential indeterminacy.
Journal of Philosophy, 89: 344–61.
330
Van Fraassen, Bas C. (1966). Singular terms, truth-value gaps and free logic.
Journal of Philosophy, 63: 481–95.
Varzi, Achille C. (2001). Vagueness in geography. Philosophy & Geography, 4:
49–65.
(2003a). Higher-order vagueness and the vagueness of ‘vague’. Mind, 112:
295–9.
(2003b). Indeterminate identities and semantic indeterminacy. Typescript
(available at <http://www.columbia.edu/∼av72/papers/Parsons 2003.pdf>);
text of a talk presented at The Philosophy of Terence Parsons: Logic,
Metaphysics, and Natural Language, University of Notre Dame, 8 February.
(2007). Supervaluationism and its logics. Mind, 116: 633–75.
Wallace, John (1977). Only in the context of a sentence do words have any
meaning. Midwest Studies in Philosophy, 2: 144–64.
Weatherson, Brain (2003a). Epistemicism, parasites, and vague names. Australasian
Journal of Philosophy, 81: 276–9.
(2003b). Many many problems. Philosophical Quarterly, 53: 481–501.
(2004). Vagueness and pragmatics. Typescript (available at <http://brian.
weatherson.org/vagueprag.pdf>).
(2005). True, truer, truest. Philosophical Studies, 123: 47–70.
(2006). Vagueness as indeterminacy. Typescript (available at <http://brian.
weatherson.org/vai-oct.pdf>).
Weiner, Joan (2004). Frege Explained: From Arithmetic to Analytic Philosophy. Open
Court, Chicago.
Williamson, Timothy (1992). Inexact knowledge. Mind, 101: 217–42.
(1997 [1992]). Vagueness and ignorance. In Keefe and Smith (1997b), 265–80.
(1994). Vagueness. Routledge, London.
(1996a). Putnam on the sorites paradox. Philosophical Papers, 25: 47–56.
(1996b). What makes it a heap? Erkenntnis, 44: 327–39.
(1999). On the structure of higher-order vagueness. Mind, 108: 127–43.
(2000). Knowledge and its Limits. Oxford University Press, Oxford.
(2003). Vagueness in reality. In Michael J. Loux and Dean W. Zimmerman
(eds.), The Oxford Handbook of Metaphysics, 690–715. Oxford University Press,
Oxford.
Wright, Crispin (1973). On the coherence of vague predicates. Synthese, 30:
325–65.
(1987). Further reflections on the sorites paradox. Philosophical Topics, 15:
227–90.
(1992). Is higher-order vagueness coherent? Analysis, 52: 129–39.
331
Greenough, P. 131 n. 2, 132 n., 137–8, admissible 81, 96, 98; see also extension
157 n. 50, 182–3, 204 of an interpretation
Gregory, D. 46 n. 18 correct, see intended interpretation;
Grice, P. 226 interpretation, acceptable
intended, see intended interpretation
Haack, S. 11, 210 n. 2, 276, 277 partial 71–2, 76–7, 81, 89 n. 64, 96–7,
Hájek, A. vi 101–3, 115–16, 192
Hájek, P. 32, 60 n., 67 n., 68, 69 n. 48, intersection 19, 25
149 n., 276 n. 88 involution 24
Halmos, P. 27 n. 6 Islam, A. vi, 32, 46 n. 18, 101 n. 75, 213 n.,
Hansson, B. 302 n. 23 219 n. 13
Hardin, C. 140 n. 23 Ismael, J. vi
Hart, W. 142 n. 26
Heck, R. 174 n. 66 Jaśkowski, S. 94
hedging response 1, 37–8, 41, 58, Jeffrey, R. vi
86 Johnston, M. vi, 136 n. 15
Hesperus 245 jolt problem 58, 59 n. 34, 117, 172–4,
Hohwy, J. 282 n. 7 192–8, 292, 317–19
Horgan, T. 118, 305 n. judgement-dependence 115 n. 85
Horwich, P. 34 n. 2, 274 n. Julius Caesar 1
Humberstone, L. 86 n. 59 jump point 177, 187–8, 190; see also jolt
Hyde, D. 82 n. 54, 94–5, 135–6 n. 11 problem
vagueness: Wallace, J. 47 n. 20
definition of 1–2, 8–9, 127–40, 152–5, WAM (warranted assertibility
180–5, 206, 278–9, 291; see also manoeuvre) 255–6
closeness definition of vagueness Weatherson, B. 41, 137 n., 149 n., 150 n.,
higher-order 57–8, 117 n. 89, 134, 155 n. 46, 158, 257 n. 68, 259, 260,
172–4, 192–3, 197, 290, 304 n., 262, 273 n., 290–1
310–11, 317–18; see also jolt Weiner, J. 32
problem; location problem wf 26, 29, 254
linguistic, see semantic indeterminacy atomic 30, 65, 72, 78
metaphysical 71, 157–8, see also object; closed 30 n.
worldly vagueness interpreted 112 n. 80, 254, 263 n., 272,
partial versus total 155–7 288–9
theory of 3, 15–16, 127, 129–32 Williams, R. 41 n. 13, 148 n.
validity 10, 35, 60, 82, 95, 98, 123, 130, Williamson, T. 15, 16, 34 n. 2, 35 n., 36
168–70, 212, 220–4, 265–6, 270–2, n. 7, 39–44, 82 n. 55, 85, 123 n., 131
288–9, 293; see also consequence n. 2, 134 n. 9, 173, 181, 183 n., 185 n.,
global and local 82 n. 55 194 n., 224 n. 19, 250 n. 56, 251, 252
value: n. 59, 256, 258, 260 n. 72, 261, 266 n.,
designated 220–2 274–5, 279 n., 290 n., 294, 304 n.,
of a function 20 305, 307–8
variable 29 witness 83
free 30 n. problem of missing 83–4, 87, 95–6,
Varzi, A. 100, 136 n. 13, 305 n. 109–10
vector space 145 n. 31 worldly vagueness 4–6, 44–5, 50, 70–1,
VFS (vagueness-free situation) 237–8, 76, 96–8, 121–2, 289, 292
241–4 Wright, C. 7, 36 n. 6, 123, 138 n. 20, 139,
Viale, R. 252 n. 59 159–65, 174 n. 66, 250–1, 267, 304 n.
Vidler, C. vi
vocabulary 26, 29
logical 257 n. 67, 271 Yager, R. 67, 68
VPB (vagueness-related partial Yuan, B. 60 n., 67, 68, 69 n. 47
belief) 230–1
Zadeh, L. 60 n., 275–6
Walker, E. 60 n., 63, 67 n., 68 n. 45 Zimmerman, D. 46 n. 18
Walker, K. vi Zinnes, J. 300 n. 20