Brain, Behavior, and Immunity 110 (2023) 152–154

Contents lists available at ScienceDirect

Brain Behavior and Immunity

journal homepage: www.elsevier.com/locate/ybrbi

Viewpoint

A Chat(GPT) about the future of scientific publishing

Elisa L. Hill-Yardin a, b, 1, Mark R. Hutchinson c, d, 1, Robin Laycock a, 1, Sarah J. Spencer a, *, 1
a
School of Health and Biomedical Sciences, RMIT University, Bundoora, Melbourne, Victoria 3083, Australia
b
Department of Anatomy & Physiology, The University of Melbourne, Parkville, Victoria 3010, Australia
c
Adelaide Medical School, Davies Livestock Research Centre, Institute for Photonics & Advanced Sensing, Robinson Research Institute, University of Adelaide, South
Australia 5005, Australia
d
Australian Research Council Centre of Excellence for Nanoscale BioPhotonics, Australia

Recently, AI in science has taken another leap forward in the release discussion of our mistakes or unsupported hypotheses, which will have
of the large language model software, ChatGPT, a publicly accessible other uses for the field. It is foreseeable that the AI will catch up in this
tool developed by OpenAI (https://chat.openai.com). In an article regard, though. In particular, when it can scrape individual databases,
published in February of this year (Manohar and Prasad, 2023), the such as a researcher’s publication record or email correspondence, it
authors played with the idea of having the software write part of their may be able to start injecting individuality and producing nuanced
paper for them, and explored the implications of this. Others have fol­ diction reflective of a specific person’s style. An interesting side note
lowed suit, with at least one journal article listing the software as an here is the question of how AI learns and whether AI may in fact be
author and then correcting that listing after a good deal of discussion in training us to write differently. Predictive text software already uses AI
both the general media and journal offices (O’Connor, 2023a,b). With to suggest a word or phrase that is easier, less vulgar (i.e., not an
authors already adapting artwork from providers like BioRender (Tor­ expletive) or quicker to type than the precise message we intended. Do
onto, Canada), with due acknowledgement, and relying on search en­ we accept this correction, or do we persist with the more laborious
gines like Google for decades now, is borrowing generic text from AI- choice? Is the suggestion more apt, or merely quicker? And if the latter,
generated software really any different? And what does this new AI is AI actually training us to use a less extensive vocabulary that will
mean for BBI and the field of psychoneuroimmunology? affect linguistic diversity on a population basis in the future? Using
Apart from disturbingly fabricated references, the text generated by predictive text in a casual email is one thing but constraining us all into
the software for Manohar and Prasad’s publication (Manohar and Pra­ restrictive and predictable language patterns in scientific publishing is
sad, 2023) is actually pretty good. It’s a fairly accurate, logical, gram­ another.
matically correct introduction to the interaction between lupus and HIV. As well as lacking in a writing “voice”, the text generated by ChatGPT
It is written in simple language, so has potential to sit well with the is too shallow for our purposes as neuroscientists. The human scientist
general public or non-specialists. It is uncontroversial, presenting what needs to have detailed knowledge of the latest relevant findings and
appears to be a mainstream consensus. And, although not shown in the trends, including their strengths and weaknesses and to be able to sift
paper in question, the software is quick, taking less than two minutes to through a lot of noise to make conclusions. All studies are not created
generate 500 words. equal, and it’s essential to be able to interpret the relative importance
Despite its pros, the generated text is largely shallow, bland, dry, and and accuracy of individual works. This leads to perhaps the fundamental
generic, lacking a distinct “voice” - rather robotic really. Unfortunately, reason why ChatGPT should not be used uncritically in academic pub­
this style may be somewhat indistinguishable from (and perhaps accu­ lishing: the tool cannot be accountable for decisions or opinions put
rately mimicking) most writing in scientific articles, which is often forward. To add to this, the nature of the technology is such that there is
bland, dry, formulaic, and lacking in superlatives. Perhaps a way to no transparency. For example, if you ask for a description of how a
“beat” the software, or at least allow us to tell the difference, is to particular neural mechanism operates, we don’t actually know which
introduce more diversity in our writing styles to allow our human-ness sources were used to generate the text, or on what basis decisions about
to be more evident. This approach could also extend to more which elements to include or not include were made. Sometimes the

* Corresponding author.
E-mail address: [email protected] (S.J. Spencer).
1
Equal contribution – authors are listed in alphabetical order.

https://doi.org/10.1016/j.bbi.2023.02.022
Received 20 February 2023; Accepted 25 February 2023
Available online 1 March 2023
0889-1591/© 2023 Elsevier Inc. All rights reserved.
E.L. Hill-Yardin et al. Brain Behavior and Immunity 110 (2023) 152–154

Box 1. ChatGPT was given the command: “Write 250 words on ChatGPT’s usefulness to neuroscience, in the style of Charles Dickens”. Note, ChatGPT’s output was
350 words, here we have cut it to approximately 250 (cuts indicated by [ ]).

statement will not only be out of date or not ‘best practise’ but will activities. It is not all that long ago that a single ANOVA with post hoc test
simply be wrong. These issues will be magnified in specialised scientific was the work of an entire Ph.D project to code and execute. Today we do
disciplines such as neuroscience that evolve very quickly, reducing the many of these statistical tasks within seconds. What ANOVA-like tasks
likelihood that emerging knowledge will be captured. Crucially, science will we be freed from and what biological problems might we solve by
is about new information and linking complex ideas together in new working together with this new technology? In any case, we have
ways to create new innovation. This software is not capable of such high- generated a test case arguing the usefulness of ChatGPT in neuroscience
level critical thinking yet. In its current form it might be useful in research and invite you to judge for yourselves as to whether we’re all
identifying convergent concepts from diverse data and literature sources out of a job just yet (Box 1).
and presenting this information to experienced users who could then
develop testable hypotheses. The regularity with which this AI itself is
trained, however, may even hold back innovation - the current forms of Declaration of Competing Interest
the technology that are publicly available were built using resources that
are 5 years out of date and its training ends at 2021 meaning data The authors declare that they have no known competing financial
published after this time are not currently incorporated (https://www. interests or personal relationships that could have appeared to influence
demandsage.com). the work reported in this paper.
So far it seems that software like ChatGPT is a (figuratively) smart
tool that will support the dissemination of relatively simple, accessible Data availability
information. It is a useful tool to get us started, but, like running an
immunohistochemistry experiment using an antibody of questionable No data was used for the research described in the article.
specificity, without probing, integrating, and non-linear minds to assess
the content, the outputs might be of little value. Science has always been Acknowledgements
a human endeavour; we should embrace new technologies rather than
running scared, but not by removing the human element of science. How The authors would like to acknowledge ChatGPT for providing the
we deal with the next conceptual step of a critically thinking, analytical “in the style of Charles Dickens” paragraph and for hours of entertain­
AI contributing to science in the future is worth planning for now. We ment in trying to get it to mimic Yoda. Any resemblances to the actual
might think we will all be out of a job when this technology gets here, yet works of Mr Dickens are due entirely to the workings of the algorithm
if we harness this potential, it may free us to perform even higher value and it was not our intention to replicate any part of Mr Dickens’ pub­
lications. No artificial intelligences were harmed in the making of this

153
E.L. Hill-Yardin et al. Brain Behavior and Immunity 110 (2023) 152–154

work. Manohar, N., Prasad, S.S., 2023. Use of ChatGPT in academic publishing: a rare case of
seronegative systemic lupus erythematosus in a patient with HIV infection. Cureus
15 (2), e34616.
References O’Connor, S., Corrigendum to “Open artificial intelligence platforms in nursing
education: Tools for academic progress or abuse?” [Nurse Educ. Pract. 66 (2023)
https://chat.openai.com. 103537]. Nurse Educ. Pract., 2023: p. 103572.
https://www.demandsage.com/. O’Connor, S., ChatGpt, Open artificial intelligence platforms in nursing education: Tools
for academic progress or abuse? Nurse Educ. Pract., 2023. 66: p. 103537.

154