In late January 2023, two leading science journals made up their minds about ChatGPT within days of each other. Nature set out ground rules for its own pages and for every Springer Nature journal on 24 January [1]. Two days later, Science published an editorial by its editor-in-chief, Holden Thorp, with a blunt title: “ChatGPT is fun, but not an author” [2].
There was a reason for the hurry. Nature noted that several preprints and published articles had already listed ChatGPT as a formal author [1].
It is tempting to treat this as a question about the byline, as if it were about credit. I think the more useful question is about the work that starts once a paper is out. Somebody has to check that the results are right, that the references exist and that confidential data were handled properly, and somebody has to answer for mistakes. A chatbot can help write the text. It cannot do any of those jobs in the way an author is expected to, and that is the real reason it cannot sign.
Same verdict, different rules
The two journals agreed on authorship and disagreed on almost everything else. Nature said researchers may use large language models (LLMs), the kind of AI behind ChatGPT, as long as they document that use in the methods or acknowledgements section [1]. It described the big worry in the research community: that people would pass off chatbot text as their own, or lean on the tools for shortcuts such as an incomplete literature review and “produce work that is unreliable” [1].
Science went further. Thorp announced that text generated by ChatGPT or any other AI tool could not be used in the work, and neither could figures or images made by such tools [2]. Breaking that rule would count as scientific misconduct, in the same category as altered images or plagiarism [2].
So one journal said use it but tell us, and the other said keep it out of the writing. Both still agreed about the byline. That agreement is what interests me, because it rests on something older than ChatGPT.
What signing a paper actually means
The International Committee of Medical Journal Editors (ICMJE) has four criteria for authorship, and every author is supposed to meet all of them [3]. The first three are about doing the work: contributing substantially to its conception, design or data, drafting or critically revising it, and approving the final version [3]. The fourth is the one I think gets the least attention. An author must agree to be accountable for all aspects of the work, so that questions about the accuracy or integrity of any part of it are “appropriately investigated and resolved” [3].
That fourth criterion is a promise about the future. The ICMJE also expects the corresponding author, the person who deals with the journal, to be available after publication to respond to critiques and to cooperate with requests for data if questions come up [3]. A paper is a claim that someone will keep standing behind.
Imagine a graduate student who finds, two years after a paper came out, that one number in its main table cannot be right. She writes to the journal. Someone has to open the original files, find the error, work out whether it changes the conclusion and sign the correction. A chatbot that drafted the results section cannot do any of that, and it will not be there when the email arrives.
In February 2023, the Committee on Publication Ethics (COPE) made the same point in plainer terms. AI tools cannot take responsibility for submitted work, and because they are not legal entities, they cannot declare conflicts of interest or manage copyright and license agreements [4]. Its conclusion was that authors are fully responsible for their manuscript, even the parts an AI tool produced [4].
Four jobs that stay with a person
When I think about what accountability means in practice, I end up with four ordinary jobs.
The first is checking the results. A language model produces fluent text, and fluent text can be wrong. Thorp quoted ChatGPT’s own website, which warned that it “sometimes writes plausible-sounding but incorrect or nonsensical answers” [2]. Someone has to compare every sentence of a results section with the actual analysis.
The second is checking the references. Thorp noted examples of ChatGPT citing a scientific study that does not exist [2], and Nature pointed out that these tools could not yet cite sources to back up what they wrote [1]. A reference list is only as good as the person who opened each paper.
The third is protecting confidential material. This one is my own view, not something the editorials spell out. I think unpublished data, patient information and a colleague’s draft should not go into any tool until you have checked how it handles them. That judgment belongs to the researcher, and so, I would argue, does the blame if it goes wrong.
The fourth is correcting mistakes. Errors are normal in science. What matters is that someone finds them, admits them and fixes them in public. That is close to the duty the ICMJE attaches to authorship, making sure questions about accuracy are investigated and resolved [3], and it needs a person who can be found.
Thorp made an observation that fits all four. Most of the misconduct cases the Science journals deal with, he wrote, “occur because of an inadequate amount of human attention” [2]. In my view, AI does not create that problem, but it makes it easier to assemble a finished-looking paper with less attention than it needs.
The case for letting AI do more
The fair objection is that a flat ban is too blunt, and I agree with part of it. Science‘s January 2023 rule barred AI-generated text outright, while Nature allowed use with disclosure [1, 2]. For a researcher who writes in a second language, or who uses a tool to tighten a clumsy paragraph, the difference matters. Even Science carved out legitimate data sets generated by AI as part of the research itself [2]. And near the end of his editorial, Thorp says that machines play an important role, as tools for the people who pose hypotheses, design experiments and make sense of the results [2].
I build large language model tools that pull structured information out of medical records, so I am not arguing that these systems are useless. They can save a great deal of time. My point is narrower. Whether a journal allows AI help with the text or not, the list of people who answer for the paper should not change. Disclosure tells readers what tool was used. It does not tell them who checked the output, and that is the part that matters.
Use AI to help write, if your journal allows it and you say so. But before you submit, make sure a named person has checked every result, opened every reference and would answer the email if something turns out to be wrong.
For anyone writing a paper with AI help, the practical test is simple. Go through the ICMJE’s fourth criterion line by line and ask who, by name, is accountable for each part [3]. If the honest answer for any section is that the chatbot wrote it and nobody really checked, the paper is not ready. As Nature put it, research needs “integrity and truth from authors” [1], and only a person can offer that.
References
[1] “Tools such as ChatGPT threaten transparent science; here are our ground rules for their use,” Nature, vol. 613, no. 7945, p. 612, Jan. 2023, doi: 10.1038/d41586-023-00191-1.
[2] H. H. Thorp, “ChatGPT is fun, but not an author,” Science, vol. 379, no. 6630, p. 313, Jan. 2023, doi: 10.1126/science.adg7879.
[3] International Committee of Medical Journal Editors, “Defining the role of authors and contributors,” ICMJE Recommendations. Accessed as archived on Jan. 23, 2023. [Online]. Available: https://www.icmje.org/recommendations/browse/roles-and-responsibilities/defining-the-role-of-authors-and-contributors.html
[4] Committee on Publication Ethics, “Authorship and AI tools: COPE position statement,” Feb. 13, 2023. [Online]. Available: https://publicationethics.org/cope-position-statements/ai-author
Related Articles
