Publisher policy on AI in submitted work
Each publisher and journal decides what an author must declare about artificial intelligence in a submitted manuscript, what the tool may have done and where the declaration goes. Those rules change quickly, so every statement of them belongs to its date.
What the publishers agreed on
In 2024 publisher policies on AI agreed that generative AI cannot be listed as an author and that its use has to be disclosed. Organisations with policies or guidance of their own include COPE, the ICMJE, the publication professionals' society, the AMA through its author guidance, and EMWA through its AI working group. Most publisher policies include guidance for editors and peer reviewers as well as for authors. Taylor & Francis expanded its generative AI policy in June 2024, so that besides authorship and disclosure it covers how the tools should and should not be used and how to disclose them. By May 2024 most universities also had guidelines on how to disclose the use of AI in papers.
By September 2023, ChatGPT had been banned as a co-author of scientific papers, and many publishers were asking authors to state clearly whether and how they had used AI tools in their writing. Guidance on AI in manuscripts by then included the JAMA guidance for authors, peer reviewers and editors of July 2023, the Journal of Global Health's policy, the WAME recommendations on chatbots and scholarly manuscripts, COPE's position statement on authorship and AI tools, the ICMJE's section on AI-assisted technology in its definition of authors' and contributors' roles, The Lancet's information for authors, and NEJM's editorial policy on AI-assisted technologies. Springer Nature, Elsevier, Taylor & Francis, the JAMA Network, WAME with the BMJ, and the ICMJE had by then published guidelines intended to regulate the use of text-generation tools for publications.
Under publisher policies, the authors are ultimately responsible for the originality, validity and integrity of the content in their manuscript. JAMA reminds authors who use AI tools that they must take responsibility for any content generated, including its integrity, and for the correctness of the information in the manuscript even where it was taken from an AI tool. Under The Lancet's guidelines on AI and AI-assisted technologies in scientific writing, as they stood in December 2023, the authors are ultimately responsible and accountable for the originality, accuracy and integrity of the work.
Why an AI tool cannot be an author
Around January 2023 a paper published in Nature named ChatGPT as one of its authors. Nature reported two preprints and two published articles in the science and health fields that carried ChatGPT as a bylined author, each with an affiliation for it and one with a non-human address and a conflict-of-interest disclosure for OpenAI. All four were indexed in both PubMed and Google Scholar. Their fates differed: one preprint lost its non-human author when it was published as a journal article; the human author of one article, whom one editor took for a provocateur, argued in the text that the tool should be an author; and the fourth was apparently an error, corrected with the second author removed.
By June 2023 the ICMJE and JAMA had issued guidance, and under JAMA's policy an AI tool cannot be an author. JAMA released the policy in January 2023, when many journals and societies released theirs very quickly, and it still held in 2024. It states that non-human artificial intelligence, machine learning and similar technologies do not qualify for authorship because they cannot be accountable.
Springer Nature's editorial policies do not attribute authorship to AI. In Springer Nature's view in 2024, AI cannot act as an author, because accountability and credibility cannot be given to it, it cannot take responsibility for errors, and authorship requires responsibility.
The Lancet's guidelines of December 2023 said AI and AI-assisted technologies should not be listed as an author or co-author, and AI should not be cited as an author.
Where the publishers disagreed
Publisher policies were broadly consistent in allowing idea generation and exploration. Taylor & Francis thought in 2024 that language improvement was the most common use it was seeing, and probably the one most often disclosed. Other uses are online searches, literature classification, coding assistance, aid in the analysis of data, and the production or editing of written and visual content, which not all publishers and journals allow.
The one point on which the policies differed in 2024 was whether generative AI may be used to generate text and images. Some journals and publishers allow it, others do not, Taylor & Francis among them, and some journals do not say. Taylor & Francis advises establishing before submission whether generated text and images may be used, and asking the target journal from the outset what is and is not allowed.
JAMA's policy, as updated by September 2023, discourages the submission and publication of content created by AI, language models, machine learning or similar technologies, and does not permit it without a clear description of the content created. As JAMA described it in 2024, the policy addresses intellectual property and image integrity by discouraging the submission and publication of content created by AI unless it is part of the formal research design.
By December 2023 The Lancet's guidelines said AI and AI-assisted technologies should only be used to improve the readability and language of the work, and not to replace researchers' tasks such as producing scientific insights and analysing and interpreting data.
The Lancet draws the line between sections of the journal. In 2024 it newly limited generative AI in its "blue" content, meaning editorials, commentaries and some correspondence, to spelling and grammar, and wants it used there in no other way. The journal sees that section as an area of human intelligence and wants it driven by human intelligence. In its experience, commentaries submitted using generative AI do not provide useful insights into significant conceptual advances, or into healthcare and medicine and what needs to happen there.
Spelling and grammar stay allowed so as not to exclude people who do not have English as a first language or who are neuro-atypical, such as dyslexic clinicians, but the tool is not to be used to draw insights on advances in research.
Images and covers
Springer Nature's editorial policies do not allow the inclusion of generative-AI images.
JAMA does not want a clinical image generated by AI presented as an actual one. It expects generated images where they are part of a study. In 2016, for example, JAMA published a study that used AI-generated retinal images in a quality-assurance process to help train physicians to recognise retinal disease, diabetic retinopathy in particular, in the hope that it could be used for training in low- and middle-income countries, and the article carries generated retinal scans.
The Lancet likes generative AI for brainstorming conceptual images and figures but does not want the generated images used, because of copyright and because it is not possible to establish fully where the material came from on the internet. An AI-generated cover would be acceptable for a special issue on generative AI in one of its journals, such as The Lancet Digital Health, and otherwise The Lancet recommends against AI-generated covers.
What to disclose, and where
Publisher policies differ in what they ask authors to disclose; between them they ask for the content created or modified, the name and version of the tool, the accuracy of the output and the reason for its use, and some specifically ask for the prompts in supplementary material. The policies agree on placement: use for language and editing is disclosed in the acknowledgements, and use as part of the research study goes in the methods. In Taylor & Francis's view, best practice is also to disclose the use in the cover letter, so that the journal's editor knows from the outset that AI was used, how, and where it has been declared in the manuscript.
Since its policy of January 2023, JAMA has asked for a description of the content created or edited, the name of the language model or tool, the version and extension numbers, the manufacturer, and the dates of generation, the same kind of reporting it expects for statistical software. The first such acknowledgement JAMA published, in September 2023, gave no date, and by 2024 JAMA asked for one because prompts change depending on how the tools are used. JAMA updated the policy in March 2024 to detail what authors are to report if they used AI to generate content or to assist with reviewing and editing it. JAMA's policies are in the instructions for authors of all its journals and in the AMA Manual of Style.
The Lancet's guidelines of December 2023 had authors disclose the use of AI and AI-assisted technologies in a statement at the end of the work. By 2024 The Lancet asked for the prompt, because no tool it knew of could accurately detect the use of generative AI and the journal relied on authors to be as transparent as possible, and it counted this among its differences from JAMA. For all its pieces The Lancet asks where in the article the tool was used, how, with what prompt and for what. Where it was used for something such as data cleaning or generating search terms, the methodology has to describe that use accurately enough for another author to reproduce it.
Springer Nature declares AI drafting in a statement published with the piece. On a prime view published in November 2024, the additional information says the piece was initially drafted by artificial intelligence and then edited by the editors to meet Nature Portfolio publication standards. The statement names who edited or curated the draft: the editors, or sometimes the authors, when the summary is shared with them to curate or edit. In Springer Nature's view, any AI use, for whatever purpose, has to be declared in the piece itself, whether a plain language summary, a research highlight, a whole review paper or a small portion of a paper.
Taylor & Francis's provisional recommendation for declaring generative AI used to draft plain language summaries places the declaration according to the document. Where AI was used for a manuscript that the authors later extensively edited and amended, the use is stated in the acknowledgements. A plain language summary that sits alongside an article is covered by the article's acknowledgements, and a standalone plain language summary of a publication has its own transparency statement in which the AI use is acknowledged.
In Taylor & Francis's view, an acknowledgement gives enough credit to a tool that inspired the wording, and giving it that credit is ethical. Using the tool resembles consulting a thesaurus for another way to say something, and what matters is whether it helps communicate better.
One review made with AI carried a completed disclosure; the first journal it went to rejected it and the second accepted it, and its author did not put the rejection down to AI. That author advises anyone using a language model to disclose it, since there is no reason not to. The same author expected in March 2024 that much more regulation of the question would come soon. In the accounts of two other practitioners, whatever a publication takes from an AI tool has to be disclosed in the methods section, and in one account the methods have to give the version of the tool, the search string and everything else related to the tool.
Three disclosures in recently published articles show the range. One is in the acknowledgements, naming ChatGPT 3.5 and its use for translation and grammatical correction. Another is in the author contribution statement under its own heading, "Declaration of generative AI and AI assisted technologies in the writing process", naming ChatGPT and DeepL, used to rephrase a handful of sentences and to help with grammar and spelling. The third, which Taylor & Francis holds up as best practice, discloses the use in the methods section of the abstract, in the acknowledgements for grammar, in the methods of the manuscript, and in the electronic supplementary file for the prompts.
Four statements published in one issue of EMWA's journal in September 2023 differ in how much they specify. An acknowledgement states that the author used ChatGPT (OpenAI, model GPT-4, version of 24 May 2023) to rephrase some passages, reviewed the generated text and edited it as needed, and takes full responsibility for the generated content. A disclaimer states the division of labour without naming a tool: a human author wrote the content, and AI was used for grammatical corrections and phraseology suggestions. A concluding remark declares that the article was written without any AI support, and adds that the Editor function of Microsoft 365 Word, an AI-enabled writing assistant, scored the text 86% correct. A fourth piece carries a joke in the form of a disclosure, that no alliteration generator was used in writing the piece. Only the first gives a tool with its version and date.
The required question at submission
JAMA added a required question to its manuscript submission form in August 2023: after a reminder of the policy, the author is asked whether AI or other technology was used to create or assist with the creation or editing of any content in the submission, and a prompted follow-up asks what was done, how, and what needs to be reported. From August 2023 into 2024, 1.6% of 61,000 submissions indicated AI use. The most common uses were language, grammar and translation, meeting word-count restrictions for titles and abstracts, and studies of which AI was itself the focus. JAMA believes the share is growing, though not fast, and hopes authors are being honest.
The Lancet's equivalent is a yes/no checkbox implemented earlier in 2024, and its initial results were slightly higher: about 6% of authors said they had used generative AI in some form, with the proportion slightly higher in clinical content, meaning reviews, literature searches and seminars, and lower in original research.
Chatbots as formal references
In 2024 JAMA recommended naming a chatbot in the manuscript as statistical software is named, or in the acknowledgements, but not citing it as a formal reference, because the tools do not yet have that level of trust, and the policy could evolve. The recommendation is in line with the second of two posts by librarians in June 2024 on whether to cite chatbots, which made the case against citing them as information sources: a piece of text entirely generated by generative AI ought to be marked as such for the sake of trust and transparency, but citing a chatbot as an information source creates a precedent that undermines the responsibility of authors. A chatbot asked the same question in July 2024 answered that there was no universal consensus, and that many style guides and academic institutions recommend acknowledging the use of AI tools but not necessarily citing them as a traditional source.
The journal's policy and the publisher's
Taylor & Francis advises a writer who is unsure to ask the journal, preferably before starting, rather than after making something that then cannot be used. Many publisher policies note that individual journals may have more specific guidance, so a publisher's policy does not necessarily cover all that applies to every journal it publishes. Taylor & Francis holds that the final decision on what constitutes acceptable use of generative AI tools rests with the journal editor and the publisher. It advocates fixing the target journal early in publication planning and, where there is more than one, talking to each of those editors.
The boundaries the policies have not drawn
Authors ask whether the way they used a tool in their manuscript counts as generative AI. Sage's policy draws the line explicitly, treating the editorial suggestions a word processor makes about spelling, grammar and better ways to write a sentence as AI assistance that does not necessarily need to be disclosed, and anything from generative AI as something that does. AI assistance comes from natural language processing, of which large language models are a subset.
Authors are asked to declare generative AI tools, but Taylor & Francis has no guidance on declaring other machine-learning tools such as Grammarly or PerfectIt, which some authors acknowledge anyway. To Taylor & Francis, the question suggests stepping back to ask what should be disclosed, since authors already state in the methods which spreadsheet program or platform they used. The more transparent the report, the easier the research is to reproduce and the better its integrity and quality.
If the tools are self-learning, the same prompt may produce different results later, which raises the question whether a disclosure should also date the use of a prompt. An editor at Taylor & Francis supports dating, because a date lets a later reader place the use against where research stood at that point and how the tools may have learned since, and without a reference point there is nothing to place it against.
Where the policies converged and what stayed unsettled
By late 2024 different publishers made similar requests of authors and of reviewers, especially about generative AI. AI used to polish English and grammar was mostly not considered a problem, and the possible issues lay with generative AI used to delegate thinking almost completely to a machine.
In the year to late 2024 journals including JAMA, The Lancet, Science and Nature upgraded their guidelines and policies as custodians of the scientific literature. In 2023 the sense was perhaps, as with any new technology, to address first the threats to the status quo. By late 2024 the same publishers were also examining the practical and responsible use of these tools in their own processes.
The Lancet made its changes very rapidly over a short period, which is unusual for journals, and caveats all of them as possibly short-lived. Some people expect that before long the question will be "did you not use generative AI?", because everybody will be using it. JAMA, which usually takes two or three years to update a policy, updated its policy of January 2023 in July 2023 and again in March 2024.