AI-Use Disclosure in Manuscripts: ICMJE Rules

The one point every major body agrees on

Strip away the detail and the AI-authorship question has a settled answer. ICMJE's position is that AI tools "cannot be responsible for the accuracy, integrity, and originality of the work, and these responsibilities are required for authorship," so no AI tool, however it was used, can be listed as an author or co-author.¹ COPE's position statement says the same thing in almost the same words: AI tools "cannot take responsibility for the submitted work," so they cannot meet authorship criteria, and a human author remains fully responsible for any content an AI tool helped produce.² Neither body treats this as a close call. Authorship in medical publishing is built around accountability, someone who can stand behind the paper if a reader or an editor comes back with a question, and a language model cannot do that, no matter how the sentence it produced reads.

What must actually be disclosed

The harder question is not whether AI can be an author but what counts as AI use serious enough to disclose. ICMJE requires authors to state, at submission, whether they used AI-assisted technologies such as large language models, chatbots, or image generators "in the production of submitted work," and to describe how in the manuscript itself, writing assistance typically goes in the acknowledgement section, while AI involved in data collection, analysis, or figure generation belongs in the methods.¹ JAMA Network's author instructions, drafted by its editors in 2023 and still the operative policy, put a finer point on it: authors are asked directly whether they used AI, a language model, machine learning, or similar technology "to create or assist with creation or editing of any of the content in this submission," and if so, to name the specific tool, its version, and its manufacturer.⁴ Basic grammar and spell-check tools are explicitly carved out as not requiring disclosure, the line JAMA draws is between correcting a sentence and generating or substantially shaping one.⁴

Where the disagreement actually sits

That carve-out is exactly where journals differ from each other and from their own authors' intuitions. A tool that fixes grammar is uncontroversial. A tool that drafts a paragraph of the introduction from a set of bullet points is, under JAMA's own wording, squarely inside the disclosure requirement, but it is also the use case closest in feel to "editing," which is exempt. An author who ran a chatbot to tighten phrasing may reasonably wonder which category that falls into, and different journals answer with different specificity: some ask a single yes/no disclosure question, others require the exact prompt or tool version, and instructions vary in whether AI used only for literature searching or translation needs the same treatment as AI used to draft results text.³ None of the major bodies dispute the underlying principle, disclose substantive use, skip trivial proofreading, but "substantive" is doing a lot of work, and it is being interpreted differently at the point of submission than editors likely intend.

The gap between the policy and the paper trail

Whatever the exact boundary, the evidence so far suggests it is not being tested very often, not because AI use is rare, but because disclosure is. In a cross-sectional analysis of 25,114 original research submissions to 49 BMJ Group journals between April and November 2024, only 1,431, 5.7%, disclosed any AI use at all, a figure the authors describe as well below survey-based estimates that place actual researcher AI use somewhere between roughly a quarter and three-quarters of authors, depending on the survey.³ Where AI use was disclosed, it was overwhelmingly for improving writing quality (87% of disclosed uses), with data analysis, code writing, translation, and literature searching each accounting for a small minority.³ JAMA Network itself has published a companion analysis of author-reported AI disclosures across its 13 journals since that group's own 2023 disclosure requirement took effect, examining the same question from inside a different publisher.⁵ Neither study claims to know why the gap in the BMJ data exists, under-disclosure, cautious survey respondents overstating their own AI use, or genuine ambiguity about what counts are all plausible contributors, but the size of the gap is large enough that "the policy is written" and "the policy is followed" are clearly not the same claim yet.

Where this leaves a researcher using an AI drafting tool

None of this is a reason to treat disclosure as optional or to guess at what a specific journal wants; author instructions are the authoritative source for any individual submission, and they are revised often enough that last year's answer is not a safe assumption. What the ICMJE and COPE language does make clear is the shape of the question editors are actually asking: not "did a machine touch this paper," but "is there a specific, described role for AI in producing the content, and can a human on the author list stand behind every part of it." That is worth keeping in mind for any tool used during manuscript preparation, including one used only to help write up results. A tool whose only job is turning already-verified numbers into AMA-format prose sits in a different place on that question than one asked to generate or interpret the numbers themselves, but which disclosure category it falls into, and whether it needs disclosing at all under a given journal's instructions, is a question for that journal's current policy, not a default either way.

Frequently asked questions

Can an AI tool be listed as an author on a manuscript?

No. Both ICMJE and COPE agree that AI tools cannot take responsibility for a submitted work, which is the core requirement of authorship, so no AI tool can be listed as an author or co-author regardless of how it was used.

Does using a grammar checker need to be disclosed?

No. Basic grammar and spell-check tools are explicitly exempt from disclosure under JAMA's guidance. The requirement applies to AI used to generate or substantially shape content, not to correct it.

How much AI use in research actually gets disclosed?

Not much relative to how much AI is used. A 2026 audit of 25,114 submissions across 49 BMJ Group journals found only 5.7% disclosed any AI use, well below survey estimates of how many researchers actually use these tools.

References

1. International Committee of Medical Journal Editors. Defining the Role of Authors and Contributors. In: Recommendations for the Conduct, Reporting, Editing, and Publication of Scholarly Work in Medical Journals. Updated January 2024. icmje.org
2. Committee on Publication Ethics (COPE). Authorship and AI Tools. Position statement, version 1, published 20 August 2024. publicationethics.org
3. AlFayyad I, Zeegers MP, Bouter L, Macdonald H, Schroter S. Self-Disclosed Use of AI in Research Submissions to BMJ Journals. JAMA. 2026;335(8):719. https://doi.org/10.1001/jama.2025.25688
4. Flanagin A, Kendall-Taylor J, Bibbins-Domingo K. Guidance for Authors, Peer Reviewers, and Editors on Use of AI, Language Models, and Chatbots. JAMA. 2023;330(8):702-703. https://doi.org/10.1001/jama.2023.12500
5. Perlis RH, Flanagin A, Kendall-Taylor J, Berkwits M, Bibbins-Domingo K. Author Disclosure of Use of AI in Submissions to 13 JAMA Network Journals. JAMA. 2026;335(8):717-719. https://doi.org/10.1001/jama.2025.25300