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

For a few weeks in early 2023 this was an open question. Several preprints and at least one published paper listed a chatbot as an author, and the ensuing debate produced an unusually fast consensus. By that spring the position was fixed and it has not moved since.

The reasoning is worth understanding rather than just the conclusion, because it tells you how the rest of the rules are shaped. Authorship in scholarly publishing is not primarily a credit mechanism, it is an accountability mechanism. An author is someone who can be asked to justify the work, who can be held responsible if it is wrong, who can declare competing interests, and who can sign a licence. A tool can do none of these, so the label does not apply regardless of how much it contributed.

That reasoning also explains why disclosure requirements exist. If a tool cannot be an author, but it did affect the work, the reader still needs to know, and the disclosure statement is where that information goes.

One consequence people miss: you also should not cite AI output as though the tool were the author of a source. If a model tells you something, the thing to cite is the source it came from, once you have found and read it.

What do COPE and ICMJE say about it?

COPE stated in February 2023 that AI tools cannot be authors because they cannot take responsibility for the work or manage conflicts and copyright. ICMJE says the same and adds that authors must not cite AI as an author of any source. Its recommendations were updated in January 2026.

ICMJE also asks journals to enquire at submission whether AI-assisted technologies were used, and asks authors to describe the use in both the cover letter and the manuscript. That is a stronger position than a disclosure-on-request rule and it is spreading beyond medical journals.

Both bodies place the responsibility for verifying AI output squarely on the human authors, including accuracy, originality, attribution, and the absence of plagiarism.

A comparative review of editorial policies traces the ICMJE position to three specific incapacities: a generative AI tool cannot make creative decisions, cannot assume accountability, and cannot disclose conflicts of interest (Yoo, 2025, p. 3). The same review reports that Nature permits authors to use AI tools but prohibits naming one as a co-author, and asks that the tool's role and scope be described, particularly in the Methods section (Yoo, 2025, p. 4). Journals differ in their detailed rules, so read the one you are submitting to, but they converge on human accountability and transparent disclosure.

Where the guidance is still thin is reporting, not authorship. Mehta and colleagues confirm that ICMJE, COPE, the World Association of Medical Editors, and the European Commission all bar AI from authorship and require disclosure, then find no cross-journal consensus on which uses are acceptable and no standardized way to report AI use granularly by stage of manuscript preparation (Mehta et al., 2026, p. 1). Their proposed checklist splits AI use into conceptual contributions, linguistic assistance, and research assistance, and marks which uses in each category raise research-integrity concerns. Until journals adopt something like it, you are writing your disclosure without a template.

How do I credit AI assistance instead?

In the acknowledgments or methods, depending on what the tool did and what the journal specifies. Name the tool and version, state what it was used for and which parts of the work it affected, and confirm that the human authors reviewed and accept responsibility for the output.

Content generation and language assistance usually go in acknowledgments. Anything touching data collection, analysis, code, or figure generation belongs in methods, because it is part of how the result was produced and a reader needs it to evaluate the work.

Keep the statement factual and specific. "Generative AI was used in the preparation of this manuscript" tells a reader nothing. "A language model was used to draft the initial version of the discussion section, which the authors then rewrote and verified against the cited sources" tells them exactly what to weigh.

Treat this as one case of a wider habit. A systematic review of 123 articles reporting 118 studies on authorship found that conception or research design and manuscript writing were the contributions most disciplines counted as qualifying (Marušić et al., 2011, p. 1). A meta-analysis of 14 surveys inside it found that 29% of researchers reported authorship misuse by themselves or someone else, and the estimate rose to 55% outside US, UK, and international-journal settings, against 23% within them (Marušić et al., 2011, pp. 13-14). The same review warns that written criteria alone do not fix this, because researchers were often unfamiliar with the ICMJE criteria or considered them unrealistic or unfair (Marušić et al., 2011, p. 14). The AI rule is settled, the underlying practice of assigning credit honestly is not.