When large language models became capable of producing fluent scholarly prose, an obvious question followed: should the tool be listed as an author? The answer, reached quickly and with rare unanimity across publishers and integrity bodies, is no. But the reason for that answer matters more than the answer itself, because the reasoning tells authors exactly what they should do instead. This article sets out both, drawing on the position at AI authorship and the practical guidance at AI disclosure for authors.
The consensus, and the bodies behind it
The major standard-setters and publishers have converged on a clear rule: a generative AI system cannot be listed as an author. The ICMJE recommendations state it directly; the Committee on Publication Ethics (COPE) takes the same position; and the author instructions of Nature, Science, the major university presses, and the large commercial publishers all say the same thing. This is not a contested or emerging view. It is settled, and it is worth understanding why the agreement was so swift.
The reasoning: authorship is accountability
The argument is short and it rests entirely on one of the authorship criteria. To be an author is, among other things, to be accountable for the work — to take responsibility for its integrity, to be able to answer for it, and to stand behind what it asserts. The ICMJE criteria make this explicit: an author agrees to be accountable for all aspects of the work, ensuring that questions about its accuracy and integrity are investigated and resolved.
A generative AI system cannot do any of this. It cannot take responsibility, cannot be answerable, cannot approve a final version in any meaningful sense, and cannot be held to account if the work proves to be wrong or fabricated. It has no standing to agree to anything. Authorship is therefore categorically unavailable to it — not because of a rule that might be relaxed later, but because the tool lacks the one property authorship is built on. This is the same logic that underpins all of authorship and accountability: a name on the author line is a claim of responsibility, and a tool cannot make that claim.
The test is not “did it contribute to the text?” — plainly a model can. The test is “can it answer for the work?” A tool cannot. That single question settles the authorship question completely.
The corollary: humans remain fully responsible
The accountability argument has a sharp consequence that authors sometimes miss. Because the AI cannot be accountable, the human authors are fully accountable for everything the manuscript asserts, including anything the AI produced. A fabricated citation, an invented statistic, a plausible but wrong sentence, a subtly distorted summary — these are the authors’ errors regardless of which tool generated them. Using a generative tool does not divide responsibility; it concentrates it on the humans who chose to use the tool and chose to publish its output. Disclosure makes the workflow transparent, but it transfers none of the responsibility.
What to disclose instead
If the AI cannot be an author, where does it go in the published record? It is disclosed as a tool or a method, never as a contributor on the author line. Practice has converged on a clear shape for that disclosure, and a good AI use statement answers three questions:
- Which tool. Name the specific generative system used, not a generic category. “A large language model” is not a disclosure; the named tool and, where relevant, its provider, is.
- Where in the workflow. State the stage at which it was used — drafting or editing text, assisting analysis, generating code, producing images — so a reader can locate the boundary of its involvement. Image generation is often held to stricter rules because of the integrity risks around figures.
- How much, and how verified. Indicate the extent of use and, crucially, state that the authors checked the output. A disclosure that the authors verified the tool’s contributions is what connects the transparency back to the accountability that justified excluding the tool from authorship in the first place.
On placement, convention is settling: analytical uses belong in the methods section, where they bear on reproducibility; writing assistance belongs in a dedicated statement near the acknowledgements. A generic line that an AI tool was “used to improve readability” fails the test — it names neither the tool nor the boundary of its use.
What is exempt
Not every interaction with a computational tool is a disclosable use of generative AI. Most policies exempt tools that do not produce novel content: a spell-checker, a grammar corrector, a reference manager, or basic translation of the author’s own words. The line is whether the tool produced novel content that materially shaped the published work. Author-written text whose grammar was tidied is not AI-assisted writing in the disclosable sense. The boundary is not always crisp — substantive rewriting shades into drafting — and where there is genuine doubt, the safe and professional practice is to disclose.
Not authorship, and not ghost-writing either
There is a subtler trap to avoid. Just as a human ghost-writer must be disclosed rather than hidden, an AI tool that substantially drafted a manuscript must be disclosed rather than quietly passed off as the authors’ unaided work. Undisclosed AI drafting is structurally the same failure as undisclosed human ghost-writing: it conceals how the text was produced. The fix is the same — name the tool, state its role, take responsibility for the result.
Where shared vocabulary fits
Publishers’ AI policies agree on the principles but differ in wording, threshold, and placement, which means a disclosure written for one venue does not always mean the same thing when read by another system. What is missing is not more policy but a shared definitional layer: agreed terms for AI-assisted writing, AI-assisted analysis, the exempt category, and the rest, so a disclosure carries the same meaning wherever it is read. Supplying that layer — federating to ICMJE and COPE for the normative content rather than inventing it — is the convening role the CASRAI dictionary is built for; the relevant terms sit in the generative AI disclosure domain.
What to do now
For authors: never list an AI tool as an author; disclose its use as a tool, naming which tool, where, how much, and that you verified the output; and remember you are accountable for everything it produced. For editors: specify where the AI use statement belongs and ask for the specifics, not a vague line. For standards work: pin down shared definitions of the disclosable categories and the exempt threshold so disclosures mean the same thing across venues.
