Researchers and authorship: the settled rule and the unsettled rest

Most AI questions in research are unsettled. Authorship is the exception: the rule is written down, it is precise, and the reasoning behind it does the real work. Read that reasoning once and most edge cases answer themselves.

Why an AI cannot be an author

ICMJE authorship requires meeting all four criteria, and the fourth is the one that matters here:

"Agreement to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved."

Accountability is not something a model can agree to. ICMJE draws the conclusion explicitly: chatbots "should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work, and these responsibilities are required for authorship".

COPE — the Committee on Publication Ethics, whose guidance journals and publishers across disciplines subscribe to — reaches the identical conclusion from the identical premise: "AI tools cannot meet the requirements for authorship as they cannot take responsibility for the submitted work", and "Authors are fully responsible for the content of their manuscript, even those parts produced by an AI tool, and are thus liable for any breach of publication ethics".

Note the shape of the argument. It is not that the contribution is too small or too mechanical. It is that authorship is accountability, and there is nobody behind the model to hold accountable. That is why the rule will survive better models — and why ICMJE and COPE, writing independently for different communities, land on the same sentence.

What to disclose, and where

ICMJE splits disclosure by what the AI did. Follow the split exactly, because journals check against it:

• Writing assistance — report it in the acknowledgements.
• Data collection, analysis, or figure generation — describe it in the methods.
• In general — disclose use in both the cover letter and the submitted work at submission.

The logic is the same one the criteria use: acknowledgements credit contributions that fall short of authorship; methods describe how results were produced so they can be examined. AI use lands in one or the other depending on which kind of contribution it was.

What stays yours

Disclosure does not transfer responsibility. ICMJE is direct: "humans are responsible for any submitted material that included the use of AI-assisted technologies".

Three specific duties follow, all stated in the recommendations:

Review and edit the output, because AI "can generate authoritative-sounding output that can be incorrect, incomplete, or biased". That is not caution boilerplate. It is the documented failure mode, and the training incentives behind it are described in the same literature this site covers.

Verify every citation. Authors must ensure "appropriate attribution of all quoted material, including full citations". A fabricated reference in a submitted manuscript is the author's fabrication, whatever generated it. The mechanical check — open each source, find the sentence — is in verifying citations.

Check for plagiarism, including in AI output. The recommendations extend this to "text and images produced by the AI". A model can reproduce phrasing from its training data without marking it; the author who submits it owns that.

The nearest neighbour

The accountability logic here is the same one lawyers and regulated professions work under, where an unopened citation is a sanction rather than a correction.

The unsettled rest

Beyond authorship, positions vary by journal and funder: whether AI may be used in peer review, whether it may appear in grant applications, what counts as "use" worth declaring. Check the specific policy of the specific venue, and note the date you checked, because these are moving.

The practical reading order from this site: verifying what you cannot judge for output outside your own methods competence · what never goes in a prompt before unpublished data or a manuscript under review goes into any account · the training toggle (Claude, ChatGPT, Gemini) for why the account type matters for embargoed work · when to not use it at all, which includes the case where the task is the learning.

What goes wrong

Listing the tool as a contributor. The rule is not "credit it somewhere". It is that authorship criteria cannot be met, and disclosure happens in acknowledgements or methods instead.

Disclosing in the wrong section. Writing help in methods, or analysis in acknowledgements, reads as not knowing the framework rather than as honesty.

Submitting unverified citations. The most common concrete failure, and the one with a misconduct process attached to it.

Treating disclosure as absolution. Every duty of accuracy, integrity and originality survives it, undivided.

Pasting a manuscript under review into a consumer account. An account-terms question that most researchers never think to ask.

How to check it worked

Before submission, take your reference list and verify every citation you did not personally read against its source. Then re-read your acknowledgements and methods and confirm the AI use you actually made is described in the section ICMJE assigns it. Both checks are mechanical, both are fast, and both are the difference between disclosure that protects you and disclosure that documents a problem.