AI Wrote Part of the Paper. The Journal Wants to Know Which Part.
No publisher will put a model in the byline. Several will retract a paper for a model that was never mentioned. The gap between those two positions is a disclosure statement, a copyright warranty, and a question about who is accountable when the citations turn out not to exist.
The Authorship Rule Is About Accountability, Not Writing
The instinct is to treat authorship as credit for producing text. The publishing standards treat it as the opposite: authorship is the assumption of liability for the work. The ICMJE criteria require an author to approve the final version and to agree to be accountable for all aspects of it, including investigating questions about accuracy or integrity after publication.
A language model can satisfy none of that. It cannot approve, cannot be held to an agreement, and cannot answer a post-publication query. This is why the answer to "can I list the model as a co-author" is structurally no rather than a policy preference that might soften as the tools improve. Better models do not acquire the capacity to be accountable.
The practical consequence runs the other direction, and it is the part authors underestimate: because the model cannot hold any share of the accountability, all of it stays with the humans in the byline. A fabricated citation produced by a tool is the corresponding author's fabricated citation. There is no partial defence available.
What Actually Triggers a Disclosure Obligation
Publisher policies converge on a line between assistive and generative use. The line is not about which product you opened; it is about whether the tool produced content that survived into the submission.
The Copyright Transfer You Are Signing
The publishing agreement is where AI use stops being an ethics question and becomes a contract question. Whether you sign a copyright transfer or grant an exclusive licence under a CC-BY open access agreement, the document almost always contains two clauses that generative AI puts under strain.
The ownership warranty
You warrant that the work is original and that you own or control the rights being transferred. Under the US Copyright Office's human authorship requirement, material generated by a model without sufficient human creative control is not copyrightable by anyone. You cannot transfer rights in it because there are no rights in it to transfer. A manuscript with a machine-drafted discussion section is a mosaic: protected where you wrote, selected and arranged, unprotected where you did not.
The non-infringement warranty
You warrant the work does not infringe any third party's copyright. Generated text can reproduce training material closely enough to matter, particularly for boilerplate methodological description and for anything the model has seen many times. The warranty is yours, the output was the tool's, and the vendor indemnity in a consumer chatbot subscription — if one exists at all — will not name your publisher.
The indemnity
Most agreements make the authors indemnify the publisher against claims arising from breach of those warranties. This is the clause that converts a disclosure problem into a personal financial exposure, and it is typically joint and several across the author list.
The registration question
If the work is later registered with the US Copyright Office, AI-generated material must be disclaimed and the human contribution described. An institution registering a monograph or dataset built on undisclosed generated content inherits an inaccurate registration.
None of this makes AI-assisted writing unpublishable. It makes the accuracy of the disclosure statement load-bearing. A publisher that knew about the generated section before accepting the paper has no warranty claim about it; a publisher that finds out from a reader does.
Peer Review Is the Bright Line
Everything above is a spectrum. Peer review is not. An unpublished manuscript sent to you as a reviewer is confidential material entrusted to you personally, and uploading it to a third-party service is a disclosure to that service — before you get to any question about whether the provider trains on the input.
- NIH prohibits the use of generative AI in grant peer review outright, and treats a violation as a confidentiality breach.
- Publishers increasingly extend the same rule to manuscript review, and report violations to the reviewer's institution rather than simply declining the review.
- Editors are covered too: a confidential submission does not become shareable because an editor rather than a reviewer is the one pasting it in.
- An institutional deployment with a no-training contract narrows the data-protection problem but does not cure the confidentiality one, and most policies are drafted as flat prohibitions rather than vendor-conditional ones.
What a Usable Disclosure Statement Contains
"The authors used ChatGPT in preparing this manuscript" satisfies almost no policy, because it does not let a reader assess reliability. A statement that works names four things:
- The tool and version — the specific model, not the vendor brand, because behaviour differs across versions.
- What it was used for — drafting the discussion, generating analysis code, summarising the literature, translating.
- Which parts of the output survived — whether the generated text was rewritten, or appears substantially as produced.
- How it was verified — who checked the citations against the primary sources, and who validated the code against a known result.
That last element is the one editors actually use. The recurring failure mode in AI-related retractions is not stylistic — it is a reference list containing plausible papers that were never written, or a statistical result no one re-ran. A disclosure that names a verification step is a disclosure that survives scrutiny.
Frequently Asked Questions
Our co-author used AI and did not tell the rest of us. Where does that leave the group?
In the same place as any other undisclosed conduct by a co-author: the corresponding author signed the warranties on behalf of the group, and integrity investigations examine the paper rather than apportioning blame at the outset. The practical protection is procedural — ask the question explicitly before submission, record the answers, and keep the drafting history. A version history that shows which sections appeared fully formed is the evidence that separates the author who used a tool from the author who concealed it.
Does an institutional AI deployment with a no-training agreement change the analysis?
It changes the data-protection analysis and it changes almost nothing else. A no-training contract addresses where your inputs go; it does not create authorship, it does not make generated text copyrightable, and it does not remove a disclosure obligation that exists because the reader needs to assess reliability. For peer review it is the weakest of the arguments, because most policies prohibit the practice rather than the vendor.
The journal accepted the paper before we realised the reference list contained fabricated citations. What now?
Contact the editor immediately and in writing, before publication if it has not appeared and immediately after if it has. Self-reported errors are corrected; discovered ones are investigated. Fabricated references generally draw a correction if isolated and a retraction if they are load-bearing for the argument, and the difference between the two outcomes turns substantially on whether the authors or a reader raised it first.
Can we publish a paper where AI generated a large share of the text if we disclose it fully?
At most journals, yes in principle and rarely in practice. Disclosure resolves the ethics question but not the substantive one: the human authors still have to satisfy the authorship criteria on their own contribution, and a paper whose intellectual content is largely generated struggles to identify anyone who made a substantial contribution to conception, design or interpretation. Full disclosure of heavy generative use tends to convert an integrity question into a peer-review question about whether the work is original scholarship.
Does any of this apply to preprints and institutional repositories?
The copyright and human-authorship points apply identically, since they arise from the law rather than from a journal policy. The disclosure norms are catching up: major preprint servers now ask for an AI-use statement at deposit, and an institutional repository that hosts a thesis containing undisclosed generated content inherits the accuracy problem when the institution later asserts rights over the work.
The Rule That Covers Almost Every Case
If the tool produced something a reader would want to know about in order to judge the work, say so in the methods. If the tool touched a document someone else entrusted to you in confidence, do not use it at all. Nearly every publisher policy in force in 2026 is a longer statement of those two sentences.
The manuscript is the easy half. The copyright transfer you sign afterwards is where an undisclosed generated section stops being an editorial matter and becomes a warranty you have personally given.