Generative AI disclosure requirements arrived quickly, were written independently by each publisher, and differ from one another in ways that matter. There is no single rule you can learn once. There is a set of questions you should ask of every journal you submit to, and the answers change.
The one point of universal agreement
Every major publisher agrees on this: an AI system cannot be an author. Authorship entails accountability for the work, and a tool cannot be accountable. This is settled across COPE, ICMJE, and the author guidelines of every large publisher.
Beyond that, agreement thins out fast.
The questions to ask of each journal
What uses require disclosure?
Policies vary from “any use of generative AI at any stage” to “AI use in the generation of content, but not in language editing”. Under the broader policies, grammar and translation tools may be disclosable. Under the narrower ones, they are explicitly exempt. Read the exemption, because it is where the practical difference lives.
Where does the disclosure go?
The common locations are a dedicated statement before the references, a line in the methods section, the acknowledgements, or a field in the submission system. Some publishers require it in two places. Putting it in the wrong place is a compliance failure even when the disclosure itself is complete.
What must the disclosure contain?
Most policies want the tool and version, what it was used for, and a statement that the authors reviewed and take responsibility for the output. Some want the prompt strategy described. Very few want the prompts themselves.
Does the policy extend to figures?
This is where policies are strictest and least well known. AI-generated or AI-enhanced images are prohibited outright by several publishers, with narrow exceptions for work where the image generation is itself the object of study. The prohibition often covers enhancement of real data images, not only wholly synthetic ones.
Does it extend to peer review?
Increasingly yes, and increasingly restrictively. Several publishers prohibit uploading a manuscript under review to any external AI system, on confidentiality grounds. If you review, this applies to you.
Writing a disclosure that will not be queried
A usable disclosure is short, specific and unhedged. It names the tool, the use, and the accountability.
During the preparation of this manuscript the authors used [tool and version] to [specific task, for example: improve the readability of the discussion section]. The authors reviewed and edited the output and take full responsibility for the content of the publication.
What weakens a disclosure: vagueness about which sections, passive phrasing that avoids saying who used the tool, and omitting the accountability sentence, which is the part editors are actually checking for.
The judgement call on language editing
Many authors, particularly those writing in a second language, use AI tools for language improvement and are unsure whether this needs declaring. The honest position: under a policy that exempts language editing, it does not; under a policy covering “any use”, it does.
Where the policy is ambiguous, disclose. A disclosure that turns out to be unnecessary costs you one sentence. An omission that turns out to be required is a correction notice on a published paper.
Why this keeps changing
These policies are two or three years old and being revised continuously as publishers observe what actually arrives in submissions. A policy you read eighteen months ago is likely to have changed. Check the version and date on the policy page itself rather than relying on a summary, including this one.