Do I need to disclose AI use in a research paper?

Publisher policies converged faster than most people realize. Between 2023 and 2026 the major houses arrived at roughly the same three rules: an AI tool cannot be an author, substantive use must be declared, and the human authors remain fully responsible for everything in the paper including anything a tool produced.

Where they still differ is on placement and on the size of the exemption. Elsevier asks for a dedicated declaration immediately before the reference list. Springer Nature and Nature Portfolio titles want substantive use documented in Methods. PLOS requires a Methods subsection naming the tool, its use, and how the output was validated. IEEE puts it in Acknowledgments with the affected sections named. Science asks for the cover letter and the acknowledgments, and for the full prompt where the tool produced part of the work.

The practical consequence is that you cannot write one disclosure and reuse it. Check the target journal's instructions for authors immediately before submission, because publisher-level policy is a floor and individual journals impose more.

A comparative review of editorial policies shows how far the required detail varies. ICMJE asks for the purpose and extent of AI use, Nature places the role and scope in Methods, BMJ asks for the tool, the purpose, and the content generated, and Science requires the tool name, version, and the prompts in Methods, with disclosure repeated in the cover letter and the acknowledgments. Science also prohibits AI-generated citations and may decline to send a manuscript for review when a tool has been used inappropriately (Yoo, 2025, p. 6).

The reason for all this is accountability rather than suspicion of the tools. Disclosure lets readers judge what the AI actually did, while the human authors stay responsible for accuracy, integrity, and originality (Yoo, 2025, p. 5). Undisclosed use becomes evidence in a research-misconduct investigation when the tool contributed to errors, copyright infringement, or data manipulation (Yoo, 2025, p. 9). Some publishers now go further and ask authors to explain how the tool was used, and the requirement extends across CAAA Journals, Elsevier, IEEE, Frontiers, PLOS, ICMJE, COPE, and WAME, while JAMA Network, Nature, and Cell Press prohibit AI authorship outright (Helmy et al., 2025, p. 6).

What to record in the disclosure is more consistent: the tool and version, what it was used for, which parts of the manuscript it affected, and a statement that you reviewed and take responsibility for the output. Keeping notes as you go makes this a five-minute task rather than a reconstruction.

What do the major publishers require?

All of them require disclosure of substantive use and all refuse AI authorship. Elsevier wants a declaration section before the references. Springer Nature, PLOS, and Nature want it in Methods. Wiley, IEEE, and Taylor and Francis accept Acknowledgments or Methods depending on the use. Science wants the cover letter as well.

Medical journals are stricter. JAMA discourages AI-created content in submissions unless it is part of the study design, and asks for the tool, version, manufacturer, dates used, and the content affected. BMJ extends its policy to AI used in the underlying work and, as far as authors know, in important cited sources.

The direction of travel is towards more detail rather than less, so a paper being written now should record more than the current rule requires.

Which uses are exempt from declaration?

Usually basic spelling and grammar checking, reference formatting, and light copy-editing of text you wrote. Springer Nature exempts AI-assisted copy editing explicitly. Wiley exempts simple grammar and limited word-choice help. Anything that changes meaning, generates content, or touches analysis is not exempt anywhere.

Translation sits awkwardly. Some publishers treat it as assistive and some as substantive, and at least one university prohibits AI translation of a thesis drafted in another language. If translation was material to your text, declare it.

One review sorts 11 use scenarios into three tiers, which is a usable test when a journal's own wording runs out. Generating research hypotheses, writing manuscript sections, collecting or analyzing data, and producing figures or tables all require mandatory disclosure. Grammar correction and sentence refinement are optionally disclosable as long as you keep control of the content. Spell-checking, reference formatting, and literature searching are routine and normally need no declaration (Yoo, 2025, p. 7).

Do not expect the boundary to hold still. There is still no cross-journal consensus and no standard way to report AI use across the different stages of preparing a manuscript, even though ICMJE, COPE, WAME, and the European Commission all support disclosure and reject AI authorship (Mehta et al., 2026, p. 1). One proposed fix is a disclosure form covering three domains of use that outputs a standardized statement with a persistent, citable reference. Until something like that is normal, follow the stricter of your institution's rule and the journal's rule.

When you are unsure, declare. No paper has been rejected for disclosing a use that turned out to be exempt, and undisclosed use has produced retractions.