The AI-disclosure wordings across major journals have converged on similar substance but continue to differ in specific requirements. An author submitting to multiple journals needs to know not just “disclose AI use” but the specific structure each journal expects. This post compares five major journals’ wordings as of mid-2026 — Nature, Cell, NEJM, JAMA, and BMJ — and draws out the practical implications for authors.
The shared substance
All five journals require: disclosure of any LLM or generative-AI use in the manuscript; identification of the tool by name and version; description of how it was used; confirmation that the human authors take responsibility for the content; refusal of AI as an author or co-author. The shared substance is settled. The differences are in the structure of the disclosure and the specific items required beyond the shared core.
Nature
Nature’s policy requires disclosure of LLM use in the Methods section (for research articles) or the acknowledgements (for editorial and review content). The required elements are: tool name, version, date of use, purpose. Nature does not permit AI-generated images or figures except where the AI generation itself is the subject of the research. The submission system asks a yes/no AI-use question and, if yes, a structured-purpose dropdown.
Nature’s idiosyncrasy is the placement in Methods rather than a dedicated declaration. The argument is that AI use is part of the research methods; the practical implication is that the disclosure is in the body of the paper rather than in a separate front-matter declaration.
Cell
Cell Press uses a dedicated AI declaration that sits alongside competing interests and funding statements. The declaration is structured: type of tool, purpose (language polishing, literature search, code generation, image analysis), confirmation that the authors reviewed and edited the output, and a verification statement.
Cell’s idiosyncrasy is the explicit verification language and the structured purpose categorisation. Cell also requires that AI use in peer review be disclosed by the reviewer to the editor — a position several other publishers have moved toward.
NEJM
NEJM’s wording is conservative and explicit. Authors must declare any use of AI tools in the preparation of the manuscript with specific identification of the tool and its application. NEJM additionally requires a writing-assistance declaration that is separate from the AI declaration; the writing-assistance declaration covers professional medical writers (often funded by sponsors) who do not meet ICMJE authorship criteria but contributed to the manuscript.
NEJM’s idiosyncrasy is the explicit separation of AI assistance from human writing assistance, both of which are required to be disclosed but in different declarations. The two are commonly conflated by authors; NEJM’s submission system enforces the distinction.
JAMA
JAMA’s policy requires structured AI-use disclosure tied to specific manuscript elements: which sections used AI, what tools were used, what their version was, what verification was performed. JAMA’s submission system asks the questions at submission and renders the disclosure in a standardised format in the published article.
JAMA’s idiosyncrasy is the section-level granularity: rather than a manuscript-level disclosure, JAMA wants to know which parts of the paper involved AI. This produces a more informative disclosure but more submission overhead.
BMJ
BMJ’s policy is among the most detailed. AI use must be disclosed in a dedicated declaration with: tool name and version; date of use; specific purpose (with controlled-vocabulary categorisation); confirmation that the AI did not propose substantive intellectual content; confirmation that the authors reviewed and verified the output; declaration of any image generation (BMJ does not permit AI-generated photographic-style images of people or events).
BMJ’s idiosyncrasies are the prohibition on AI-generated photographic images and the requirement for authors to assert that AI did not propose substantive intellectual content. The latter is a stronger version of the human-responsibility language than the other journals use.
The common-form opportunity
The variation across journals is real and creates substantial submission overhead for authors targeting multiple journals. The community would benefit from a common-form AI disclosure analogous to the Common Forms work for federal-funder personnel disclosures. NISO has begun discussion of an AI-disclosure standard; COPE has produced guidance; the publisher community has been slow to converge.
The CASRAI AI disclosure helper tool takes the author’s per-journal target and the manuscript’s AI-use facts and assembles the publisher-specific declaration. The tool is a workaround for the lack of a common form; the longer-term answer is convergence.
What authors should do
For authors submitting to a single journal, follow that journal’s specific requirements precisely. The submission system will likely reject a non-compliant disclosure.
For authors submitting to multiple journals (revisions, resubmissions after rejection), maintain a structured internal record of AI use that captures all the elements the major journals might require: tool, version, date, purpose, section-level scope, verification statement, image-generation status. The internal record is the source from which each journal-specific declaration is assembled.
For authors uncertain about whether their AI use crosses the disclosure threshold, default to disclosure. The downside of over-disclosure is minimal; the downside of under-disclosure can be retraction.
What publishers should do
Three recommendations. First, converge on a common-form structure. The disclosure substance is shared; the wording differences are largely cosmetic. The convergence work coordinated by NISO and COPE would substantially reduce author burden. Second, integrate the disclosure into the submission system as structured data, not free text. Structured data supports cross-publisher aggregation, machine-readable disclosure, and post-publication queries. Third, retain prompts and outputs where the disclosure threshold warrants it — major medical journals are already moving this direction; broader convergence would be helpful.
What the integrity community needs
The disclosure variation has implications beyond author burden. Integrity investigators looking at a retraction trail across journals face inconsistent disclosure formats; aggregating AI-use patterns across the literature requires substantial normalisation work. The case for a common-form structure is partly an integrity-research argument.
For the CASRAI community, the practical posture is to support the convergence work, to maintain the AI-disclosure helper as a stopgap, and to track the publisher-by-publisher evolution. The CASRAI GenAI disclosure domain catalogues the current wordings; we update quarterly.
