Publisher AI disclosure policies, compared

Elsevier, Springer Nature, Wiley, Taylor & Francis, SAGE and ICMJE all reach the same authorship rule: generative AI cannot be listed as an author or co-author.

The reasoning is uniform across their policies - authorship carries accountability for the work, including for its accuracy and for responding to queries after publication, and a tool cannot hold that. This is the one part of the landscape with no ambiguity in it.

Everything after this point differs by publisher. There is no general rule to follow, only a specific one to look up.

Human accountability continues beyond the byline. Authors remain responsible for verifying facts and references, protecting confidential material, checking licences and intellectual-property terms, and confirming that the final manuscript represents their work. A disclosure records use; it does not transfer responsibility to the tool or excuse an error.

• No AI authorship. A tool cannot approve the manuscript, manage conflicts of interest or answer post-publication questions.
• Human verification. Generated claims, code, images and references must be checked against original sources and research records.
• Confidentiality still applies. Unpublished manuscripts, peer-review material and sensitive data should not be entered into tools unless the relevant policy, consent and service terms allow it.
• Journal rules can be stricter. A publisher-wide policy is a baseline, not a promise that every use is acceptable in every title.

A declaration can be substantively complete and still be placed in the wrong part of the submission. Check both the publisher policy and the target journal's author instructions, including any submission-system question or cover-letter requirement.

• Elsevier - a statement at the end of the manuscript, immediately before the references, in a new section headed Declaration of generative AI and AI-assisted technologies in the manuscript preparation process. The journals policy was updated in June 2026. Watch which page you are reading: Elsevier maintains a separate policy for books and commissioned content, and the two differ.
• Springer Nature / Nature Portfolio - use of an LLM should be documented in Methods, or another appropriate section where Methods does not apply; AI-assisted copy editing as narrowly defined there is exempt.
• Wiley - in Methods or Acknowledgements, described transparently and in detail; the policy also tells authors to check the tool’s own terms for intellectual-property implications.
• Taylor & Francis - its authorship policy requires authors to acknowledge generative AI use with the tool's name and purpose, subject to the journal's instructions.
• ICMJE - its AI-assisted technology guidance says writing assistance belongs in Acknowledgements, while data collection, analysis or figure generation belongs in Methods; it also calls for disclosure in the cover letter and submitted work.
• SAGE - generative uses that affect methodology, analysis, results or conclusions require disclosure, with its current policy pointing to Methods or Acknowledgements as appropriate.

Note that the ICMJE split cuts across the others. If you used a model to help phrase a discussion section and also to help clean a dataset, those are two disclosures in two places.

A single workflow can also cross policy categories over time. Asking for an outline, generating analysis code, translating interview material and copy-editing a finished paragraph are four different uses even if they happen in one chat. Keep a contemporaneous use log rather than trying to reconstruct the workflow from memory at submission.

Several policies carve out limited assistive uses, but the definitions are not the same size. An exemption from disclosure is not an exemption from accuracy, confidentiality or human-responsibility requirements.

• Elsevier exempts basic checks of grammar, spelling and punctuation, basic reference-manager functions that do not invoke generative AI, and AI inside disability-related assistive technology used solely for accessibility.
• Springer Nature exempts AI-assisted copy editing - improvements to readability, style and error correction on your own text do not require disclosure.
• Taylor & Francis exempts grammar and spelling tools.
• SAGE distinguishes assistive tools that suggest improvements to the author's work, which do not require disclosure, from generative uses affecting research content, which require disclosure. Its current policy says authors should cite original sources rather than treating a model as the primary source; tool citation may be appropriate for representative outputs such as generated images.

A common boundary is between limited correction of human-authored text and generation that changes or adds substantive content. But the wording matters: "grammar and spelling" is narrower than "AI-assisted copy editing," and autonomous rewriting may fall outside both even when it begins with the author's paragraph.

Classify the use, not the product name. The same application can perform a spelling check in one session and generate a literature-review paragraph in another. Do not assume that starting from your own text makes a rewrite exempt: compare the actual degree of generation with the policy definition, and ask the editorial office when the boundary is unclear. Our ethics page describes HumanPen's scope, but it cannot override a journal's disclosure rule.

When reasonable readers could classify a use differently, a concise, accurate disclosure is safer than relying on a broad interpretation of an exemption.

A PNAS analysis of 5,114 journals and more than 5.2 million papers found that around 70 percent of journals had adopted some form of AI policy. Among 75,172 post-2023 papers for which the researchers could inspect full text, roughly 76 - about 0.1 percent - contained an identified disclosure statement. The study also found no statistically significant difference in the estimated rise of AI-assisted writing between journals that required disclosure and those that did not.

That result measures detectable statements, not every use of AI. Some authors may have used no tool, some uses may have fallen under exemptions, some declarations may have appeared in forms or supplements outside the searchable text, and some wording may have escaped the researchers' identification method. Absence of a detected declaration is not proof of undisclosed use.

The finding is best read as an implementation gap: policies have spread faster than visible reporting practices. It does not tell an individual author to imitate the majority. Your obligation comes from the policy governing your submission, and an editor can ask about a use regardless of how rarely other papers disclose it.

The lack of a policy effect on estimated AI-assisted writing also needs restraint. A non-significant difference does not prove policies are useless; journals may differ in discipline, enforcement, timing and author population, and text-based estimates have their own error. It shows that publishing a rule alone did not produce a clearly measurable difference in that study.

• Read your target journal’s author guidelines, not the publisher’s general page. Individual journals tighten the publisher default. The publisher policy is the floor.
• Keep a use log while working. Record tool and version, date, purpose, affected section, substantive output retained and human verification performed.
• Disclose per purpose, not merely per tool. Under ICMJE the same model used for writing and data analysis belongs in two sections, plus the cover letter.
• Describe actions precisely. "Used AI" is too vague; state whether it corrected grammar, translated author-written text, generated code, suggested an outline or drafted prose.
• Document verification and responsibility. Note that authors reviewed the output, checked original sources and take responsibility for the final content where the journal's template calls for it.
• Check non-text rules separately. Image generation, data analysis, code, peer-review material and personal data often carry restrictions beyond the manuscript declaration.
• Confirm placement at submission. Manuscript section, acknowledgements, methods, cover letter and portal questions may all be required.

A useful declaration answers four questions without marketing language: which tool and version, what it did, where it affected the work, and how humans checked the result. For example, a journal-specific statement might say that a named tool was used on author-written text in the Discussion for language editing, after which the authors reviewed every change and verified all citations. Use the journal's required template when one exists.

If the use may be prohibited - for example uploading confidential peer-review material or generating a research image contrary to policy - disclosure does not retroactively make it permitted. Stop and contact the editor or research-integrity office with an accurate account before submission.

Build a policy snapshot for the submission file. Save the publisher policy, journal instructions and portal wording with access dates; note who checked them and any answer received from the editorial office. This matters when a general policy changes between drafting and acceptance or when the portal asks a narrower question than the author guide. The record shows which rule the authors reasonably applied at each stage.

Recheck after revision, not only before first submission. A reviewer may ask for new analysis, a translated abstract, a graphical summary or extensive language changes, creating a new AI use that was not covered by the original declaration. Update the use log and statement with the manuscript rather than treating disclosure as a one-time checkbox.

Policies change. Elsevier revised its journals policy in June 2026, and SAGE's current wording differs from earlier summaries that told authors simply to cite generated content. This page was checked on 2 August 2026; save the policy and journal instructions you relied on, and recheck them immediately before submission.