Publishers do not agree on what AI writing help must disclose

Start here

Someone who used an assistant to help write a paper, a book or an acknowledgment section has one practical question before they submit anything: what exactly has to be disclosed, and where. Three publishing bodies with different audiences — an academic journal group, an authorship standards body, and a self-publishing platform — each answer that question in writing, and their answers are not identical.

What the documents say

Nature Portfolio's editorial policy on AI, a living page retrieved 16 September 2026, requires ‘transparent declaration of AI’ from authors, reviewers and editors, states that ‘accountability for scholarly content, evaluation, and editorial decisions cannot be delegated to AI systems,’ and forbids sharing manuscripts or peer review material with ‘unsecured or public AI systems.’ The ICMJE Recommendations state that ‘chatbots should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work,’ and direct authors to disclose writing assistance in the acknowledgment section and analytical AI use in the methods section. Amazon's KDP AI Content Guidance draws a narrower line for authors publishing books: ‘AI-generated’ content — text, images or translations an AI tool created — must be disclosed when publishing or republishing, while ‘AI-assisted’ content, where a person used AI only to ‘edit, refine, error-check, or otherwise improve’ their own material, requires no disclosure at all.

Check this

Before publishing anything with any AI involvement, check which of these three models the venue actually follows: does it want disclosure of any AI role, the way Nature and ICMJE do, or only of AI-generated content as distinct from AI-assisted editing, the way KDP does? The mechanism to check is where the line falls between ‘wrote’ and ‘improved,’ because that is exactly where these three policies diverge, and a submission built for one standard can fail another.

What holds and what fails

What holds across all three is that responsibility for accuracy stays with the human author regardless of disclosure; none of the three lets AI assistance excuse an error. What fails is any assumption that ‘disclosure’ means the same threshold everywhere — a use that KDP would wave through as AI-assisted editing could be exactly what ICMJE expects named in an acknowledgment. This is an editorial reading: check the specific venue's own policy text before publishing, rather than applying whichever standard is most familiar.

• Find the exact disclosure policy of the venue you are submitting to, rather than assuming a general rule.
• Note in your own file which AI tool did what, before you forget the distinction between drafting and editing.
• If a venue's policy is silent on AI, treat that as unresolved rather than as permission.

None of these three organisations claims to speak for the others, and none is a law. They are three working definitions of a disclosure line, retrieved as each stated it on 16 September 2026, and a person moving between academic publishing, journals and self-publishing needs to know they are not the same line.