ICMJE Authorship Criteria and AI Disclosure: What Changed in 2025

The four-part authorship test remains unchanged

Authorship under the ICMJE rules still requires meeting all four criteria, not just most of them.

• Substantial contributions to the conception or design of the work, or to the acquisition, analysis, or interpretation of data.
• Drafting the work or reviewing it critically for important intellectual content.
• Final approval of the version to be published.
• Agreement to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part are appropriately investigated and resolved.

A researcher who secures funding but does not draft or approve cannot be an author. A tool that generates text cannot satisfy the final two criteria at all, because it has no legal or professional standing and cannot take responsibility. That is why the AI-specific guidance rarely needs to argue against machine authorship; the existing test already excludes it.

Disclosure rules for AI-assisted writing

The AI-specific language appears in the section of the ICMJE recommendations on artificial intelligence-assisted technology. It asks authors to disclose the use of AI in the cover letter and, if AI was used, to specify in the methods or acknowledgements section, or in another section the journal accepts, which tool was used and for what purpose. Authors retain full responsibility for every part of the manuscript, including parts produced by an AI-assisted technology, and must be able to assert that there is no plagiarism in text or images produced by AI.

The disclosure standard separates spelling and grammar checkers, which typically do not require reporting, from tools that produce substantive sentences, tables, abstracts, or images. Editors want the second kind flagged at submission, not after review. COPE's guidance on AI tools follows the same line and adds that human accountability should extend to any automated content that reaches a reader.

Photo by Markus Winkler on Unsplash

The base rate, and Dr. V's two-day detour

A 2023 Nature survey of roughly 1,600 researchers found just under a third had used generative AI in manuscript preparation. That is the base rate before the 2025 ICMJE update landed, and it explains why editors now ask the AI question as a matter of routine. The exception is rare: a researcher who keeps a dated prompt log and a record of which paragraphs came from a tool can answer a disclosure request in minutes. Most people cannot, because drafting happens across platforms and versions.

A colleague — call her Dr. V — ran a discussion section through a commercial language model to smooth the syntax, then answered 'no' on the submission form because she considered the change cosmetic. The journal's editor asked which tool she used and which sentences it had touched. Reconstructing that took her two days, mostly spent comparing old and new drafts against the raw data. The manuscript was accepted, but the process added a demand that could have been handled in one sentence at submission. The policy did not fault the use; it faulted the missing trail.

Journals agree on accountability but differ on placement

Journals that follow the ICMJE framework now ask for AI disclosure in their submission forms, while some also require a version in the methods or acknowledgements. A few journals direct authors to report tool names, prompt scope, human verification, and a named fallback for checking generated text. That variation is less significant than the shared bottom line: the human authors remain responsible for every sentence and image. The same concern appears in discussions of generative AI in peer review policies and in publisher efforts to detect AI citation stuffing in manuscripts, where generated references add volume without adding verification.

Because the ICMJE recommendations are not a law, each journal can impose stricter rules. Some require authors to state that no AI was used if that is true, closing the silent middle ground. Others ask only that authors describe substantive use and leave grammar-checking tools unmentioned. The safest reading for a submitting author is to over-disclose slightly: a sentence in the cover letter costs less than a revised disclosure later.

What this means for your lab

The fix is a habit, not a policy document. For each manuscript, keep one disclosure note with the tool name, version date, and the section it touched. Paste that note into the cover letter and into the methods or acknowledgements in the form the journal asks for. The corresponding author should confirm that every AI-generated sentence has been checked against the source data, especially for numbers, p-values, and names.

Photo by Jonathan Kemper on Unsplash

• Draft the AI disclosure while the methods section is still open, before submission forms force a binary answer.
• Record the tool version and the date of use, because some journals ask for both.
• Use a shared folder or a single lab document so a departure doesn't take the disclosure trail with it.
• Review the journal's AI question when you upload files, not after the editor writes back.

Start with the next submission, not the backlog

Choose the next manuscript your group sends out and write one disclosure sentence while the methods section is open. That sentence needs four facts: the tool, the purpose, the section affected, and the date of use. Paste it into the cover letter before the editor has to ask. It removes the most common friction researchers encounter under the 2025 ICMJE update and takes less time than reconstructing a draft two days later.