The ICMJE Added a Section on Artificial Intelligence in January 2026

The ICMJE is a small working group of medical journal editors, not a regulator. It has no power to compel anybody. What gives the Recommendations their force is adoption: a journal writes in its own instructions for authors that it follows them, and from that moment they are that journal’s rules and a submission can be returned for breaching them.

The document covers who may be named an author, what has to be disclosed, when a trial has to be registered, how participants are protected, and what a journal may require of a sponsor. The January 2026 revision keeps that shape and adds two sections.

Section V, artificial intelligence. The ICMJE announcement describes it as “expanded guidance on the use of artificial intelligence (AI) in publishing”. The Recommendations’ existing guidance already states the hard rule: chatbots such as ChatGPT should not be listed as authors, because they “cannot be responsible for the accuracy, integrity, and originality of the work” — and authorship requires exactly that responsibility.

Section III.L.2, authors’ access to data. This one is quotable in full and worth quoting: “All authors should have the ability to review the data supporting study results.” In academic–industry collaborations, “at least one author from the academic institution should have accessed the primary data and be able to participate in the analysis.” And with sponsored research, that access “should be expressly provided for in any research agreements.” Editors “may ask which specific authors had access and participated in the analysis.”

Both additions answer the same problem from opposite directions: who is answerable for what appears in a paper.

A generative model can produce a methods section, a discussion and a reference list in a minute, and none of it can be held to account. Naming it as an author would put a signature on the paper that cannot be asked a question. So the rule is not that AI is forbidden — it is that a human stays answerable and the use is declared.

Authors’ access to data answers the mirror image. A paper can carry ten academic names while the analysis sat entirely with a sponsor, and the named authors could not have checked the result if asked. Writing data access into the research agreement, before the study starts, is the only point at which that is fixable.

How a journal applies it. The ICMJE maintains a list of journals that have told it they follow the Recommendations, but the operative test is simpler: read your target journal’s instructions for authors. If the phrase “ICMJE” appears there, the Recommendations are part of that journal’s submission conditions and an editor can hold you to them.

In practice journals implement the same rules through different mechanics. Some ask for a tick-box AI declaration in the submission system; some want a named statement in the manuscript; some ask for the ICMJE disclosure form from every author. The obligation is the same; only the paperwork differs, which is why checking the specific journal still matters.