AI use in a medical paper should be disclosed according to the destination journal's requirements, with humans retaining responsibility for the submitted work. A polished paragraph does not establish that its references exist, its analysis is correct or patient information was handled appropriately. Keep a record of what the tool actually did before deciding how to describe it.
The disclosure starts with the workflow
An author who says "AI helped with the paper" has not yet given an editor useful information. Assistance might mean correcting grammar, suggesting an outline, extracting information, generating code or drafting an interpretation. Those uses create different questions about reproducibility and verification.
At the start of a project, keep a short activity log: tool and version where available, date, purpose, material supplied and output retained. Add the human checks performed. This is a proposed working record, not a universal journal-mandated form.
The log prevents a common problem near submission: nobody remembers whether a sentence came from the original analysis, a collaborator or a generated draft. Reconstructing the provenance at that stage is much harder than recording it during the work.
What ICMJE currently requires
The ICMJE recommendations on AI use by authors, checked on 19 September 2026, call for disclosure at submission and a description in the cover letter and relevant part of the manuscript. They retain human responsibility, exclude AI tools from authorship and require authors to check generated material and attribution. AI-generated content is not acceptable as the primary source for a substantive claim.
The destination journal may provide additional instructions. Read those alongside the recommendations rather than assuming that a disclosure accepted by one journal will satisfy another. Where the policy is unclear about a material use, ask the editorial office a specific question before submission.
Disclosure is not an alternative to verification. "Generated with AI" does not transfer responsibility for a false reference or an unsupported conclusion away from the authors.
A writing edit and an analytical contribution are not the same
Imagine an original fictional project with two uses of AI. In the first, a tool rewrites an author-written introduction for clarity. In the second, it generates code that derives the primary outcome from a dataset.
For the first use, the authors need to check that meaning, qualifications and attribution survived the edit. A grammatical improvement can still change "was associated with" into "caused", or remove a limitation that made the sentence accurate.
For the second, checking the prose is insufficient. The authors need to inspect the data definitions, exclusions, calculations and outputs. Code that runs without an error message may still implement the wrong denominator or analyse the wrong population.
These examples are not observations from a tested product. They illustrate why "we reviewed the AI output" should describe a task-appropriate check rather than an unspecified read-through.
Verify references at the claim level
For each reference supplied or influenced by AI, locate the actual source. Confirm its identity, inspect the relevant passage and determine whether it supports the statement being made. A real paper can still be the wrong citation.
Pay particular attention to numerical claims, causal language, novelty statements and assertions about clinical practice. Check whether a result concerns the same population and product version, and whether an apparent finding is actually speculation in the discussion.
Publication status also matters. Crossmark and publisher notices can identify updates, but the underlying record needs inspection when a claim is important. Crossref's Crossmark documentation describes that update pathway; it is not a substitute for reading the notice.
Do not cite the AI conversation as though it were the primary evidence for a clinical claim. Cite the underlying research or authoritative source, and describe AI assistance separately where required.
A case report needs a separate confidentiality decision
A patient's story can remain identifiable after removing a name. A distinctive timeline, unusual combination of events or recognisable image may matter. The ICO's anonymisation guidance, checked on 19 September 2026, distinguishes effective anonymisation from pseudonymisation and considers identifiability in context.
Permission to publish a case should not be assumed to authorise uploading its source records to any external AI service. Establish the approved environment, applicable permissions and institutional requirements before sharing material. Check images, attachments and metadata separately from the narrative.
For an early writing exercise, a fully invented case may be preferable. Label it as fictional and do not later present it as a real patient's course, a clinical case report or evidence of treatment effectiveness.
Example disclosure wording, with honest limits
For a hypothetical language-editing use, an author might write: "The authors used an AI-assisted tool to suggest language edits to an author-written draft. All suggested changes were reviewed against the original meaning and sources. The authors take responsibility for the final manuscript."
That example intentionally leaves the actual tool, version, dates and manuscript location to be completed according to the journal's policy. It should not be copied as a claim that no other AI use occurred.
For analytical use, a generic writing statement would be inadequate. Describe the relevant role in the methods and retain the code, prompts or other information required for the work to be inspected. Do not claim reproducibility from a tool name alone when the input, workflow and version are unspecified.
Keep the contribution and the evidence separate
An AI tool may make a writing task more convenient without becoming an author or a source of evidence. A reference platform may help identify an explanation without replacing the cited guideline or paper.
iatroX's appropriate role is clinical reference and learning. Its methodology available on 19 September 2026 describes source-linked retrieval and checking processes, not authorship or guaranteed verification of a manuscript. A clinician can use the resulting learning to understand a concept, while still checking the underlying sources and making the research judgements personally.
Before submission, the author group should be able to explain the central result without relying on the tool's authority. If nobody can defend the method or trace the evidence, the problem is not a missing disclosure sentence; the work needs further review.
Frequently asked questions
Can an AI tool be listed as an author?
ICMJE does not permit AI tools to qualify as authors because they cannot take the required responsibility. Human authors remain accountable for the submitted material.
Is disclosure enough if the AI supplied the references?
No: verify each reference and the claim it is intended to support. Disclosure does not excuse inaccurate or inappropriate citation.
Can I use one standard disclosure for every journal?
Use the destination journal's current instructions and describe the actual workflow. A generic statement may omit information needed for a particular use or submission.
Explore clinical research and advisory support through iatroX Insights →