The BMJ's AI Policy: Transparency-First Rules for Medical Authors

The policy page does not describe a separate AI question on the submission form. It asks authors to put the declaration into the material they submit. Consider also including a summary of the input, the output, and the way the authors reviewed the AI output, as supplementary files or additional information for the editor. The editor may ask for more information, and may ask for information to be added to the content for internal use or for publication.

Why this matters: if your contributor-section acknowledgement, your Methods description, and any information you give the editor tell different stories, the inconsistency is easy to spot. Keep the three versions identical in substance.

The BMJ's policy asks for the name of the technology, why it was used, and how it was used. Given the journal's transparency ethos, err on the side of more detail rather than less.

Good disclosure:

"During the preparation of this manuscript, the authors used ChatGPT (GPT-4, OpenAI) to improve the readability and language of the Results and Discussion sections. The tool was not used to analyse data, interpret findings, or generate clinical recommendations. All AI-suggested revisions were reviewed by the corresponding author (A.B.) and the senior author (C.D.), both of whom verified that the revised text accurately reflected the study findings. The authors take full responsibility for the published content."

Why this works for The BMJ:

• Specifies the tool, version, and manufacturer

• Names exactly which sections were touched

• Explicitly excludes data analysis and clinical interpretation

• Identifies which authors reviewed the AI output

• Includes the responsibility statement

What wouldn't pass:

"AI was used for editorial assistance."

The BMJ's editors may ask for more information. Which tool? Which sections? What kind of assistance?

Clinical trials are The BMJ's bread and butter. Here's what a disclosure might look like for a randomized controlled trial:

*"The authors used Claude (Claude 3.5, Anthropic) to improve the language clarity of the Introduction and Discussion sections. ChatGPT (GPT-4, OpenAI) was used to assist with formatting the CONSORT flow diagram data into a table. No AI tools were used for statistical analysis, outcome adjudication, or interpretation of clinical results. The trial data were analysed using Stata 18 (StataCorp) by the study biostatistician (E.F.).

All AI-generated text was reviewed against the original draft and the trial protocol by three authors (A.B., C.D., G.H.). The authors take full responsibility for the published content."*

This level of detail might seem excessive, but for a BMJ clinical trial report that could influence NICE guidelines or WHO recommendations, the editorial team appreciates specificity.

BMJ defines AI broadly, to include large language models such as ChatGPT and any other technologies that use machine learning, deep learning, natural-language processing and similar techniques. The policy page lists no exempt tools or uses, so the table below shows how each common use maps onto the policy rather than a published exemption list. If you are unsure whether a tool counts, ask the editorial office at papersadmin@bmj.com before you submit.

The BMJ asks for a Patient and Public Involvement section in research papers. Using AI to generate a PPI section would undermine its purpose. If patients contributed to your study, their voices should come through in their own words, not filtered through an LLM.

BMJ's policy sets out a short list of consequences:

During editorial assessment and peer review:

• BMJ considers whether the way AI was used and declared is reasonable and consistent with its publication policies and practice

• The editor may ask for more information, or for information to be added to the content

• Content may be rejected on the basis of an inadequate declaration or the particular circumstances of the AI use

• BMJ may use screening tools, systematically or case by case, to help identify content generated in whole or in part by AI before publication

After publication:

• BMJ may also use screening tools after publication

• Content may be subject to post-publication changes on the basis of an inadequate declaration or the particular circumstances of the AI use

The policy does not publish a graded ladder of corrections, expressions of concern and retractions for AI cases. Treat an undeclared, discoverable AI use as a risk to the paper's standing.

BMJ's policy says it may screen submissions for AI-generated content, on a systematic or case-by-case basis, before publication and also after publication. The same policy covers reviewers: if reviewers use AI to improve word processing and language they should declare it, and they should not put unpublished manuscripts they are reviewing into publicly available AI tools where the security of confidential information cannot be guaranteed.

BMJ also says AI is developing continuously and that it will review the policy as necessary, or every six months until further notice. This is a policy to re-read before each submission.

Before writing:

• Check the BMJ author hub AI use policy for the latest wording. BMJ says it will review the policy as necessary, or every six months until further notice.

• Decide upfront which AI tools you'll use and document them as you go: the name of the technology, why you used it, and how. You don't want to reconstruct your AI use history from memory.

During manuscript preparation:

• Keep AI tools away from clinical interpretation and recommendation sections where you can. Authors are responsible for the accuracy of AI-produced content.

• If you're writing a systematic review or meta-analysis and use AI to screen or extract data, describe it in Methods and report how it was validated.

• For CONSORT, PRISMA, STROBE, and other reporting checklist compliance, AI can help format these, but the underlying data must be human-verified.

At submission:

• Include an acknowledgement of AI use in the contributor section, with a fuller description in the Methods section if the AI use was part of the research.

• Consider adding a summary of the input, the output, and how the authors reviewed the output as supplementary files or additional information for the editor.

• If you used AI for the plain language summary, say so.

After acceptance:

• If you used AI during revisions (responding to reviewer comments, rewriting sections based on feedback), update your disclosure before the revised manuscript is finalized.

• Don't assume that AI use during revision is somehow exempt from disclosure. It isn't.

In our review work on BMJ submissions, AI-policy problems usually appear as consistency failures rather than obvious misconduct. The common pattern is a manuscript that mentions language editing in the acknowledgments, describes AI use differently in the contributor statement, and leaves the Methods section silent. We see this most often when authors treat AI disclosure as a compliance sentence instead of a traceable part of the manuscript record.

Three failure patterns matter most before upload:

• BMJ AI policy disclosure names a tool but not the affected manuscript section. The statement says ChatGPT, Claude, Gemini, or Grammarly was used, but it does not identify whether the abstract, Methods, Results, Discussion, response letter, table legends, PPI section, or plain-language summary changed. Editors cannot tell whether AI touched interpretation, clinical recommendations, or only language.

• BMJ AI policy is not carried into the revision response. The Methods section says no AI was used for analysis, but the response to reviewers or revised Discussion uses AI-assisted wording. In an open peer-review model, this matters because accepted papers can expose the review history and make inconsistencies easier to spot.

• BMJ PPI or plain-language summary reads like generic AI prose. The patient and public involvement section names a patient-facing process, but the language is generic enough that it no longer sounds connected to actual patient input. That conflicts with BMJ's transparency culture even when the study design, ethics approval, and statistics are sound.

The fix is a short audit before submission: list every tool, every affected section, and every author who reviewed the output, then make the contributor-section acknowledgement and the Methods statement say the same thing. For The BMJ, the disclosure should connect the AI tool to a manuscript component and a human accountability step. A good internal check reads like this: tool used, section affected, purpose, reviewer, excluded tasks, and final author responsibility.

This is also why authors should not copy an AI disclosure from another BMJ-family journal without checking the flagship BMJ context. BMJ Open, Heart, Gut, and other BMJ titles share publisher-level rules, but The BMJ's combination of clinical consequence, open peer review, PPI expectations, and public correction culture makes inconsistencies more visible.

Common mistakes to avoid:

• Declaring no AI use because you "only used Grammarly." If you used AI-powered features of Grammarly (like the AI rewriting tool), describe them. BMJ defines AI broadly.

• Describing AI use in only one place. The policy asks for an acknowledgement in the contributor section and, where the AI use was part of the research, a fuller description in the Methods section.

• Forgetting to update the AI disclosure after revision rounds. If you used ChatGPT to redraft sections in response to reviewer comments, add that to the disclosure.

• Using AI to write the patient and public involvement section. This undermines a core BMJ value.

A The BMJ submission readiness check can check whether your BMJ submission meets the journal's transparency and disclosure requirements before you enter the submission system.

Comply proactively if:

• You used any AI tool (ChatGPT, Grammarly, Copilot) during manuscript preparation

• The policy asks for AI use to be described in the contributor section and, for research, the methods section

• Your institution has its own AI use policy that may be stricter

Less concerned if:

• You used only basic grammar and spell checking with no AI features (BMJ defines AI broadly, so check the tool)

• Your use was limited to reference management with no AI features