How to disclose AI use in biomedical and clinical research papers

Biomedical AI disclosure has two jobs

Biomedical researchers now face a disclosure problem with two layers.

First, you may use AI while writing, coding, translating, checking references, drawing figures, cleaning tables, or preparing a cover letter. That use belongs in an AI usage disclosure.

Second, you may study an AI system as part of the research itself. That system may diagnose, triage, predict risk, segment images, suggest treatment, draft clinical notes, or support a clinician. That belongs in the methods, results, limitations, and sometimes the title or abstract.

Do not mix these up.

A paper can need both. A clinical prediction paper may report a machine learning model under TRIPOD+AI and still disclose that the authors used an LLM to edit the introduction. A radiology paper may follow CLAIM for the imaging model and also add an [AI Usage Card](/ai-disclosure-for-social-science-research/) for manuscript writing.

If you want a short starting point, read What are AI Usage Cards? first. If you need to check whether your journal asks for a specific statement, use [[[AI disclosure](/how-to-disclose-microsoft-copilot-use-in-academic-writing/)](/how-to-disclose-ai-use-for-neurips-icml-and-acl-submissions/) policies by major journals](/ai-disclosure-policies-by-journal/) alongside the instructions for authors.

Start with authorship

AI tools do not belong in the author list.

The ICMJE says authors should not list AI or AI-assisted technologies as authors. It also says journals should ask authors to disclose AI-assisted technologies, such as LLMs, chatbots, and image creators, when authors use them in the production of submitted work. Authors who use such tools should describe how they used them in the cover letter and in the submitted work when the journal asks for it. (icmje.org)

That rule matters in biomedical papers because authorship carries responsibility. A chatbot cannot approve a final version, answer editorial questions, verify patient consent, or take responsibility for wrong claims.

Write the disclosure around the people who used the tool.

Use this shape:

"The authors used [tool name and version] for [task]. The authors reviewed and edited the output. The authors take responsibility for the final manuscript."

That sentence does more work than a vague line like "AI was used during preparation." It tells the editor what happened. It also names the human authors as the responsible actors.

For more general wording patterns, see How to disclose ChatGPT usage in academic papers and Do I need to disclose AI usage in my paper?.

Record AI use before submission week

Most disclosure mistakes happen because authors try to reconstruct AI use after the paper feels done.

Do not wait.

Create a small running log when you start the project. The log does not need to become a second methods section. It needs enough detail that you can write a clean statement later.

Record the tool name, version if available, date range, task, input type, output type, and review step. If you used a tool only for grammar, say that. If you used it to draft a paragraph, say that. If you used it to create code for analysis, put the code review and testing step in the record.

The AI tools for research disclosure log template gives you a practical format for this. You can also generate an AI Usage Card once the manuscript stabilizes, then paste the card text into the acknowledgments, methods, appendix, or cover letter.

A biomedical editor does not need your full prompt history in most cases. They need enough information to judge whether AI touched the scientific claims, patient data, analysis, images, or wording.

Separate writing help from scientific work

Many biomedical teams use LLMs for language editing. That can fit inside a short acknowledgment or declaration if the tool did not change the study design, analysis, interpretation, or conclusions.

Example:

"The authors used DeepL Write and ChatGPT-5 to improve sentence clarity in the introduction and discussion. The authors reviewed all edits and approved the final text."

That statement tells the editor that the AI helped with wording. It does not pretend the tool did science.

If you used AI to design search strings, screen abstracts, extract data, write code, classify adverse events, label images, generate synthetic data, or interpret results, do not hide that in an acknowledgment. Put the use where a reader can evaluate it.

A systematic review, for example, should explain any AI-assisted screening or extraction in the methods. See AI disclosure in systematic reviews and meta-analyses for a deeper treatment.

A clinical data paper should explain AI-assisted coding, validation, and error checks near the analysis workflow. A reader cannot assess the methods if the AI step sits in a one-line note at the end.

Protect patient data before you paste anything

Biomedical disclosure starts before publication ethics. It starts with data handling.

If your prompt contains patient text, images, dates, locations, rare diagnoses, device identifiers, or linked study IDs, stop and check your protocol, data use agreement, IRB approval, institutional policy, and the tool terms.

In U.S. HIPAA settings, HHS describes two de-identification routes under the Privacy Rule: Expert Determination and Safe Harbor. Safe Harbor requires removal of specified identifiers, and the covered entity must have no actual knowledge that the remaining information could identify the person. (hhs.gov)

Do not treat "I removed the name" as de-identification. A rare disease, exact admission date, small clinic location, and age over 89 can identify a patient when combined.

Your AI Usage Card should not expose patient data. It should describe the kind of input at a safe level:

"The authors used an institutionally approved LLM interface to summarize de-identified interview excerpts. The prompts did not include names, medical record numbers, dates of service, contact information, or free-text identifiers. Two authors checked each summary against the source excerpt."

That wording gives the reader enough context without leaking the data you worked to protect.

If the paper studies an AI system, use the right reporting guideline

Generative AI disclosure does not replace clinical AI reporting.

If your research evaluates an AI intervention in a clinical trial, look at CONSORT-AI for trial reports and SPIRIT-AI for trial protocols. CONSORT-AI extends CONSORT for clinical trials with an AI component and asks authors to report details such as how the AI intervention entered the trial setting and whether the AI version changed during the study. (equator-network.org)

If your paper reports a clinical prediction model, TRIPOD+AI gives updated guidance for prediction models that use regression or machine learning methods. The BMJ published the TRIPOD+AI statement in 2024. (bmj.com)

If your study evaluates an AI-based decision support system during early live clinical use, DECIDE-AI fits that stage. The guideline covers early-stage clinical evaluation of AI decision support systems, regardless of the study design. (pmc.ncbi.nlm.nih.gov)

If your paper studies diagnostic accuracy, STARD-AI now covers AI-centered diagnostic accuracy studies. Nature Medicine published the STARD-AI reporting guideline in 2025. (nature.com)

If your paper focuses on medical imaging AI, CLAIM gives authors and reviewers a checklist for AI in medical imaging. RSNA introduced CLAIM in 2020 and later posted a 2024 update. (pubs.rsna.org)

These guidelines answer a different question than an AI Usage Card answers.

A reporting guideline asks: "Did you report the AI system and study design well enough?"

An AI Usage Card asks: "How did the authors use AI tools while producing this scholarly work?"

Use both when both apply. For a broader map of documentation types, see AI documentation frameworks compared, AI Usage Cards vs Model Cards, and AI Usage Cards vs System Cards.

Put the disclosure where readers will look

Biomedical journals vary. Some ask for an AI statement in the cover letter. Some ask for it in the acknowledgments. Some require a declaration section. Some fold it into methods.

Follow the journal first. Then make the disclosure easy to find.

For manuscript writing help, the acknowledgment or declaration section often works. For AI-assisted methods, write the details inside the relevant methods subsection. For AI-generated or AI-edited figures, describe the tool in the figure legend or methods and check whether the journal allows such images.

If you need a submission checklist, use AI transparency requirements for journal submissions before you upload files.

Example disclosure for a biomedical manuscript

Use this as a model, not as boilerplate.

"The authors used ChatGPT-5 through an institutionally approved account between May and July 2026 to improve sentence clarity in the introduction and discussion. The authors did not enter patient-level data, clinical notes, dates of service, medical record numbers, images, or identifiable information into the tool. The authors reviewed all suggested edits and take responsibility for the final text. No AI tool designed the study, analyzed data, interpreted results, or wrote the conclusions."

That statement gives the editor four facts: tool, task, data boundary, and human review.

If the tool touched analysis, add the analysis details:

"The authors used GitHub Copilot in RStudio to draft helper functions for data cleaning. Two authors reviewed the generated code, tested it on simulated data, and reran the analysis from raw study tables. The final analysis scripts appear in the project repository."

Now the reader can inspect the risk.

LaTeX example for a medical paper

If you write in LaTeX, add a short AI use section near declarations or acknowledgments. The exact location depends on the journal template.

\section*{AI use disclosure}
 
The authors used ChatGPT-5 through an institutionally approved account
to improve sentence clarity in the Introduction and Discussion.
The authors did not enter patient-level data, clinical notes, images,
dates of service, medical record numbers, or other identifiable information
into the tool. The authors reviewed all suggested edits and take
responsibility for the final manuscript.
 
\noindent\textbf{AI Usage Card:}
An AI Usage Card for this manuscript was generated at ai-cards.org
and is available in the supplementary materials.

If your journal allows appendices, you can place the full card in supplementary material. The LaTeX tutorial for AI Usage Cards and the Overleaf guide show how to format the card without disrupting the manuscript template.

A short checklist before you submit

Before submission, ask four questions.

Did AI touch patient data, images, interviews, code, analysis, figures, search strategy, screening, extraction, interpretation, or manuscript wording?

Did you record the tool name, version if available, task, date range, input type, and human review?

Did you follow the journal policy and any field reporting guideline, such as CONSORT-AI, TRIPOD+AI, STARD-AI, DECIDE-AI, or CLAIM?

Did you generate an AI Usage Card that matches what actually happened?

If the answer to the last question is no, generate one now. Use the free AI Usage Card generator, copy the text into your disclosure section, and keep the card with your submission records. Biomedical readers do not need mystery around AI use. They need a clear account of what you did, what you did not do, and who checked the work.