Disclosing AI Use in Your Manuscript: What Journals Require

Journals that follow ICMJE should ask whether you used AI tools. The principles are settled, but which uses count and where you declare them still differ by journal.

Guide · 10 min read · Directive Publications Editorial Team

What the main standards agree on

Three documents underpin many medical journal AI policies. The ICMJE Recommendations, in the version updated January 2026, cover AI in the authorship section and in a dedicated section on the use of AI in publishing. The COPE position statement on authorship and AI tools and WAME's recommendations on chatbots and generative AI, revised in May 2023, point the same way.

Across all three, the core points are the same:

1. AI tools cannot be authors.
2. If you used AI, say which tool and how you used it.
3. You are responsible for everything the tool produced, including errors, bias and plagiarism.
4. Where AI touched the research itself, describe that use in the methods.

What varies is the detail: which uses count, which section holds the statement, and what is allowed in figures. Read the target journal's current instructions before you write your statement, not after — the rest of this guide is about the differences you will meet when you do.

Which uses need a statement

WAME sorts chatbot use into three kinds: simple word-processing tasks, generating ideas and text, and substantive research. That split is a practical way to audit what you and your co-authors actually did.

Springer Nature takes a risk-based approach. Using an AI tool to polish language is low risk and permitted, extensive copy editing or writing support calls for caution and disclosure, and fabricating data or citations, or using a tool to generate core research reasoning without disclosure, is not permitted.

Decision rule. If a tool changed what your paper says, how it argues, or what it shows, disclose it. If it only corrected spelling and punctuation, follow the journal's policy, and where the policy is silent, add a one-line statement.

Audit by task, not by tool. The same chatbot may have done three different jobs on one manuscript — tidying a co-author's English, drafting a paragraph of the discussion, and writing the script that cleaned your dataset — and each lands in a different row of that table. The reference manager's AI summariser and the plug-in inside your word processor count too, even though neither feels like "using AI". For each manuscript, disclose at the level of the most substantive use.

Where the statement goes

The ICMJE Recommendations tie the location to the purpose:

• Writing assistance belongs in the acknowledgments section.
• Data collection, analysis or figure generation belongs in the methods, described in enough detail to replicate the approach, including the tool, its version and the prompts where applicable.
• The cover letter should also describe how you used AI. ICMJE asks for the description in both the cover letter and the manuscript.

Publishers adapt this. Elsevier asks for a separate declaration section at the end of the manuscript, immediately before the references, and for AI-assisted explanatory images, a note in each caption as well. The Science journals ask authors to note AI use in the cover letter and in the acknowledgments, and to give details in the methods, including the full prompt, the tool and its version. COPE's statement points to the materials and methods or a similar section.

ICMJE also says journals should require authors to disclose at submission whether they used AI, so the submission system may ask you directly. Your answer there, your cover letter and your manuscript statement must say the same thing. If you are drafting the letter now, our guide to writing a journal cover letter includes a declarations paragraph where one AI sentence can go.

What to write, and what to keep

A useful statement answers five questions: which tool and version, made by whom, for what task, on which parts of the manuscript, and how the authors checked the output. For research uses, WAME adds the full prompt, the time and date of the query, and the tool's version.

Adapt these rather than copying them word for word.

Acknowledgments, for writing assistance:

Methods, for research use:

Cover letter:

Keep the evidence while you work. Elsevier recommends a separate record of the tool and model used, saved prompts and outputs, and tracked changes or annotated drafts showing how the text was reviewed, and says authors should be able to provide this documentation if editors ask. For medical research, the EQUATOR Network lists GAMER, a reporting guideline for generative AI use in medical research published in 2025. Its nine items are a sound template for a detailed disclosure: general declaration, tool specifications, prompting techniques, the tool's role in the study, any new models developed, AI-assisted sections of the manuscript, content verification, data privacy, and impact on conclusions.

As the senior author, set one rule for the group: every co-author reports their AI use to the corresponding author before the manuscript is finalised, and it goes into a log kept with the project files.

Why AI cannot be an author, or a source

ICMJE's reasoning is direct. Chatbots cannot be responsible for the accuracy, integrity and originality of the work, and authorship requires those responsibilities. COPE adds that AI tools, as non-legal entities, cannot assert the presence or absence of conflicts of interest or manage copyright and licence agreements.

That reasoning has two practical consequences.

Do not cite AI as a source. ICMJE says referencing AI-generated material as the primary source is not acceptable, and that authors should be able to attest that the references they cite support the associated statement. Elsevier warns that AI-generated references can be incorrect or fabricated. Check every reference against the original, not against the tool's summary of it.

Your sign-off covers everything. Approval of the final version is one of the four ICMJE authorship criteria. When you approve a manuscript, you approve the passage a junior colleague drafted with a chatbot as well. Our guide to ORCID, affiliations and authorship covers the four ICMJE criteria and how to record each contributor's role.

AI-generated images and figures

• The Science journals do not permit AI-generated images and other multimedia without explicit permission from the editors. Exceptions may be granted case by case, for example in manuscripts specifically about AI or machine learning, and should be disclosed at submission.
• Elsevier sorts images into explanatory images, research and data images, and graphical abstracts and cover art. AI tools may support explanatory images such as flow charts, decision trees and schematics, with disclosure in each caption. Within research and data images, data visualisations may use AI only when the output is directly derived from the underlying data, and primary research images, including microscopy, histology, western blots, radiology scans and patient images, must not be created or altered with AI tools. General-purpose generative image tools must not be used for graphical abstracts; cover art may be permitted, but only with prior permission from the journal editor and the publisher.
• Springer Nature lists creating photorealistic deepfake images among the uses that are not permitted.

A workable rule for medical authors:

• Clinical images, scans, blots and micrographs: no generative AI editing. If AI-assisted imaging is part of your research method, describe it in the methods as a method.
• Schematics and study flow diagrams: AI assistance is the most likely to be permitted here, and the caption is where you say so.
• Graphical abstracts and cover art: assume a general-purpose image generator is not acceptable unless the journal explicitly says otherwise.

Confidentiality: what not to paste into a tool

Elsevier tells authors to check a tool's terms and conditions to protect the confidentiality of their data and unpublished manuscripts, to take particular care with personally identifiable data, and not to grant the tool rights to train on their material. Springer Nature says manuscripts, peer review reports and sensitive data must not be shared with unsecured or public AI systems.

For clinical research, never paste identifiable participant data into a consumer tool. Check your ethics approval, consent terms and institutional data policy first, and ask whether your institution provides an approved tool.

The same applies when you review. ICMJE says reviewers must follow the journal's AI policy or request permission before using AI, should not upload a manuscript into an AI system where confidentiality cannot be assured, and should disclose any use to the journal.

If undisclosed AI use comes to light

ICMJE states that nondisclosure of AI use may require corrective action and may be construed as misconduct in some circumstances. COPE's statement says authors are liable for any breach of publication ethics, including in parts produced by an AI tool. The Science journals state that editors may decline to move forward with manuscripts if AI is used inappropriately.

What happens next depends on what the tool did:

• Missing note about language help, found before acceptance: tell the editor and update the statement in your revision.
• Missing statement, found after publication: contact the journal, explain exactly what was used, and ask how they want the record updated.
• Fabricated references, altered images or AI-generated analysis presented as your own: this goes to the reliability of the paper. Under ICMJE, errors serious enough to invalidate results may require retraction, and editors may publish an expression of concern while misconduct is investigated.

In every case, raise it yourself before a reader does. Our guide to correcting the published record walks through the options.

Before you submit

1. Every co-author has reported which AI tools they used, and for what
2. You have read the target journal's current AI policy, including its image rules
3. The statement names the tool, version, developer, task and sections affected
4. Any research use is described in the methods, with prompts and dates
5. The submission form, cover letter and manuscript say the same thing
6. Every reference has been checked against the original source
7. No primary research image has been created or altered with an AI tool
8. Prompts, outputs and tracked changes are saved with the project files

The standard to hold yourself to

Good AI disclosure in research papers is short and specific. Write the statement you would be comfortable defending to an editor who has just asked for your prompts and drafts, because Elsevier, for example, says authors should be able to provide documentation of their AI use if editors request it.

Directive Publications sets out its author guidance and publication ethics policies on its For Authors page. Check them against your manuscript, as you would for any journal, before you submit your manuscript.