A reviewer facing a dense manuscript and a deadline reaches for the obvious shortcut: paste it into an AI assistant and ask for a summary. It feels harmless. It is, in fact, one of the easiest ways to break confidentiality in peer review today — and unlike most breaches, it happens silently and at scale.
Why uploading a manuscript is the problem
When you agree to review, you accept a confidentiality obligation: the manuscript is unpublished, privileged work belonging to its authors. Pasting it into a general AI service transfers that work to a third party, potentially onto servers that retain it and may use it to train future models. The authors never consented to that, and you cannot undo it. Depending on the content, it can also expose unpublished clinical data or personal information.
The rule is simple: confidential manuscript text does not leave the review system. That includes abstracts, figures, tables and supplementary files.
What editorial bodies say about AI in peer review
| Body | Position relevant to reviewers |
|---|---|
| International Committee of Medical Journal Editors (ICMJE) | Reviewers must keep the manuscript confidential, which may rule out uploading it to software or other AI technology where confidentiality cannot be assured, and must ask the journal's permission before using AI to help with a review |
| World Association of Medical Editors (WAME) | Editors and reviewers should disclose any use of chatbots in evaluating manuscripts and in writing reviews and correspondence |
| Committee on Publication Ethics (COPE) | AI tools cannot be authors, because they cannot take responsibility for the submitted work |
| US National Institutes of Health (NIH) | Prohibits its own grant reviewers from using generative AI to analyse applications or write critiques; this covers NIH funding review, not journals |
COPE's point about authors applies equally to reviewers: a tool cannot answer for an assessment. The ICMJE Recommendations are updated periodically, so read the current text on the ICMJE website. Where a journal's own instructions are stricter, follow the journal.
What AI can legitimately help with
The line falls between your own words and the authors' words:
| Generally acceptable | Not acceptable |
|---|---|
| Polishing the grammar and tone of comments you wrote yourself, provided they do not quote or describe the authors' unpublished data | Uploading the manuscript, figures or data for summary or critique |
| Asking general questions about a method, in the abstract, without manuscript text | Asking a tool to generate the review |
| Using a journal-provided tool that is contractually bound to confidentiality | Using consumer tools that may retain inputs |
The deciding questions are: does this send the authors' unpublished content anywhere, and is the judgement still mine?
Five questions before you use any AI tool on a review
- Does the journal allow it? If its instructions say nothing, ask the editor first.
- Will any manuscript text, figure, table, data or supplementary file enter the tool? If so, stop, unless the journal supplied the tool.
- Could your own comments identify the study, its data or its authors? If so, remove those details before any editing tool sees them.
- Is every judgement in the report one you reached yourself and can defend?
- Will you tell the editor which tool you used and for what?
Why the judgement must stay human
Peer review is an accountable expert opinion. A language model can produce a confident, well-structured critique that is subtly wrong — inventing references, missing a fatal design flaw, or objecting to something standard in the field. It cannot declare a conflict of interest, cannot be held responsible, and cannot know what it does not know. An editor weighing your recommendation is relying on your expertise; substituting a generated text misrepresents what you supplied.
Disclose what you used
If you used any AI assistance, tell the editor and say how. Undisclosed AI-written reviews are being detected — through characteristic phrasing and impossibly fast turnaround — and they damage the reviewer's standing far more than a late or short review ever would. Journals are actively developing policy here, so check the specific journal's instructions before you start.
A short statement in your confidential comments to the editor is enough:
AI use: I used [tool name] only to correct the grammar of comments I had already written. No manuscript text, figures, data or author details were entered into the tool. All assessments and recommendations in this report are my own.
For editors
State a policy in your reviewer instructions rather than assuming reviewers know. Make clear that manuscripts must not be uploaded to external tools, ask reviewers to disclose assistance, and be alert to reports that are fluent, generic and free of the specific detail that shows the manuscript was actually read. Our guides on reviewing well and AI in research writing cover the author-facing side.
Sample wording for reviewer instructions
Manuscripts sent for review are confidential. Do not upload any part of a manuscript, including figures, tables and supplementary files, to an AI tool or other external service. If you use AI to edit the wording of your own comments, tell us which tool and how you used it. The assessment and recommendation must be your own. Contact [editorial office email] before using any other tool.
When a report looks machine-written
Do not treat an AI-detection score as proof. Check instead whether the report mentions content the manuscript does not contain, such as references, analyses or figures. Then ask the reviewer, neutrally, how the report was prepared. If the report cannot be relied on, set it aside and invite another reviewer rather than deciding on thinner evidence. Our guides to running fair and ethical peer review and to deciding between conflicting reports cover the next steps.
Updated 17 September 2026: We expanded this guide with a table of ICMJE, WAME, COPE and NIH positions, a five-question check, a disclosure template and editor wording.