Can I use AI to help write a peer review?

Peer review is treated differently from authorship, and more strictly, for one reason: the manuscript is not yours. Authors submitted it in confidence, they have not published it, and in many cases they have not yet secured priority for the finding. Sending that text anywhere outside the review process is a disclosure of someone else's unpublished work.

Funders moved first and clearly here. The US National Science Foundation announced in December 2023 that reviewers may not upload proposal or review content to non-approved generative AI tools. The European Research Council clarified in March 2026 that reviewers may not use AI to summarize proposals, assess merit, or draft evaluations, allowing only limited language polishing that does not involve proposal content or personal data.

The list is longer than those two bodies. The NIH prohibits AI from analyzing or generating review content, the CIHR College of Reviewers prevents generative AI use in research-proposal review, and the European Research Council states that AI must not replace humans in grant evaluation (Helmy et al., 2025, p. 4). Reviewer guidance written for computational biologists puts it more directly, describing the use of language models for reviews as almost always strictly forbidden, because sending manuscript content to an external agent can breach the venue's privacy policy (Rahman and DeBlasio, 2026, p. 2). The same guidance treats confidentiality as a core reviewer duty rather than a formality (Rahman and DeBlasio, 2026, p. 3).

Journals have followed the same logic. The consistent requirements are that the assessment must be the reviewer's own, the manuscript must not be uploaded, and any assistance should be disclosed to the editor.

There is also a quality argument worth stating separately. The value of a review is a specialist's judgment about whether the work is sound and what it contributes. A generated review reads as competent and generic, and editors have become good at spotting it, which is a reputational risk to set against the time saved.

The measurements back that up. Liang and colleagues ran GPT-4 over 3,096 papers from 15 Nature-family journals and 1,709 ICLR papers and compared its comments against the human reviews. Overlap with human feedback averaged 30.85% for the Nature papers and 39.23% for ICLR, against 28.58% and 35.25% between two human reviewers (Liang et al., 2023, p. 1). What it misses matters more than that headline number. GPT-4 raised research implications 7.27 times more often than human reviewers and raised novelty 10.69 times less often, and it was 2.19 times more likely to ask for experiments on additional datasets while humans were 6.71 times more likely to ask for ablations (Liang et al., 2023, p. 17). Novelty is usually the exact question the editor needs you to answer.

What does uploading a manuscript to a chatbot risk?

A confidentiality breach, which is the reviewer obligation taken most seriously. You are handling unpublished work that the authors have not released, and sending it to an external service is a disclosure regardless of what the service does with it. Several funders now prohibit this explicitly.

The consequences are real. Editors remove reviewers, and a demonstrated breach can affect your standing with a journal or a funding body for a long time. Grant review has formal enforcement attached.

The obligation covers the review you write as well as the manuscript, since a detailed review discloses the manuscript's content by implication.

What do publishers currently allow?

Positions vary and are tightening. The common ground is that the review must be your own judgment, the manuscript must not be uploaded to external tools, and any AI assistance should be disclosed to the editor. Check the specific journal's reviewer guidance, since several now ask directly.

Where assistance is permitted, it is usually confined to language: improving the clarity of a review you have already written, with no manuscript content passed to the tool. That is a narrow allowance and it is worth reading the exact wording before relying on it.

Where a venue does permit assistance, validation is part of the work. The recommended practice is to refine your prompts, look at several outputs, and cross-check anything generated against empirical evidence, domain expertise, or established knowledge before using it, then disclose the platform, the version, and the role it played (Helmy et al., 2025, p. 9). The sources do not agree on how far this extends. Liang and colleagues found that 57.4% of 308 surveyed researchers rated GPT-4 feedback helpful or very helpful, and 82.4% found it more useful than feedback from at least some human reviewers, but they position that tool for early manuscript preparation when timely expert feedback is unavailable, which is a different situation from a formal review you agreed to deliver (Liang et al., 2023, p. 1).

If you are short of time, the better response is to decline the review or ask for an extension. Editors much prefer a late review or a decline to one they cannot trust.

One use is uncontroversial and worth mentioning: checking a claim in the manuscript against the literature you already hold, without sending the manuscript anywhere. Verifying that a cited source says what the authors report is exactly the reviewing work that gets skipped under time pressure.