03 · What You Need to Know
Where Generative AI May Help and Where Peer-Review Responsibilities Set Limits
Peer Review Is an Independent Scholarly Assessment
Peer review involves critically examining a submitted manuscript to help editors assess its suitability for publication and help authors improve their work. Depending on the journal and discipline, reviewers may evaluate methodological appropriateness, evidentiary support, reporting completeness, originality, and the interpretation of findings.
The reviewer's responsibility extends beyond producing a polished report. Reviewers are expected to understand the work sufficiently to offer informed criticism, identify the basis for their concerns, and distinguish substantial weaknesses from matters of preference.
Generative AI can produce text resembling a peer-review report. That capability does not establish that the system has evaluated the manuscript accurately or that its comments are relevant.
The ICMJE's recommendations emphasize that peer reviewers must provide constructive, honest assessments while maintaining the confidentiality of submitted manuscripts. These obligations remain with the human reviewer when AI assistance is involved.
Which Peer-Review Writing Tasks Might Benefit From AI?
The suitability of AI assistance depends on the task and the journal's permission. Some activities concern presentation rather than scientific evaluation, making them potentially easier to separate from confidential manuscript content.
| Potential AI Assistance |
Possible Benefit |
Necessary Safeguard |
| Organizing independently written comments |
Improves structure and readability |
Obtain required permission and avoid unauthorized disclosure |
| Improving grammar and clarity |
Makes feedback easier to understand |
Preserve the reviewer's intended meaning and comply with journal policy |
| Suggesting a generic review-report structure |
Provides a framework for presenting comments |
Use no confidential manuscript information without authorization |
| Rephrasing independently established criticisms |
May improve precision and professional tone |
Verify that the criticism has not been strengthened or altered |
| Generating methodological criticisms |
May suggest questions for further consideration |
Requires permitted use, independent verification, and relevant expertise |
| Producing an entire review from a manuscript |
Automates substantial review drafting |
Creates significant confidentiality and accountability concerns and may be prohibited |
These are potential uses, not a list of activities universally authorized by journals. A journal may prohibit sending reviewer comments or manuscript-derived information to external AI services even when the reviewer intends only to improve grammar.
Check the Journal's AI Policy Before Using Any Tool
Journal policies govern whether and how AI may be used during peer review. Reviewers should consult the journal's current reviewer instructions and any AI-specific policy before involving a generative AI system.
The ICMJE's January 2026 recommendations state that reviewers must follow the journal's policy or request permission before using AI to facilitate their review. Reviewers who use AI should disclose that use and ensure the resulting content is appropriate and valid. The official ICMJE guidance on AI use by reviewers sets out these expectations.
These recommendations are influential in biomedical publishing, but they should not be represented as binding rules for every journal or discipline. The journal's actual policy remains the immediate authority for the review assignment.
If the policy is silent or ambiguous, reviewers should ask the handling editor before using AI. Silence should not be interpreted as permission.
Why Manuscript Confidentiality Is the Central Constraint
Peer reviewers receive access to manuscripts because they have accepted a confidential editorial responsibility. Submitted manuscripts may contain unpublished findings, sensitive data, commercially valuable information, or ideas that have not yet entered the public record.
Uploading a manuscript to an external AI service may transmit that material beyond the journal's authorized review environment.
Potential concerns include provider retention practices, access controls, secondary processing, contractual restrictions, and uncertainty about how submitted material is handled.
The ICMJE explicitly warns that manuscripts are privileged communications and that AI use may violate confidentiality when appropriate protections are absent.
Reviewers should therefore establish whether uploading a confidential manuscript to an AI system is permitted before sharing any part of the submission.
Watch Out
Removing the authors' names does not necessarily make an unpublished manuscript safe to upload. Its research questions, methods, datasets, findings, or distinctive descriptions may remain confidential or identifiable. Anonymization is not a substitute for journal authorization.
Can Reviewers Use AI Without Uploading the Manuscript?
Potentially, depending on the journal's policy.
A reviewer might ask a permitted AI tool for a generic report structure or for general guidance on presenting constructive academic criticism, without supplying manuscript content or identifiable details.
For example, requesting a generic framework that separates major methodological concerns from minor editorial suggestions does not necessarily disclose information about a particular submission.
However, a prompt containing the manuscript's research question, distinctive methods, unpublished results, or detailed criticisms may reveal confidential information even without including the complete document.
The reviewer must consider what information is being transmitted, not merely whether a file has been uploaded.
Moreover, some journals may restrict AI assistance irrespective of whether confidential content is included. Reviewers should obtain any required permission before using the tool for the review assignment.
AI Can Improve Wording Without Improving Scientific Judgment
One potentially useful application is refining comments that the reviewer has already developed independently.
For instance, a reviewer may identify that the manuscript's causal interpretation exceeds what its observational design supports. AI could potentially help express that concern in a more concise and constructive manner.
Yet language refinement introduces its own risk. The system might transform a qualified concern into an absolute criticism, introduce a methodological claim the reviewer did not make, or suggest revisions beyond what the evidence warrants.
Reviewers should compare any AI-assisted wording with their original observations and verify that the final comment accurately represents their judgment.
The distinction between improving the wording of reviewer comments and delegating the assessment itself is methodologically important.
Can AI Help Identify Methodological Problems?
Generative AI may suggest methodological questions or point toward general reporting considerations. Its usefulness depends on the task, the quality of the supplied information, and the reviewer's ability to evaluate its suggestions.
For example, a model might suggest checking whether a study's design supports causal claims or whether missing-data handling was adequately described.
These suggestions are not evidence that the manuscript contains those problems.
Reviewers should independently examine the actual methods, results, and supporting material before raising a criticism. An AI-generated concern should never be included simply because it sounds technically sophisticated.
Any use involving confidential manuscript content must first be authorized. The detailed question of whether AI can identify methodological weaknesses during peer review therefore involves both scientific reliability and editorial confidentiality.
Why AI-Generated Criticisms Require Independent Verification
Generative AI can produce authoritative-sounding statements that are incorrect, irrelevant, or insufficiently supported.
A model may criticize a manuscript for omitting a statistical procedure that was actually reported. It may recommend an inappropriate analysis, misunderstand a disciplinary convention, or invent a limitation unrelated to the study design.
Such comments can burden authors with unnecessary revisions and potentially influence editorial decisions unfairly.
The reviewer should be able to identify the evidence supporting every substantive criticism and explain why the requested change is appropriate.
If a criticism cannot be independently understood and defended, it should not appear in the submitted report. This principle applies regardless of how persuasive the generated language sounds.
AI Assistance Should Not Replace the Reviewer's Own Evaluation
Peer review involves judgments that cannot be reduced to the production of plausible comments.
A reviewer must consider whether the methods are suitable for the research question, whether the conclusions follow from the evidence, and whether limitations have been interpreted appropriately.
These judgments require disciplinary understanding and accountability. AI may support particular tasks where permitted, but the reviewer remains responsible for the scientific assessment.
For this reason, the boundary between AI assistance and human judgment should remain explicit.
A reviewer who submits an AI-generated report without independently examining and evaluating the manuscript has not fulfilled the same scholarly responsibility as one who conducts a substantive review.
Should Reviewers Disclose AI Assistance?
Disclosure requirements depend on the journal, but reviewers should not assume that AI assistance can remain undisclosed.
The ICMJE recommends that reviewers disclose AI use to the journal and identify the assistance involved. Its recommendations also require adherence to journal policy or permission before use.
Disclosure may identify the tool and explain whether it was used for organization, language editing, or another authorized purpose.
Importantly, disclosure does not retroactively authorize a prohibited activity. A reviewer who has shared confidential manuscript material without permission cannot resolve the confidentiality issue merely by acknowledging the upload afterward.
Researchers should follow the journal's requirements for disclosing AI assistance in peer review and seek clarification when those requirements are uncertain.
Does the Type of AI System Matter?
Yes, although the journal's authorization remains essential.
Different systems may offer different arrangements for data retention, access control, processing, and institutional oversight. An institutionally managed tool may provide stronger contractual or technical safeguards than a general consumer service.
However, describing a tool as private, secure, enterprise-grade, or locally deployed does not automatically establish that its use is permitted for confidential peer review.
Reviewers should verify the applicable journal policy and any relevant institutional requirements. When the tool's data handling cannot be established, confidential manuscript content should not be supplied.
What Should the Reviewer Verify Before Submitting an AI-Assisted Report?
Before submission, the reviewer should reread the final report against the manuscript and their original notes.
Each substantive criticism should correspond to an identifiable issue in the submitted work. Suggested revisions should be proportionate to the concern, and any statistical or methodological recommendation should be appropriate to the research design.
The reviewer should also check that AI-assisted editing has not introduced invented citations, exaggerated claims, contradictory recommendations, or unnecessarily hostile language.
Finally, the report should comply with the journal's confidentiality and disclosure requirements.
AI assistance may reduce the effort required to organize or express a review, but it does not transfer responsibility for the report to the technology.
06 · What This Means for You
How to Decide Whether AI Assistance Is Appropriate for Your Next Peer Review
Before using AI, determine whether the journal authorizes the intended activity and whether it can be performed without compromising confidentiality.
The fact that a tool can produce polished reviewer comments is not sufficient justification for using it. The relevant questions concern permission, information handling, and the independence of the scholarly assessment.
A practical decision framework
If the journal explicitly prohibits AI assistance
Do not use generative AI for the prohibited review activities, even if the intended purpose is only language editing.
If the journal's policy is silent or unclear
Request clarification or permission from the handling editor before using AI to facilitate the review.
If AI-assisted language editing is permitted
Use only authorized tools and material, preserve the intended meaning, and independently verify the final comments.
If the intended use requires confidential manuscript content
Confirm explicit authorization and appropriate confidentiality protections before supplying the content. Otherwise, do not upload it.
If AI suggests a new methodological criticism
Evaluate the concern independently and include it only if you understand it and can substantiate it from the manuscript and relevant methodological standards.
Keep the Scientific Evaluation Separate From Writing Assistance
A defensible workflow begins with the reviewer reading the manuscript, examining the evidence, and developing an independent assessment.
Any permitted AI assistance should support the presentation of that assessment rather than determine the scientific judgment.
Reviewers should retain control over the substance of their comments, including which concerns are major, which are minor, and what revisions are reasonably necessary.
They should also remember that an editor's decision is not determined solely by the reviewer's recommendation. The review is one contribution to a broader editorial process.
Watch Out
Do not paste an unpublished manuscript, its figures, detailed results, or confidential reviewer notes into a general-purpose AI service without the required authorization and confidentiality safeguards. A tool's usefulness for writing does not override the obligations attached to privileged manuscript access.
The safest role for generative AI in peer-review writing is one that remains explicitly authorized, protects confidential material, and leaves the scientific assessment under the reviewer's independent control.
