What one journal calls assistance, a funder calls a confidentiality violation.

That gap exists because most policies governing AI peer review were written for a single behavior, and reviewers are now doing two very different things under one name. 

In one, a reviewer uses a model as an instrument while forming the judgment themselves. In the other, the model forms the judgment and the reviewer submits it. The first is increasingly permitted, within limits. The second is getting reviews struck and papers desk-rejected.

Venues and publishers have started drawing this line in writing, some with enforcement attached. Reviewers who cannot say which side of it their workflow falls on are now carrying real professional risk.

A defensible AI policy starts with a clear boundary between assistance and generation, plus a disclosure standard that makes that boundary visible to an editor.

‍

Why the Distinction Matters

More than half of roughly 1,600 researchers surveyed across 111 countries say they have used AI while peer reviewing a manuscript, often against the guidance of the venues they review for. The number is less alarming than the ambiguity inside it, as the survey covers everything from polishing a sentence to producing an entire report.

Policy treats those behaviors very differently. The NIH prohibits reviewers from using generative AI to analyze applications or formulate critiques, and treats uploading application content to an online AI tool as a confidentiality violation in its own right. Major publishers hold a narrower line, allowing limited help on the reviewer’s own text when disclosed and keeping the manuscript out of public tools.

Both behaviors fall under the same label, and the label hides everything that matters. A reviewer who says they used AI has told an editor almost nothing. Where the use falls on the line determines whether the review is compliant, whether the report can be trusted, and, at some venues, whether it survives at all.

‍

‍

What AI-Assisted Peer Review Means

In AI-assisted peer review, the reviewer reads the manuscript, forms the judgment, and writes the substance of the report. The model operates as an instrument, under the reviewer's direction, using the reviewer's own words and questions. In practice, assistance inside the line looks like this:

• Polishing your own draft: improving the clarity, tone, or structure of comments you have already written.
• Checking your understanding: asking about a statistical method or a field convention, without sharing manuscript content.
• Pressure-testing your own critique: having a model challenge the points you drafted, then deciding which challenges hold.

Used this way, AI measurably improves reviews. In a randomized controlled study of more than 20,000 reviews at ICLR 2025, a machine-learning conference, 27% of reviewers who received automated feedback on their drafted reviews revised them, and blinded evaluation found the revised reviews more informative. The judgment stayed human, but the artifact improved.

‍

What AI-Generated Peer Review Means

In AI-generated peer review, the model reads the manuscript and produces the evaluation. The human contribution shrinks to selecting the output and submitting it. However fluent the result, no expert judgment was formed anywhere in the process.

That difference is visible in the reports themselves. When authors at ICLR 2026 raised concerns about suspiciously long, vague reviews with hallucinated citations, an analysis by the detection firm Pangram Labs, reported in Nature, flagged 21% of roughly 75,800 reviews as fully AI-generated. The flagged reviews shared a signature: exhaustive lists of weaknesses, low information density, and criticisms of methods the papers never used.

Machine-learning venues produced these numbers first because their review systems are public and measurable. Life-science journals face the same pressure without the public record.

‍

The Main Differences Side by Side

Every dimension of the comparison reduces to the same question: where is the judgment formed?

The rows are different symptoms of one variable: autonomy. Move the judgment from the human to the model, and every downstream property of the review changes.

‍

Why AI-Generated Reviews Break Peer Review

Peer review integrity rests on a specific promise. A qualified expert reads the work and stands behind the criticism. Generation removes exactly that promise, and the documented harms all follow from its absence. 

• Accountability collapses: a review nobody authored is a review nobody can defend in discussion with authors or editors. Venues now treat this as misconduct: ICML 2026 desk-rejected 497 papers, about 2% of submissions, tied to reviewers who violated an LLM policy they had explicitly agreed to, with every flagged review manually verified before action.
• Confidentiality breaks: an unpublished manuscript or proposal leaves the review circle the moment it is uploaded to a public tool, whatever happens to the output afterward. The NIH classifies the upload itself as the violation.
• Scores drift: undisclosed reliance shifts outcomes, even short of full generation. At one machine-learning venue, reviews flagged as AI-assisted assigned higher scores and raised paper acceptance probability by 3.1%, a lottery effect that rewards authors for the reviewer they happened to draw.
• Trust erodes: authors who suspect a generated review stop believing the process evaluated their work at all, and every future critical review inherits that suspicion.

‍

How to Use AI in Review Without Crossing the Line

The line is workable in daily practice. It asks reviewers to keep the judgment and editors to make the expectations explicit.

‍

For Reviewers

Read the venue's written policy before touching a tool. Even a single conference can hold two lines at once. ICML 2026 ran one policy banning all LLM use and one permitting limited help, and reviewers committed to one in advance. 

Never upload manuscript or proposal content to a public tool. Work only on text you drafted yourself, and disclose what you did. Springer Nature, for example, asks reviewers to declare any AI support of the evaluation in the review report itself.

‍

For Editors

State the policy in the review invitation rather than the aftermath, and ask for disclosure instead of relying on detection. A disclosed, narrow use of AI in a reviewer's own text is a different event from silence, and treating them the same pushes honest reviewers to say nothing.

The same principle shapes assistive tools built for this. QED Science's Paper Review works on the author's side. The platform breaks a manuscript into individual claims, evaluates each against a validated claim-level quality metric, and returns structured findings for the author to inspect and decide on before submission. The judgment stays with the scientist, and the confidentiality at stake is the author's own to give.

‍

Strengthen the Manuscript Before Anyone Reviews It

Every rule above protects someone else's unpublished work. An author's own manuscript carries no such restriction, and the pre-submission window is the one place where claim-level AI rigor is entirely the author's call.

That window is where the blind, anonymized evaluation infrastructure behind QED Score does its most practical work, surfacing the gaps a reviewer would catch, while there is still time to fix them.

Upload a manuscript and put the toughest questions to your work before a reviewer does.

‍

FAQs