Reviewer3: AI Peer Review for Research Papers

Which Reviewer3 checks are available to authors and reviewers, and to editors and publishers.
Check	Authors	Editors
Submission Screening Catches coursework, partial drafts, leftover AI prompts, and rewrites of already-published work.	Not included	Included
Reference Check Finds hallucinated, retracted, irrelevant, and self-cited references.	Included	Included
Fatal Flaws Surfaces unsupported conclusions, impossible numbers, and circular reasoning.	Included	Included
Claim Review Checks the evidence behind each claim, plus the methods, statistics, derivations, and internal consistency.	Included	Included
Code Replication Runs the code and compares its output with the paper.	Included	Included

Authors

Editors

Submission Screening

Catches coursework, partial drafts, leftover AI prompts, and rewrites of already-published work.

Reference Check

Finds hallucinated, retracted, irrelevant, and self-cited references.

Fatal Flaws

Surfaces unsupported conclusions, impossible numbers, and circular reasoning.

Claim Review

Checks the evidence behind each claim, plus the methods, statistics, derivations, and internal consistency.

Code Replication

Runs the code and compares its output with the paper.

AI Peer Review Checks What's Checkable

An AI reviewer reads a paper the way a referee would. It breaks the paper into the claims it makes, then checks the evidence offered for each one. Does the analysis support the conclusion? Do the claims agree with each other? Are the cited sources real? It won't tell you if the work matters. It catches the problems a human reviewer shouldn't have to spend their time on.

We Verify. We Don't Generate.

Most AI writing tools produce text. Reviewer3 checks text that already exists, and that difference decides what it's good for. Ask a model to “review this paper” and it'll hand you something review-shaped whether or not it found anything, because producing plausible text is the job it's doing. Verification means actually going and looking. Searching for the cited source. Checking claims against the data reported. Saying nothing is wrong when nothing is wrong.

What It Can't Do

It can't tell you whether a contribution is novel, or whether a field will care. That takes knowing what a community values, and it's still a human question. What it's good at is the work where nothing can be sampled. Every reference verified instead of a few. Every claim checked against every other.

Claims First, Then the Evidence for Each

A manuscript is decomposed into the claims it makes, and every claim is checked against the evidence offered for it. Whether the analysis supports the conclusion. Whether the statistics hold. Whether the claims contradict each other or the data reported. Separately, a reference chain pulls out every citation, goes looking for a real source, and decides whether the match is genuine. Stages run in parallel where they can and wait where one needs another's result, and an editor stage pulls all of it into one report, ordered by what matters most.

Who Uses It

Researchers run it on their own work before they submit, to see what a reviewer is likely to raise while there's still time to do something about it. Journals and editorial teams run it on incoming submissions, so a screening decision comes with the methodology and citation checks already done. Same pipeline, different modes.

How do you know it works?

Because we published the tests. The reference checker is measured against a dataset of real and fabricated citations where the answer is known in advance, and ReviewBench compares over 145,000 human and AI review comments. Both are on the benchmarks page, with the datasets and methods, so the claims can be checked rather than taken on trust.

Why does peer review need help at all?

Because the volume has outrun the reviewers. Published articles have roughly doubled since 2010, while 20% of scientists carry up to 94% of the reviewing and fewer than half of review invitations are accepted. That arithmetic does not resolve itself. Our write-up on AI peer review sets out the numbers behind the strain, why we were skeptical that AI could help, and how the review stages were built and measured.

Catch critical issues in research before publication

Technical checks to support human judgment

Evidence from 145,000+ review comments

What researchers and editors are saying

Built for how you work

For authors

For reviewers

For editors

For institutions

Ready to strengthen your research?