The Stanford Agentic Reviewer, at paperreview.ai, is the free AI reviewer built by Yixing Jiang and Andrew Ng at the Stanford ML Group, and it sits at the top of the results for “AI peer review”. It has earned that place: it costs nothing, its method is published, and its authors measured it against human reviewers. It was also built for one particular world, machine-learning and computer-science conferences, and that shapes what it can and cannot do with a paper in psychology, nursing or medicine.
This page compares it with Previum, which is mine, so read it as written by an interested party. Everything said here about paperreview.ai comes from its own homepage and technical overview, read on 6 October 2026; where I could not confirm something, I say so instead of guessing.
The short version: if you are submitting to ICLR, NeurIPS or another ML venue, use paperreview.ai, it is free and made for you. If your paper goes to a health or social science journal, read on.
It is free, with no catch: no account wall, no paid tier pushing you along. Its authors published how it works and measured it: on 147 ICLR 2025 papers held out from training, the correlation between its rating and the human reviewers’ is 0.42, while two human reviewers agree at 0.41. That is a serious result, and it comes from one of the best-known machine-learning groups in the world.
It also does something mine does not: it searches the literature. It generates queries, pulls related papers from arXiv, summarises them and uses them to judge whether your contribution is new and well positioned. In machine learning, where nearly everything appears on arXiv first, that is exactly the context a reviewer brings. Previum checks the references you cite, but it does not go looking for the paper you should have cited.
paperreview.ai · Its technical overview, with the evaluation
paperreview.ai analyses the first 15 pages of the PDF. For a conference paper, eight or nine pages plus references, that covers almost everything. For a journal article in health or social sciences, double-spaced as most journals ask at submission, the Method often starts well into the document and the Results and Discussion come later still.
That is where the objections that sink a manuscript live: sample size, measurement, analysis, what the design can support. A review of the first fifteen pages of a journal manuscript can read as complete while the part that gets you rejected was never read. If you use it on a long manuscript, check where page fifteen falls in yours.
Its own overview says it works best in fields like AI, where recent research is freely available on arXiv. In clinical psychology, nursing, physiotherapy or medicine, most of the literature you would be compared against is published in journals and indexed in PubMed, and preprints are less common. The related-work search that makes it sharp in machine learning has less to stand on there.
Its scoring is also calibrated on ICLR reviews: the regression that turns its seven dimension scores into a rating was trained on 150 ICLR submissions. That is the right calibration for a conference reviewer and a different one from what a health sciences journal editor weighs.
It reads the whole manuscript, Word or PDF. The free Basic report gives you the finding that decides the review, with the sentence it objects to and the fix, plus three checks run with code, not with a model: your references against eight databases (Crossref, OpenAlex, PubMed, Semantic Scholar, DOAJ, OpenLibrary, Google Books and arXiv), your p values recomputed from their degrees of freedom, and figures that contradict each other.
The Professional report is the full simulated review: two reviewers, one from your field and one on methods, and an editor who signs a decision, every objection anchored to its sentence, your Word file back with the comments in the margin, and a second round within 30 days to cross the new version against the objections of the first. The manuscript is not stored; your account keeps the review.
Our benchmark against Reviewer3, step by step · A sample report on a published nursing paper (CC BY)
If your paper is going to an ML or CS conference, use paperreview.ai: it is free, it is built for that, and its arXiv search is an advantage I cannot match. If it is going to a journal in health or social sciences, run both free layers on the same manuscript and compare where each one lands: on method and analysis, with a sentence of yours quoted, or on generalities. That comparison takes ten minutes and tells you more than any table, this one included.
Yes. On 6 October 2026 its homepage offered free AI feedback on a PDF of up to 10 MB, analysing the first 15 pages, with an email notification when the review is ready.
Yixing Jiang and Andrew Ng, at the Stanford ML Group. Its technical overview describes the pipeline: the PDF converted to text, a literature search on arXiv, a synthesis of related work, and a review scored on seven dimensions.
It accepts an “Other” venue, but its literature search runs on arXiv and its rating is calibrated on ICLR reviews; its own overview says it works best where recent research is freely on arXiv. For a long journal manuscript, also check that your Method and Results fall within the first 15 pages it reads.
On 147 ICLR 2025 test papers, the Spearman correlation between its ratings and human ratings was 0.42, against 0.41 between two human reviewers, according to its technical overview. That measures agreement on overall scores, not whether each comment is right.
Its authors discourage using it as a conference reviewer where that violates the conference’s policies. A manuscript you were sent to review is confidential: check the venue’s rules before uploading it anywhere.
Previum reads the whole manuscript, checks your references against eight databases and recomputes your p values with code in the free Basic report, and in the Professional report returns two simulated reviewers and an editor, with your Word file commented in the margin.