Google Research has opened Deep Review to the public, an AI system that hunts for mathematical errors, logical gaps and unsupported claims in long technical documents. Google announced the release on October 14. Under the hood it is a team of cooperating agents, not one model reading a document in a single pass.
The verification gap
The pitch is simple. Generative AI now helps write papers, proposals and contracts faster than people can check them. Google’s researchers Rajesh Jayaram and Vincent Cohen-Addad call this a validation bottleneck: information is produced faster than anyone can review it properly.
Academic peer review shows the strain clearly. Google says submissions to three major AI conferences (ICLR, NeurIPS and ICML) passed 74,000 this year, up from roughly 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023.
From Paper Assistant to Deep Review
Deep Review grew out of the Paper Assistant Tool, or PAT, an experiment to let authors stress-test manuscripts before submission. Google ran pilots with four computer science venues: STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027. Authors opted in for private feedback in the weeks before each deadline, kept entirely separate from the official review. Google puts the number of manuscripts at more than 16,000, though its own social post cites over 15,000.
In a test on 26 mathematics and computer science papers known to contain errors, PAT caught 89.7% of the 29 errors. Gemini 3.1 Pro used in the ordinary way caught 55.2%. Among ICML 2026 survey respondents, 35.4% said the tool had flagged significant theoretical problems, and 31% ran new experiments because of its feedback. Some authors found algorithmic flaws they had missed for months and had to rewrite large parts of their papers.
How it works
The technical reality is a pipeline. The document is split into sections such as introduction, methods, results and proofs, and harder material, mathematical proofs especially, gets a bigger share of the compute. Specialist agents then check sections in parallel. Each can see the whole document, so contradictions between chapters can surface. A final agent merges the findings, removes duplicates and cross-checks against outside information to cut false alarms.
It is not limited to academia. Google says it can be pointed at software design documents, slide decks and graphs, legal clauses and financial statements. The tool now supports Gemini 4 Argon, which Google says improves on older PAT results. A related multi-agent review pattern has also gone into its Antigravity development environment.
Limits and access
As ever, the detail that matters is the failure mode. During the pilots PAT sometimes raised problems that did not exist or marked correct proofs as wrong, so Google presents Deep Review as an aid to expert judgment, not a replacement for it.
It is available worldwide. The free tier allows two reviews a week, and heavier users can connect a Gemini API key or use the Deep Review API. At launch it accepts only PDF and plain text, with multi-file review promised later. Google says uploaded documents are not used for model training, though they are stored for a year so users can view them in the interface.
Sources
- Analyzing technical documents with Deep Review
- Google has made 'Deep Review,' an AI-powered tool for verifying errors in academic papers and technical documents, publicly available for free use up to twice a week.
