Google Research has opened Deep Review to the public, an AI system built to check complex documents for logical gaps, shaky math and unsupported claims. It began as the Paper Assistant Tool, or PAT, an experiment to help computer science researchers stress-test manuscripts before submitting them. It now covers other fields too, including legal and finance work, and can also review Google Docs and presentations. Authors can use it to improve a draft, and reviewers can use it to verify claims.
Tested on 16,000 papers
Over the past year Google ran private pilots with four flagship conferences: STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027. Authors who opted in received feedback on their papers in the weeks before each deadline, kept apart from the official review process. According to Google, PAT collected no data for training, and papers were deleted from its systems soon after each program ended. More than 16,000 manuscripts went through. Author reactions were “extremely positive,” the company said, and the tool has improved since with the addition of Gemini 4 Argon support. Details of the pilots appear in a manuscript titled Towards Automating Scientific Review with Google’s Paper Assistant Tool.
The review bottleneck
The researchers behind it, Rajesh Jayaram and Vincent Cohen-Addad, describe a growing mismatch: AI speeds up writing, but the human capacity to check that work has stayed roughly fixed. Reviewing a long manuscript or a complicated legal agreement takes specialist knowledge. Academic peer review shows the strain most clearly. This year ICLR, NeurIPS and ICML together drew more than 74,000 submissions, up from about 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023, and each needs several careful reviews.
Deep Review checks logical consistency and mathematical statements, grounds facts, compares a document with existing work and suggests improvements. It leaves higher-level judgments to people. Google says a single prompt to a language model often fails on long documents, because it runs out of reasoning budget, misses links between sections or accepts subtle claims unchallenged, so Deep Review runs a four-stage agentic pipeline instead.
Who could benefit
Strong pre-submission critique has usually depended on senior faculty, well-funded labs or large networks of peers. For researchers without those, an automated check on their reasoning could make a real difference. Google is also introducing a related Teamwork Deep Document Review Pattern for Antigravity, invoked with the /teamwork-preview command. Multi-file support is coming soon, which will let users run consistency checks across whole sets of documents, spreadsheets and exhibits at once.
