Analyzing technical documents with Deep Review
Google Research
Google released Deep Review, an AI tool for checking technical docs. It’s meant to spot gaps and mistakes faster than human reviewers can.
Based on reporting by Google Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google Research has opened up Deep Review, an agentic AI system built to give technical documents the kind of scrutiny that usually takes seasoned experts. It started with computer science papers, but Google says it now reaches into legal and finance too. The pitch is simple: AI can help write and draft faster, but human review has not scaled at the same pace.
That mismatch is getting painful in academia. Google points to this year’s AI conference load: more than 74,000 submissions to ICLR, NeurIPS, and ICML, up from about 45,000 in 2025, 32,000 in 2024, and 23,000 in 2023. Each paper needs multiple serious reviews, and Deep Review is aimed at the grinding parts of that job: checking logic, math, facts, and alignment with related work, while leaving final judgment to people.
The system grew out of Paper Assistant Tool, or PAT, which Google ran in pilots with STOC 2026, ICML 2026, NeurIPS 2026, and ICLR 2027. More than 16,000 manuscripts were reviewed in those private, opt-in programs before submission. Google says those papers were not used for training and were deleted from its systems after each program ended.
Deep Review now comes with two free document reviews per week, plus higher limits through a Gemini API key and an API for programmatic use. It currently accepts PDFs and plain text, with multi-file support coming later. Google also says uploaded documents are not used to train or fine-tune its models, and are stored for one year only so users can view them in the Deep Review interface.
Under the hood, Google says the system breaks a document into segments, assigns compute where it is needed most, runs specialized reasoning agents in parallel, then synthesizes the results and performs grounding checks before returning feedback. And the company is already folding the same idea into Antigravity through a Teamwork Deep Document Review Pattern, using the /teamwork-preview command. This is not being sold as magic. It is being sold as relief for a review bottleneck that has become too obvious to ignore.
My take — AI-written commentary, not fact-checked reporting
This is the sensible AI story: less sparkle, more bookkeeping. The real innovation isn’t that a model can praise your paper; it’s that it can catch the boring inconsistencies people miss when deadlines turn everyone into a speed-reader. The bigger question is whether institutions will use tools like this to raise standards, not just to shovel more drafts through the same broken process.
Read more about this at: Google Research