Google Launches Deep Review: An AI That Grades Other AI's Homework
Google launched Deep Review, an AI that checks technical documents for math errors and logic gaps. But humans must verify the verifier.
Tools · Source: Google Research Blog
What happened
Google Research just released Deep Review. It is an agentic AI system built to catch mathematical errors, logical gaps, and unsupported claims in technical documents. The tool launched publicly on October 14 at the Research@ event in Mountain View, California. It grew out of a pilot program called the Paper Assistant Tool. That early version gave authors private feedback before submission deadlines at major computer science conferences.
The system works by breaking a document into sections. It throws more computing power at the hardest parts. Multiple AI agents scan the text while a synthesis agent combines their notes. It even uses Google Search to check for AI hallucinations. A full review takes ten to twenty minutes. The feedback anchors directly to the text, highlighting major issues, minor issues, and suggestions.
Google tested this extensively. The pilots provided feedback on more than fifteen thousand submissions at top conferences like NeurIPS, ICLR, and ICML. In one benchmark covering twenty nine errors across twenty six papers, the system caught nearly ninety percent of known math and proof errors. A standard Gemini 3.1 Pro prompt only caught fifty five percent. But the system is not perfect. It sometimes flags correct proofs as wrong, meaning humans still have to grade the grader.
Key facts
- 15,000+ — Manuscripts reviewed during the pilot phase
- 89.7% — Known math and proof errors detected by the Paper Assistant Tool
- 55.2% — Errors detected by a single Gemini 3.1 Pro response
- 10 to 20 minutes — Time required to complete a full document review
Why it matters
AI makes it incredibly easy to generate polished drafts at the push of a button. As UC Berkeley computer scientist Nika Haghtalab noted, AI produces papers that look polished but have weak substance. Meanwhile, human capacity to verify those drafts has stayed completely fixed. Deep Review closes this exact verification gap for builders and researchers. You can now run complex PDFs, Word documents, PowerPoint presentations, or plain text through an automated peer review before you publish. It saves hours of manual fact checking and protects your reputation.
The broader shift is the rise of agentic verification. We are moving past single prompt outputs. Deep Review uses multiple agents working together and searching the web to validate claims. This means the bottleneck in AI is no longer generation. The new bottleneck is verification. As AI generates more weak substance disguised as polished text, tools that grade AI homework will become mandatory infrastructure. We are entering an era where AI checkers are just as important as AI writers.
For builders
Automate your technical document reviews
Google gives you two free reviews a week using 4 Argon on Medium compute. You can unlock higher limits by connecting your own Gemini API key. This is a cheap, scalable way to sanity check your whitepapers or technical reports before investors see them.
Keep your proprietary data safe
Google explicitly states that uploaded documents are excluded from model training. They retain files for one year just so you can access them in the interface. Founders can safely upload sensitive technical specs without feeding the global AI brain or leaking proprietary data.
Do not blindly trust the output
Pilot users reported the AI sometimes incorrectly challenged valid arguments. You still pay the cost of human oversight. The tool highlights exactly where to look, but you still need an expert to make the final call on whether the AI is right or hallucinating.
My take
We built AI to write our papers, and now we need AI to read them. Deep Review proves that generation is cheap but verification is expensive. The future of AI is not just creating content. It is building the immune system to filter out the garbage.
Original reporting: Google Research Blog. This is my rewrite and opinion.