Blog
Google opens Deep Review, an AI agent that checks technical documents for errors
On 14 October 2026 Google Research publicly released Deep Review, an agentic system that checks long technical documents for mathematical errors, logical gaps and unsupported claims. Every user gets two free reviews a week, with higher limits and an API through a Gemini API key. Accuracy claims come from Google's own testing.
On 14 October 2026 Google Research made Deep Review publicly available. It is an agentic AI system that checks long technical documents for mathematical errors, logical gaps and unsupported claims. Google announced it at its Research@ event in Mountain View, in a blog post by research scientists Rajesh Jayaram and Vincent Cohen-Addad.
Deep Review started life as the Paper Assistant Tool (PAT). Over the past year, Google offered PAT as private, opt-in, pre-submission feedback to authors at four large computer science conferences: STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027. Google says this feedback was kept separate from official review, and papers were not used for training.
The problem Google says it is solving
Google argues that AI has made writing faster than checking. Its post says ICLR, NeurIPS and ICML received more than 74,000 submissions this year, up from about 23,000 in 2023. Speaking at Research@, Jayaram said AI can now generate drafts of papers and legal documents "nearly at the push of a button", while human capacity to verify them has stayed essentially fixed, R&D World reported.
How it works
According to Google, Deep Review runs a four-stage pipeline instead of a single prompt:
- Segmentation: an agent splits the document into logical sections that together cover all of the content.
- Adaptive budgeting: sections get different amounts of compute depending on how dense they are. A section of mathematical proofs, for example, gets more than an introduction.
- Parallel verification: separate reasoning agents check each section at the same time. Each can see the whole document so it can follow cross-references.
- Synthesis and grounding: a final agent merges the reports, re-runs consistency checks between sections, and uses external checks and tool calls to guard against hallucinations.
Google says the tool now uses Gemini 4 Argon. Jayaram told R&D World that a review takes about 10 to 20 minutes and returns major and minor issues, suggestions and summaries, with comments anchored to the relevant text.
Access, limits and privacy
Google lists these terms:
- Every user gets two free document reviews per week.
- Heavier users can connect a Gemini API key for higher limits and use the Deep Review API in their own drafting or CI workflows.
- Accepted inputs include PDFs and plain text. Google says multi-file input, such as checking related documents and spreadsheets together, is coming soon.
- Google says uploaded documents are never used to train or improve its models. They are stored for one year so they can be shown to the user in the Deep Review interface.
Antigravity users also get a multi-agent review pattern inspired by Deep Review, started with the /teamwork-preview command.
Beyond research papers
Google says the same checks apply to other documents where claims need evidence: design documents (unstated assumptions and edge cases), presentations (whether slide takeaways match the charts), and legal and financial documents (consistency between clauses and data tables, and ambiguous terms).
How accurate is it?
R&D World reported that a June preprint by Google researchers found PAT detected 89.7% of known mathematical and proof errors in a selected benchmark subset, against 55.2% for a single Gemini 3.1 Pro response. That test covered only 29 errors across 26 papers. Pilot participants also reported cases where the system wrongly challenged correct proofs or arguments. Google itself describes Deep Review as handling the most tiring parts of review while "leaving higher-level judgements to humans".
Why it matters
For businesses, the practical use is likely a second reader for contracts, financial reports, specs and pitch decks before they go out. Two free reviews a week is enough to test it. The usual caution applies: a tool that can wrongly flag a correct proof can also miss a real problem. Its findings are a checklist for a human expert, not a replacement for one. Companies should also weigh the one-year storage period before uploading confidential material.