Google Research has opened Deep Review to the public, giving anyone a free, agent-powered second reader for technical and professional documents. Launched on October 14, 2026, the tool checks manuscripts, design documents, presentations and legal or financial files for logical inconsistencies, mathematical errors and unsupported claims. Here is what it does, how to use it well, and where its limits lie.
What Deep Review Is
Deep Review grew out of Google's Paper Assistant Tool (PAT), which was piloted with major computer science conferences including STOC, ICML, NeurIPS and ICLR. Google says those pilots covered more than 15,000 opt-in manuscripts, with feedback delivered privately before submission and kept separate from official peer review.
The public version broadens the scope beyond academic papers. Google says it now suits:
- Design documents, where it looks for unstated assumptions and missed edge cases
- Presentations, checking slide flow and whether takeaways match the charts
- Legal and financial documents, flagging inconsistencies between clauses and tables, ambiguous terms and claims that need grounding
Google is clear that higher-level judgments, such as whether a research idea is important or a deal is wise, remain with humans.
How It Works Under the Hood
Deep Review is an agentic system rather than a single chatbot response. According to Google's research blog, it runs in four stages:
- A segmenter agent splits the document into overlapping logical sections so nothing is missed
- Adaptive budgeting assigns more compute to dense material, such as mathematical proofs, than to introductions or related work
- Parallel verification agents review each section while retaining access to the full document for cross-references
- A synthesis agent consolidates the findings, checks consistency across sections and runs external grounding checks, including searches to catch possible hallucinations
R&D World reports that the free tier runs on Gemini 4 Argon at medium compute.
Step-by-Step: Running Your First Review
Getting started takes a few minutes:
- Go to research.google/deepreview and sign in with your Google account.
- Upload your document. Google's blog lists PDF and plain text at launch, with multi-file inputs coming soon. Some coverage, including R&D World, lists additional formats such as Word, PowerPoint and Markdown, so check the upload screen for what is currently supported.
- Choose an effort level if offered. R&D World reports users can set how much effort the model spends on the review.
- Wait for results. A review typically takes 10 to 20 minutes. Completed reviews appear in your documents list.
- Read the anchored comments. The output separates major and minor issues, offers suggestions and summaries, and pins comments directly to the relevant passages.
Limits, Pricing and the API
Every user gets two free reviews per week. Heavier users can connect a Gemini API key for higher limits, and a Deep Review API lets teams run checks programmatically, for example inside a continuous integration pipeline for technical documentation or a drafting workflow.
On privacy, Google says uploaded documents are never used to train, fine-tune or improve its models. Files are retained for one year solely so you can view them in the Deep Review interface. That said, organisations handling confidential contracts or regulated data should still check internal policy before uploading.
Tips for Getting Better Results
- Use it before, not after, sharing a draft. Deep Review is most valuable as a pre-submission or pre-circulation check, when fixes are cheap.
- Spend your two free reviews wisely. Run it on near-final drafts where structure and numbers are stable, rather than early outlines.
- Treat every flag as a lead, not a verdict. Verify each issue against the source before editing.
- Pair it with domain review. Use Deep Review for consistency and logic, and keep human experts for strategy, novelty and judgment.
How Accurate Is It?
Google has not published accuracy figures for Deep Review itself. The best available evidence comes from a June preprint on its predecessor: PAT caught 89.7% of known mathematical and proof errors in a benchmark subset, compared with 55.2% for a single Gemini 3.1 Pro response. That test was small, covering 29 errors across 26 papers.
Pilot participants also reported cases where the system wrongly challenged correct proofs or arguments. False positives can waste time or, worse, push an author to "fix" something that was right.
Why It Matters
As AI-generated text multiplies, the scarce resource in many organisations is no longer drafting but verification. Deep Review targets that bottleneck directly, giving researchers, engineers, analysts and lawyers a structured, multi-agent check that goes deeper than asking a chatbot to proofread. Its free tier makes it easy to evaluate, and its API hints at a future where automated document review becomes a standard step in professional workflows.
