Google has opened Deep Review to the public, offering an AI agent that checks complex documents for logical gaps, unsupported claims and inconsistencies. Everyone gets two free reviews a week, bringing a tool first tested on research papers into reach for legal, financial and everyday technical work.
Google DeepMind Watch analysis
What happened
Google Research says Deep Review grew out of its Paper Assistant Tool, used in opt-in pilots with more than 16,000 submissions across four computer science conferences. The public version accepts PDFs and plain text, with multi-file reviews still to come. Users with heavier needs can connect a Gemini API key for higher limits and programmatic access. Read Google Research’s announcement.
The system divides a document into sections, assigns review effort according to each section’s complexity, then runs specialised agents in parallel. A synthesis agent brings their findings together, checks consistency across the document and carries out external grounding checks. Google says it can examine mathematical statements, logical consistency, factual grounding and comparisons with existing work.
Why it matters
Deep Review is designed to make a particular kind of expert attention less scarce: the slow work of checking whether a long document’s claims hold together. Researchers could use it to stress-test a paper; teams could look for assumptions in a design document or mismatches between presentation slides and charts. It offers two free reviews each week, rather than asking every curious user to arrive with an API key in hand.
There is a meaningful boundary here. Google positions the tool as a way to surface issues, not to take over the higher-level judgement that makes a review useful. And the generous-sounding free tier comes with a data detail worth knowing: uploads are stored for one year for display in the Deep Review interface, though Google says they are not used to train, fine-tune or improve its models.
Our read
This is a practical launch with a clearer job than “AI that reads things”: help people find claims and connections that deserve another look. Two free reviews a week make it easy to try on a non-sensitive document. Given the stated one-year storage period, think twice before uploading material you would not want retained in the service.
What to watch
- How well the tool catches real errors in documents outside academic research.
- Whether multi-file support arrives, and how it handles evidence spread across documents.
- What the API limits and real-world review quality look like for frequent users.
Discussion spark: For a tool that checks complex documents, would you trust it first with research papers, business documents or something else, and what would it need to show before you relied on its findings?
Sources and evidence
- Analyzing technical documents with Deep Review – Google Research (14 October 2026, 16:17 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind