Google Deep Review Goes Public: Two Free Document Reviews a Week, Built From 16,000 Conference Papers

Google Research has opened Deep Review, its agentic document checker, to the public. In a launch post dated October 14, 2026, research scientists Rajesh Jayaram and Vincent Cohen-Addad say it is accessible worldwide, gives every user two free document reviews per week, and can be driven through an API with your own Gemini key. It started as a tool for catching errors in computer science papers. It now also covers design documents, presentations, and legal and financial documents.

The reason it matters beyond academia: AI has made documents cheap to produce, and checking them is still slow, expensive human work. Deep Review is Google's bet that verification can be automated well enough to be useful before a human ever looks.

Deep Review at a glance

What problem is Google solving?

Google calls it the validation bottleneck: information is being generated faster than it can be rigorously reviewed. The post's own number is the clearest illustration. This year there were over 74,000 submissions to three major AI conferences (ICLR, NeurIPS and ICML), up from roughly 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023. Each submission needs several careful reviews.

We have covered the same strain from other angles: arXiv capping submissions at two per month to slow the flood of AI-assisted papers, an earlier arXiv one-year ban for AI-generated errors, and research on LLM judges refereeing papers. Deep Review sits on the other side of that fight. Instead of limiting supply, it tries to add reviewing capacity.

Two reviewers inspecting a single research paper, one holding an approval stamp, illustrating Google Deep Review AI peer review

How the four-stage pipeline works

Google's argument is that a single prompt on a long manuscript fails in predictable ways: it runs out of reasoning budget, overlooks dependencies between distant sections, or accepts nuanced claims without scrutiny. Deep Review replaces the single call with a pipeline of agents.

1. Document segmentation. A segmenter agent decomposes the document into logical segments (introduction, methodology, experiments and so on). Segments may overlap but must cover all of the content.
2. Adaptive budgeting. The pipeline inspects information density and allocates compute per section. A dense mathematical proof gets more inference-scaling effort than a related-work section.
3. Deep parallel verification. Specialized reasoning agents examine each segment in parallel. Each can see the full document so it can follow cross-references, but concentrates on its own target section.
4. Global synthesis and agentic grounding. A synthesis agent merges the segment reports, runs consistency checks between sections (re-running logic checks if needed), and performs external grounding checks and tool calls as a last guard against hallucinated findings.

If you build agents, the shape will look familiar: fan out, verify in parallel, then fan in with a skeptic. It is the same pattern behind subagent delegation in coding agents, applied to reading instead of writing.

What the evidence actually shows

Google's blog post reports author reactions to the pilots as "extremely positive" and shows pilot statistics in a chart, but the clearest published numbers sit in the preprint, Towards Automating Scientific Review with Google's Paper Assistant Tool (arXiv 2606.28277, submitted June 26, 2026). Its abstract says PAT, using inference scaling, achieves a 34% improvement over zero-shot recall on mathematical errors in the SPOT benchmark, and that pilots at STOC and ICML surfaced critical errors and substantive improvements.

Some outlets have repeated a more specific detection-rate figure for Deep Review. We could not find that number in the arXiv abstract or in Google's launch post, so treat it as unconfirmed until you see it in the paper's full text.

Three caveats are worth holding on to:

• The pilots were opt-in. Authors who chose to use PAT before a deadline are not a random sample, and satisfaction among them is not the same as measured accuracy.
• The tool in the pilots is older than the tool launched. Google says Gemini 4 Argon support makes Deep Review stronger than the PAT versions described in the paper, so the paper's numbers describe a predecessor. For the model itself, see our Gemini 4 Argon launch breakdown.
• Everything here is Google-reported. There is no independent benchmark of the public product yet.

Beyond papers: design docs, slides, contracts

The most interesting part of the announcement is the generalization. The reasoning used to test whether a proof holds up, Google says, also surfaces logical gaps in any document where claims must be backed by evidence. The post lists concrete uses:

• Design documents: catch unstated architectural assumptions, spot edge cases, suggest improvements.
• Presentations: check that slide transitions flow and that bulleted takeaways match the charts.
• Legal and financial documents: check internal consistency across clauses and data tables, flag ambiguous terms and downstream consequences, and ground claims against supporting evidence.

For a builder, the design-doc case is probably the most practical. A spec that contradicts itself across sections is a common source of rework, and a review pass before sending it to teammates is cheap. For legal and finance material, treat the output as a checklist of things to ask a professional, not as advice.

Deep Review meets Antigravity

Google also introduced the Teamwork Deep Document Review Pattern, an implementation inspired by Deep Review built into Antigravity on the Teamwork multi-agent orchestration framework. It uses a multi-agent hierarchy with self-correction and segment-based rubrics, and Antigravity users invoke it with the /teamwork-preview command. If you already work in that environment, you can approximate the pipeline on local files without uploading them to the Deep Review service. Our write-up of Antigravity Teamwork explains the framework.

How to try it

1. Open the Deep Review entry on the Google Research portal linked from the launch post and upload a PDF or text file.
2. Use your two weekly reviews on documents where an error would be costly: a paper before submission, a design doc before a review meeting.
3. Read the report as a list of leads. Verify each flagged issue yourself before changing anything.
4. For volume, connect a Gemini API key and use the Deep Review API, for example in a drafting or continuous integration workflow.

A sensible prompt discipline when you read the output: for every finding, ask "can I reproduce this by reading the cited section?" If you cannot, treat it as a possible false alarm.

What to watch before relying on it

Privacy. Google says uploads are never used for model training and are kept for one year to show them in the interface. That is a clear statement, but a year of retention may still rule out confidential contracts or unpublished results at some organizations. Check your own policy before uploading.

False positives and false negatives. An automated reviewer can flag non-errors and miss real ones. Google's own framing, that high-level judgement stays with people, is the right way to use it.

Reviewer-side use. The post says the tool can serve both authors improving a draft and reviewers verifying claims. That second use is where governance questions arise: venues have to decide whether and how reviewers may use such tools. The pilots were decoupled from official conference reviewing, and Google does not announce any change to that here.

A narrow input list. PDFs and plain text only for now. Spreadsheets, exhibits and multi-document sets are promised, not shipped.

Where it fits in Google's science push

Deep Review is one of several research tools Google has been rolling out for scientists, alongside its AI co-scientist work shown at ICML and the broader trend of AI producing results that need checking, as in AlphaProof Nexus and peer-reviewed Erdos problems. The common thread is that generating claims is getting cheaper than verifying them, so verification tooling is becoming the next product category. If you read benchmark claims from vendors regularly, our guide to how to read AI benchmarks applies equally to Google's pilot statistics.

Bottom line

Deep Review is a useful, low-cost second pair of eyes: two free reviews a week, a clear pipeline, a stated privacy policy and an API. It is not an oracle, and its accuracy claims come from Google itself. Use it as a pre-flight check on documents that matter, keep a human in the loop, and wait for independent evaluations before treating its silence as proof that a document is correct.

Details reflect Google Research's launch post of October 14, 2026 and may change as the product evolves.

Related reading

• Gemini 4 Argon launch: benchmarks and pricing
• Antigravity Teamwork multi-agent framework
• arXiv two-submissions-per-month cap
• arXiv one-year ban for AI-generated errors
• LLM judges as paper referees
• Google AI scientist at ICML 2026
• AlphaProof Nexus and peer-reviewed Erdos problems
• How to read AI benchmarks