Google opens Deep Review to the public with two free document checks a week

Google released Deep Review publicly on October 14, 2026, giving users two free document reviews a week. Built from an academic manuscript feedback tool, it now targets complex documents beyond research papers. Google says the system checks reasoning, mathematical statements and supporting evidence, while leaving higher-level judgments to people.

An experiment built around the review backlog

The launch accompanied Google's Research@ 2026 event in Mountain View. Its stated catalyst is a mismatch: AI can accelerate writing, experiments and proposals, but human capacity to verify the resulting work has not kept pace. Google describes that gap as a validation bottleneck.

Academic publishing supplied the first testing ground. Google says submissions to ICLR, NeurIPS and ICML exceeded 74,000 this year, compared with approximately 45,000 in 2025. Each paper requires multiple reviews, increasing the burden on experts asked to assess its quality.

Deep Review grew out of Google's experimental Paper Assistant Tool, or PAT. Authors opted in to receive private feedback before submission deadlines through partnerships with STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027. That feedback was separate from the conferences' official review processes—not an automated decision about acceptance.

Google's detailed launch post puts participation at more than 16,000 manuscripts. Its Research@ overview gives a lower threshold of more than 15,000 submissions. Both figures are company-supplied deployment counts, rather than measures of review accuracy.

Dividing the document without losing its context

Rather than asking one model prompt to assess an entire manuscript, Deep Review divides the work into four stages. Google says a single pass can exhaust a model's reasoning budget, miss connections between distant sections or accept subtle claims without enough scrutiny.

First, an AI agent splits the document into logical sections that collectively cover all its content. The system then allocates computing effort according to each section's demands. A dense mathematical proof, for example, can receive more attention than an introduction or a discussion of related work.

Specialized agents examine those sections in parallel, retaining access to the full document so they can follow cross-references. A final agent combines their findings, checks consistency between sections and reruns logical checks when needed. It also uses external checks and tools to help guard against invented information.

Public access reaches beyond research papers

The service is available worldwide and initially accepts PDFs and plain text. For heavier workloads, users can connect a Gemini API key to obtain higher usage limits and integrate reviews into software through the Deep Review API. Google's launch post describes access through its Research Portal.

Google describes several uses beyond academic critique. These are intended applications of the review system, not guarantees that every error will be caught:

• Design documents: surface unstated architectural assumptions, edge cases and possible improvements.
• Presentations: examine whether slide transitions make sense and written takeaways match the charts.
• Legal and financial documents: check consistency across clauses and tables, flag ambiguity and compare claims with supporting evidence.

Multi-file input remains a promised next step, not a launch feature. Google says that addition will allow consistency checks across related documents, spreadsheets and exhibits. For now, the public release offers document-by-document critique, with the broader review of a connected set of materials still ahead.