Key Notes
- Google Deep Review offers two free document reviews per week, with additional usage billed through a connected Gemini API key.
- The reported 89.7% detection result came from an earlier system tested on 29 known mathematical and computer-science errors, rather than a general accuracy test.
- Uploaded documents can remain available for up to one year, and users must still verify the system’s criticisms before treating them as established errors.
Google has opened Deep Review to the public, giving researchers and other knowledge workers an AI tool for finding weaknesses in complex documents. The service offers two free reviews per week and is designed to question arguments, check calculations and identify inconsistencies that a quick read might miss.
The October 14 announcement expands an experimental research-paper assistant into a broader document-review service. PDFs and plain text are supported at launch, with multi-file input planned. Google describes uses ranging from scientific manuscripts and design proposals to legal and financial documents.
Deep Review divides a document into sections, allocates more computing effort to demanding material and assigns specialized agents to examine those sections. A synthesis stage brings their findings together and checks consistency across the document. Google says the public release adds support for Gemini 4 Argon, extending the model’s use beyond general-purpose chat and coding.
What the 89.7% Error-Detection Result Measures
The strongest headline number needs context. Google’s research paper reports 89.7% detection on a selected portion of the SPOT benchmark: 26 mathematics and computer-science papers containing 29 known equation or proof errors. That evaluation used the earlier Paper Assistant Tool, or PAT, with Gemini 3.1 Pro, rather than measuring the current Argon-powered service across all document types.
A single model call found 55.2% of those errors under the researchers’ evaluation procedure. The team used an AI grader to assess whether a report identified the known error, then had a human author audit every grade. It also changed the grading method, so the figures are not directly comparable with the original SPOT results.
The percentage measures how many known problems were detected. It does not show that 89.7% of every criticism the system makes is correct. The paper acknowledges false alarms, including cases where the model incorrectly judged a valid proof or argument to be wrong.
That distinction matters as AI produces more material for specialists to assess. Our coverage of OpenAI’s mathematics papers examined the gap between generating a proposed result and establishing confidence in it. An automated critique adds another piece of evidence to that process, rather than ending it.
Conference Pilots Tested Feedback, Not Automated Acceptance
The earlier ICML pilot provided feedback on approximately 4,500 papers before the submission deadline. Its reports were private to authors and separate from formal peer review. A survey drew 869 responses, with 92.1% saying they would use the tool again and 31% of authors with experimental results saying the feedback prompted new experiments.
Those responses indicate that authors found useful suggestions, but they are not an independent accuracy test. The practical benefit is a chance to catch avoidable weaknesses before submitting a manuscript, while authors and reviewers remain responsible for checking the claims.
Free Access Comes With Document Retention
The Deep Review portal requires a Google account and advertises two weekly reviews using Argon at its Medium setting. Users who connect a Gemini API key can pay the underlying model-call costs for additional reviews; its FAQ says there is no separate Deep Review surcharge.
Google’s launch post says uploads are not used to train, fine-tune or improve its models. The portal says documents remain private and are retained for up to one year for users to revisit, while its FAQ separately offers an opt-in allowing the research team to use data to improve Deep Review.
For unpublished research or confidential business documents, retention and any optional consent therefore deserve attention before uploading. A useful review should also lead back to the original evidence: the calculation, assumption or passage allegedly at fault. The service can help prioritize that work, but its conclusions still need checking.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
