Google Research has opened Deep Review to anyone, a system that reads a long document and returns a critique of its logic, its mathematics and its claims. In its own blog post, Google says tools like it “will play a significant role across the peer review process in the coming years.” That is a toolmaker predicting the future of someone else’s profession, and the post backs it with little that can be measured.
Start with what the thing is. Google calls it an “agentic AI system,” which in practice means several model instances working in sequence instead of one prompt. A segmenter splits the document into sections, which may overlap. A budgeting step then gives dense parts, such as mathematical proofs, more compute than an introduction. Reviewer agents examine each section in parallel while seeing the whole document, so they can follow cross-references. A final agent merges the reports, reruns logic checks where needed, and makes outside lookups and tool calls as a guard against hallucination.
The word “verification” covers less than it suggests. By Google’s own account, the system compares a document with existing literature, grounds facts, proposes improvements, and tests whether the logic and the math hold together. Those are reading tasks. The post does not say Deep Review runs a paper’s code, reproduces an experiment, or confirms that every cited reference exists and says what the authors claim it says. “Grounds facts” may cover some of that, but the post does not specify. Think of a very attentive reader, not a lab that re-runs the results.
The evidence is thin. Under its earlier name, the Paper Assistant Tool, the system gave private feedback before submission deadlines to authors of more than 16,000 manuscripts who opted in, through pilots with STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027. The one result the post states is that “author reactions to the tool have been extremely positive.” It gives no count of real errors found, no rate of false alarms, and no comparison against human reviewers. It points readers to a preprint for “pilot evaluations” but summarizes none of them.
A second gap sits in the design. Google says the pilot feedback was “completely decoupled from the official conference review process.” No reviewer used the tool inside a real review. The prediction about peer review therefore rests on a pilot that never touched peer review, run on authors who chose to take part.
Access is straightforward. Every user gets two free reviews a week, and heavier users can connect a Gemini API key for higher limits and a Deep Review API. Google says uploads are never used to train its models and are stored for one year so the user can see them in the interface. Input is limited to PDFs and plain text, one document at a time, with multi-file support “coming soon.”
That limit is bigger than it sounds. A human reviewer often reads a paper alongside its appendix, supplementary material and code. A tool that sees one file cannot check the main text against a separate appendix. Google pitches the system for contracts and financial documents too, where exhibits and spreadsheets travel with the main file, and says multi-file support will cover those.
One tool serves two users who want opposite things, and that is where the logic gets tricky. An author running it before submission removes the errors it can detect. A reviewer running it afterward hopes to find the errors that remain. If authors adopt it first, the papers reaching reviewers are cleaner on exactly the axis the tool measures, which helps unless the flaws that matter most are the ones it cannot see. Google’s post does not address that, and it offers no data on which kinds of errors the tool misses.
Nothing in the post says any journal or conference has adopted Deep Review, and it does not tell researchers to rely on it. For now it is a free second reader to try on a draft, with its hit rate unpublished in the announcement.
Based on Google Research’s blog post “Analyzing technical documents with Deep Review” (publication date not stated in the source), retrieved 15 October 2026
