Google released Deep Review for analyzing technical documents
Google Research ● Covered by 2 sources
Google released Deep Review, an AI tool that checks technical docs for logic, math, and missing evidence. It’s meant to ease the review bottleneck as AI floods papers, contracts, and other long documents.
Based on reporting by Google Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google has put Deep Review out in the open, pitching it as an AI system for the hard part of technical work: checking whether a document actually holds together. The original target was computer science papers, but Google says it now reaches into legal, finance, design docs, and presentations too.
The timing is not subtle. Google says this year there were more than 74,000 submissions to three major AI conferences — ICLR, NeurIPS, and ICML — up from about 45,000 in 2025, 32,000 in 2024, and 23,000 in 2023. That kind of growth turns review into a bottleneck fast, especially when each submission needs several strong reads from humans who already have day jobs.
Deep Review is built to do the slogging part of that job. It checks logical consistency, mathematical statements, facts, and comparisons with existing work, while leaving the final judgment to people. Google says the system breaks a document into segments, allocates more compute to denser sections, runs specialized reasoning agents in parallel, then synthesizes the results and does grounding checks to cut down on hallucinations.
This is not a brand-new idea wearing new clothes. Deep Review grows out of Google’s Paper Assistant Tool, or PAT, which was tested in pilots with STOC 2026, ICML 2026, NeurIPS 2026, and ICLR 2027. More than 16,000 manuscripts opted into private pre-submission feedback during those pilots, and Google says the feedback was kept separate from official conference review, with papers deleted promptly after each program ended.
The public version comes with some plain limits and a few useful ones. Users get two free document reviews a week, higher-volume users can hook in a Gemini API key, and the first supported inputs are PDFs and plain text. Google also says uploaded documents are not used to train or fine-tune its models, and are stored for one year only so they can be shown back in the Deep Review interface.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of AI use: tedious, checkable, and aimed at work that already has standards instead of vibes. The bigger tell is that Google is selling verification, not replacement, which is exactly where these systems belong. The industry still prefers grand promises, but the boring tools that save experts time are usually the ones that stick.
Read more about this at: Google Research