This week: Google opened up an AI that reads long, dense documents the way a picky reviewer would, and you can use it free twice a week.

Elsewhere: What happened when the authors of about 4,500 papers let the beta pick them apart, plus how to run your first review.

The New Prime Signal

Everyone has written something long and important and then wished a sharp-eyed expert would read it before anyone else did. On October 14, Google Research announced the public release of Deep Review, which it describes as "an agentic AI system engineered to provide rigorous feedback and technical verification on complex documents." It started life checking computer science papers. Now Google is pointing it at design docs, slide decks, and legal and financial documents too. It won't tell you your ideas are brilliant. That's rather the point.

What shipped: Deep Review is the public version of Google's Paper Assistant Tool (PAT), which ran pilots with four major computer science conferences: STOC 2026, ICML 2026, NeurIPS 2026, and ICLR 2027. Authors of more than 16,000 manuscripts opted in for private feedback before their deadlines. With the launch, every user gets two complimentary document reviews per week. Heavier users can connect a Gemini API key for higher limits and use the Deep Review API in their own workflows. It takes PDFs and plain text for now, with multi-file input "coming soon." Google says uploaded documents are never used to train, fine-tune, or improve its models, and are stored for one year only so they can be shown back to you in the Deep Review interface. The tool has also added Gemini 4 Argon support, which Google says raises its capabilities beyond the earlier PAT versions.

Why it matters: AI has made writing faster, but checking still happens at human speed. Google's example is AI research itself: more than 74,000 submissions to ICLR, NeurIPS, and ICML this year, up from about 45,000 in 2025, 32,000 in 2024, and 23,000 in 2023. Deep Review checks logical consistency and mathematical statements, suggests improvements, grounds facts, and compares a document with existing work, "while leaving higher-level judgements to humans." So it's a second reader, not a judge. Google also points out that this kind of thorough pre-submission critique has usually depended on access to senior faculty, well-resourced labs, or big advisory networks. A free tier puts a version of it in front of anyone with a PDF.

What's next: Google says multi-file support is on the way, so the same consistency checks can run across a whole set of related documents, spreadsheets, and exhibits at once. For people who build with Google's Antigravity agent tool, there's also a new Teamwork Deep Document Review Pattern, inspired by Deep Review and started with the /teamwork-preview command. Google says it expects tools like this to "play a significant role across the peer review process in the coming years." Worth watching: whether more conferences sign up, and how well the checks hold up on legal and finance documents, where the tool is newer.

A lot of AI news is about machines writing more. This one is about a machine helping people check what's already been written, which is the part that has been falling behind. Two free reviews a week won't replace a sharp colleague, but it's a pretty good warm-up before you face one.

Before going public, the tool ran a pilot at ICML 2026 from January 14 to 26, giving feedback on about 4,500 papers, usually within about 30 minutes, according to the ICML retrospective. In a survey that drew 869 responses, 92.1% said they would use it again and 73.3% rated the feedback "Very" or "Mostly" helpful. Only 1.6% found it not useful at all. Among authors of papers with theory, 35.4% said it identified significant theory gaps that took more than an hour to fix, and 31% of authors with experimental results said it prompted them to run new experiments. That's a survey of people who chose to take part, so read it as encouraging rather than definitive.

How it works: Google says one-shot chatbot prompts tend to stumble on long documents. They run out of reasoning budget, miss links between sections, or wave through subtle claims. Deep Review uses a four-stage pipeline instead. First, a segmenter agent splits the document into logical sections. Second, adaptive budgeting gives dense parts, like proofs, more compute than an introduction. Third, specialized agents check each section in parallel while still seeing the whole document for cross-references. Fourth, a synthesis agent merges the reports, re-runs consistency checks between sections, and does external grounding checks as a guard against hallucinations. The method and pilot results are in the preprint on arXiv.

Try it: Open the Deep Review page on Google Research and spend one of your two free weekly reviews on a PDF or plain-text file you're comfortable uploading, like a draft report, a design doc, or a proposal. Read the results like notes from a smart, very literal colleague, and check each flagged issue yourself before you change anything. Keep confidential or client files out unless you're cleared to share them, since Google keeps uploads for a year for display.