Image: Google

Google has opened up Deep Review to everyone. It's an AI tool that reads a long, complicated document from start to finish and tells you where the argument doesn't hold up. Think of it as a very picky second reader who checks the maths, the logic and the facts before your boss, your client or a journal does.

Google's Research team built it for science papers. Now it says the tool can also check design documents, slide decks, and legal and financial paperwork.

Today @GoogleResearch is launching Deep Review — an agentic system designed to stress-test manuscripts, design docs, and complex proposals before submission or publication. After pilot partnerships analyzing over 15,000 research papers, we’re making Deep Review globally accessible to help researchers, engineers, and professionals bring structured critique to their daily workflows.

— News from Google (@NewsFromGoogle) October 14, 2026

Here's a relatable example. Say you're signing off on a 40-page business proposal. The summary on page 2 promises a 15% saving, but the table on page 31 adds up to 9%. A tired human skims right past that. Google says Deep Review is built to catch exactly these mix-ups, checking that clauses, data tables and claims agree with each other and with outside evidence.

The reason it needs a special tool comes down to how chatbots read. Google says pasting a long document into a single AI prompt often fails. The model runs out of thinking room, misses links between sections, or accepts tricky claims without really checking them.

So Deep Review splits the job up. It breaks the document into sections and gives the hardest parts, like proofs or methods, more thinking time than the introduction. Separate AI agents check each section at the same time while still seeing the whole document. A final agent then pulls the findings together, removes repeats and runs web searches to check its own claims.

Image: Google

The tool has already had a big tryout. Under its old name, Paper Assistant Tool, it gave private feedback on more than 16,000 papers at four major computer science conferences, Google says. In one of them, ICML 2026, 35.4% of surveyed authors of theory papers said it found significant gaps that took over an hour to fix. Google's pitch is that human reviewers are swamped: it counts over 74,000 submissions to three big AI conferences this year, up from about 23,000 in 2023.

Anyone with a Google account can try it on the Deep Review site. You get two free reviews a week. After that, you plug in your own Gemini API key and pay only for the AI usage, and there's an API for developers. Google's blog says it takes PDFs and plain text for now, with multi-file support coming.

The fine print is where things get interesting. Google's own FAQ puts a price on a big document: roughly $3 to review 40 pages on the cheapest setting, and up to $20 on the most thorough setting with Gemini 4 Argon. That's cheap next to a human expert, but it adds up if you review every draft.

Privacy wording also differs between pages. The blog says uploads are "never" used to train Google's models. The tool's own page says they won't be used "without your permission", and its FAQ mentions an opt-in that lets the research team use your data to improve Deep Review. Either way, documents are kept for up to a year, so think twice before uploading anything confidential.

There are also limits to keep in mind. The pilot results come from surveys of authors who chose to take part, run by Google and the conferences, not independent tests. Google hasn't published accuracy figures for legal or financial documents. Even in its best survey, 97.2% of respondents said only that the "majority" of the feedback was factual. Google itself says the final judgement should stay with humans. Treat Deep Review as a sharp-eyed first pass, not the final word.