Google Research has made Deep Review available worldwide. You upload a PDF or a plain text file, and an AI system reads it section by section and reports where claims do not follow, where numbers disagree with each other and where terms are ambiguous. It began as a private tool for authors submitting papers to four computer science conferences. It now also works on legal and financial documents, presentations and software design documents.
One question to a chatbot misses too much
You could already paste a report into a chatbot and ask it to find mistakes, and that works better than most people expect. On a test set of 29 known errors in math and computer science papers that were later corrected or retracted, a single request to Gemini 3.1 Pro found 55%. An earlier version of the pipeline behind Deep Review found nearly 90%.
A single request fails for an ordinary reason. The model spreads a limited amount of effort across the whole document, loses track of what page 3 promised by the time it reaches page 40, and accepts claims that sound right. Deep Review works more like an audit team. It splits the document into sections and gives the hard ones, such as a dense proof or a pricing table, more effort than an introduction. A separate reader checks each section, and every reader can see the whole document. A last reader reconciles their notes and checks claims against outside sources.
What it changes in ordinary work
Take a finance director at a hospital group with a 45-page supplier agreement. Page 6 caps yearly price rises at 3%. A table in the appendix, built for the budget, assumes 5%. Nobody reading line by line on a Friday afternoon holds both numbers in their head. A check across clauses and tables is built for exactly this. So is a shop owner's lease whose renewal clause points to a section that was renumbered.
Done by hand, a careful pass like that takes a few hours per document, so two a week is roughly a working day back for whoever reviews the most paperwork.
What comes back is a list of suspicions, not verdicts. In the conference pilots, the model sometimes stumbled on complex notation and figures, and the authors sorted the useful flags from the wrong ones themselves. That is still a good trade. Checking a flagged point on page 6 against the appendix table takes a minute, while finding it took the afternoon nobody had.
Limits to know before you rely on it
For now it reads PDFs and plain text, so a spreadsheet has to be exported to a PDF first. Reviewing several files together, spreadsheets included, is promised later. The free allowance is two reviews a week, and heavier users connect their own Gemini API key. Uploaded documents are stored for a year so you can see them in the tool, and they are not used to train Google's models. For a confidential contract, check your company's rules before uploading.
Reviewing is now the slow step
Google built this because AI writes drafts faster than people can read them. Submissions to three big AI conferences went from about 45,000 in 2025 to about 74,000 in 2026, and each one needs several expert reviews. Offices have the same shape. AI drafts the contract, the proposal or the monthly report in minutes, and the person who must sign it has the same hours as before.
Reading faster cannot close that gap. A first pass done by a machine can, with the person left to decide what each flag means. Within a few years, sending a long document that no machine has checked will look like sending one that skipped spellcheck, and the people who receive your documents will run the check themselves. When anyone can do that in minutes, "nobody caught it" stops being an excuse and becomes a choice. The people who run the check first are the ones who get to fix the problem before anyone else sees it.