What Is Google Deep Review? Free AI Document Checker (2026)

TL;DR. Google Deep Review is a free AI tool, public since October 14, 2026, that reads a finished document and flags logic gaps, shaky math and missing evidence. You get 2 free reviews a week (Google Research, 2026). Its best results so far come from math and computer science papers, not business documents.

Last reviewed: 2026-10-19

Most AI writing tools help you start a document. Deep Review helps you finish one. You hand it something you’ve already written, and it behaves like a skeptical reader with plenty of time: it checks whether your numbers add up, whether section four contradicts section two, and whether that confident claim on page three has anything behind it.

Deep Review is a document-checking tool from Google Research. It uses a team of AI agents, not a single prompt, to review a long document section by section and then reconcile what each agent found. In plain terms: you upload a draft, wait about twenty minutes, and get back a list of problems with suggested fixes. This page is FindSkill.ai’s plain-language reference. I’ve kept to what Google, the arXiv papers and the first public hands-on reports confirm as of October 19, 2026, and I’ve marked what nobody has measured yet.

What Deep Review actually is

Deep Review is Google Research’s public version of an earlier tool called the Paper Assistant Tool, or PAT. Google first offered PAT to authors at computer science conferences as a pre-submission check, and says it reviewed more than 16,000 submissions at STOC, ICML, NeurIPS and ICLR (Google Research, October 14, 2026). On October 14 Google opened the tool to everyone and widened the scope beyond research papers.

Three details tell you what kind of tool it is:

It reviews; it doesn’t write. Deep Review takes a document you already wrote. It doesn’t draft one for you. That makes it the opposite job from the tools that generate a first draft.
It’s built from several agents. Google describes a four-stage process, covered below, instead of one big prompt.
It’s meant to challenge you. According to Google Research (October 2026), it checks logical consistency and mathematical statements, grounds facts, and compares claims against existing work, while leaving higher-level judgments to people.

Vahab Mirrokni, one of the Google researchers behind it, wrote in the launch announcement that the public version was “improved and expanded” for “all types of documents: legal, finance, presentations, design docs, and others” (X, October 15, 2026).

One naming warning before anything else. Deep Review is not Deep Research. Deep Research searches the web and writes a report about a topic. Deep Review does the reverse: it takes your own document and looks for errors in it. If you searched for one and landed on the other, you’re in the wrong product. Deep Review lives at research.google/deepreview.

How Deep Review works

Deep Review works in four stages: it splits your document into sections, decides how hard to think about each one, reviews the sections in parallel while every agent can see the whole document, and then merges the results into one report. Google’s launch post gives these four stages, and they explain both what the tool is good at and why a review takes a while.

How a Deep Review runs

Segment split into sections

Budget more effort on dense parts

Review in parallel each agent sees the whole document

Synthesize merge, rank by severity, check facts

Based on Google Research's launch post (October 14, 2026).

Segmentation. A first agent divides the document into logical sections and has to cover all of the content.
Adaptive budgeting. The system gives easy sections, such as an introduction, light attention and dense ones, such as proofs or methods, much more computing time.
Parallel review. Separate agents review each section at the same time, but each one can see the full document, so a claim on page 3 can be checked against a number on page 19.
Global synthesis. According to Google’s diagram (October 14, 2026), a final agent merges the reports, removes duplicate complaints, checks how severe each issue is, and runs searches to ground what it found.

This is the same idea behind multi-agent orchestration: several narrow agents, one coordinator. It’s also an example of agentic AI in a very practical form, since the tool plans its own review of your document instead of answering one question.

What Deep Review checks for

Deep Review looks for four kinds of problems: logic gaps, inconsistent claims, mathematical errors and missing evidence. Findings come back with severity labels such as critical, major and minor, plus suggested fixes tied to the text. A user who tested it on a 35-page technical specification reported 36 findings (6 critical, 14 major, 10 minor) including a sampling rule that needed 40 runs but allowed 30 (X, October 18, 2026). That’s one person’s account of one technical document, not a benchmark.

What it flags	Plain-language example
Logic gaps	A conclusion that doesn’t follow from the steps before it
Inconsistent claims	Section 2 says 120 participants; the budget pays for 100
Math errors	A subtotal that doesn’t match its line items
Missing evidence	A confident statistic with no source behind it

What it isn’t: an AI-content detector, a plagiarism checker, or a grammar tool. It doesn’t tell you whether a robot wrote your text. It tells you whether your document holds together.

Who can use Deep Review, and what it costs

Anyone with a Google account can use Deep Review, and every user gets two free reviews per week (Google Research, 2026). According to Google’s product page (October 2026), the free tier runs Gemini 4 Argon at Medium depth, and heavier users can add their own Gemini API key and pay only for the underlying model calls.

Route	Cost	What you get
Free tier	$0	2 reviews per week, Gemini 4 Argon at Medium depth
Your own Gemini API key	Standard API rates	More reviews, higher depth settings
Google’s cost estimate for a 40-page document	About $3 to $20	$3 for Low with Gemini 3.8 Flash, up to $20 for Argon on Extra High (Google, “only a guide”)

A few practical limits, per Google’s pages and early reports:

File types. The launch post says PDF and plain text, with multi-file uploads “coming soon.” The live page also lists Google Docs and Google Slides.
Time. According to a hands-on report from a Japanese academic (X, October 16, 2026), a review takes about 20 minutes.
Context box. There’s an optional “custom context” field where you tell the tool what the document is and who will read it.
Languages. Google hasn’t published a Deep Review-specific language list.

The two-a-week cap works like other AI usage limits: it’s a rationing system, so spend a review on the document where a mistake costs the most.

How accurate is Deep Review?

Nobody knows yet for business documents. The numbers Google and the coverage cite come from academic papers and from an earlier version of the tool, and no independent group has published a test of the public product. That’s the honest answer, and it matters more than any headline percentage.

Here is what is published:

The 89.7% figure. Several outlets say Deep Review “detected 89.7% of known errors.” That number describes the earlier PAT tool on a small test: 26 papers containing 29 known math or proof errors, of which it found 26. A single Gemini 3.1 Pro answer found 55.2% (Digital Today, October 2026). The arXiv paper itself describes the gain as a “34% improvement over zero-shot recall on mathematical errors” and doesn’t report false positives. The 89.7% is simply 26 divided by 29.
Pilot surveys. In Google’s table, 97.2% of surveyed NeurIPS and ICLR authors said most of the feedback was factual, and 58.5% of ICLR authors ran new experiments afterward (Google Research, October 14, 2026). These are self-reported answers from authors who chose to use the tool.
Grant proposals. The closest independent evidence is a March 2026 preprint that tested AI review on six real grant proposals. It found that section-by-section review worked best, that alignment problems were usually caught, and that clarity problems were mostly missed (Thorne et al., arXiv 2603.08281). It’s small and exploratory.
False positives. According to Digital Today (October 2026), there are cases where the system judges a valid proof to be an error. That’s why the output is a list of questions to check, not a verdict.

For a plain-language look at why AI tools sometimes state or flag things wrongly, see our guide to AI hallucination.

What happens to your document

According to Google’s Deep Review FAQ (2026), uploaded documents are stored for up to one year so you can see them in the interface, and are not used for model training unless you give permission. The tool may also run web searches while it works. That’s a reasonable promise for a free consumer tool, but it isn’t a confidentiality agreement.

Google’s public materials don’t mention a data-processing agreement or a zero-retention mode. So the practical rule is simple: strip names, client identifiers and anything covered by an NDA before you upload, and never upload material that isn’t yours to share. Funder and institutional rules apply on top of Google’s. For example, NSF bars its reviewers from uploading proposal content to AI tools it hasn’t approved, and a person reviewing someone else’s application should never put it into Deep Review (NSF, December 2023). Our AI Privacy 101 course covers what to keep out of any AI tool.

What Deep Review can’t do

Deep Review can’t judge whether your idea is good, whether your argument will persuade a particular reader, or whether your numbers are right when only you know the real figure. It can only tell you where the document disagrees with itself or leans on nothing.

It hasn’t been tested on most business documents. Google’s strongest results are on math and computer science papers.
It doesn’t know your reader. Unless you describe the audience in the context box, it reviews your document as a general document.
It can be wrong in both directions. It can flag something that’s correct and miss something that’s wrong.
It reads one document at a time. Multi-file input isn’t available at launch, so a budget in a separate spreadsheet has to be pasted into the same file.
It doesn’t replace a human reader. Google itself says higher-level judgment stays with people.

What this means for your profession

Deep Review is most useful wherever a document has to be internally consistent and a mistake is expensive: specs, financial statements, proposals and applications. Each section below gives a first thing to try, and none of them means skipping your own check.

What this means for developers

Developers write design documents full of assumptions and edge cases, and the cross-references between sections are where they drift. Google names design docs as a core use. The sensible first step is to run a design document through Deep Review before the review meeting and read only the critical and major findings. The 35-page spec test above is the clearest public example. Never upload code or documents that your employer’s policy keeps in-house. Our developers course and multi-agent systems course cover the engineering side of the same ideas.

What this means for accountants and finance teams

Finance documents have a built-in consistency test: totals have to match their parts, and the narrative has to match the numbers. Google lists financial documents among its examples, but the strongest published results are on math papers, so treat any finding as a prompt to recheck in your spreadsheet. First step: run a board-report draft with client names and account numbers removed, then verify every flagged total yourself. See the accountants and finance course for the broader workflow.

What this means for freelancers and consultants

A proposal or a client report is a document where a small internal contradiction costs you credibility. Deep Review is a free cold read before you send it. First step: de-identify the client, run the proposal, and fix only the findings you can reproduce. The freelancers course covers pricing and proposals, and the AI for consultants client decks course covers keeping client information out of AI tools.

What this means for nonprofit staff and grant writers

Grant proposals combine a narrative, a timeline and a budget that must all agree, which is the exact kind of cross-section problem Deep Review is built to catch. There’s no independent test on grants yet, and funders such as NSF and NIH have their own AI rules, so check them first. We wrote a full walkthrough in Google Deep Review: A Free Check for Your Grant Proposal. The AI for nonprofit grant writing quick skill covers drafting and disclosure, and the nonprofits course covers the wider picture.

Common misconceptions about Deep Review

“Deep Review is Google’s AI that writes my document.”

It doesn’t write anything. Deep Review reads what you wrote and challenges it, which is the opposite of what a drafting assistant does. Writing a first draft is a different product category, and while Deep Review suggests fixes tied to the text, it leaves the rewriting to you.

“89.7% accurate means 9 out of 10 findings are right.”

It doesn’t. The 89.7% figure is the share of 29 known math errors that the earlier tool found on a small test. It says nothing about how many of its complaints are correct, and it wasn’t measured on business documents.

“Deep Review is the same thing as Deep Research.”

They’re two separate Google products with similar names, and they do opposite jobs. Deep Research searches the web and writes a report from the sources it finds. Deep Review never writes a report from scratch: it checks a document you already have and lists the problems in it.

“A clean report means my document is good.”

A clean report only means Deep Review found nothing it could flag in the text it was given. It can’t know your audience, your real figures or your goals, and Google reports no independent accuracy test yet, so a clean result is encouraging but never proof that a document is good.

The bottom line

Deep Review is worth one free try on a document where consistency matters and a mistake is costly. Take the confidential details out first, verify every finding before you change anything, and treat the result as a list of questions, not a grade. It’s strongest for math-heavy and technical writing today, and it’s unproven for most other documents.

Use your two weekly reviews on the documents that matter most. And if you want to get better at checking AI output in general, our fact-checking AI course teaches a quick verify habit that applies to every tool on this site, and AI Fundamentals is the place to start if the terms on this page are new.

FAQ

What is Google Deep Review in simple terms?

Deep Review is a free Google tool that reads a document you’ve already written and tries to find what’s wrong with it. It looks for logic gaps, claims that contradict each other, math that doesn’t add up and statements with no evidence behind them. Google opened it to the public on October 14, 2026.

Is Google Deep Review free?

Yes, up to a point. Every user gets two free reviews per week, run on Gemini 4 Argon at Medium depth. If you need more, you can connect your own Gemini API key and pay standard rates. Google estimates about $3 to $20 for a 40-page document, depending on the settings.

Is Deep Review the same as Google Deep Research?

No. Deep Research searches the web and writes a report on a topic you ask about. Deep Review takes a document you already wrote and checks it for errors. They are separate products with similar names, and they do opposite jobs.

How accurate is Google Deep Review?

Nobody has measured it independently yet. Google’s earlier version of the tool caught 26 of 29 known math errors in a small test of 26 papers, and most surveyed pilot authors said the feedback was factual. Those results come from Google and from research papers, and the tool sometimes flags correct material, so every finding needs a human check.

Does Google use my documents to train its AI?

Google says uploaded documents are not used for model training without your permission, and you can choose to opt in. Documents are stored for up to one year so you can view them in the Deep Review interface. Remove names and confidential details before uploading anything sensitive.

What kinds of documents can Deep Review check?

Google’s launch post says PDF and plain text at launch, and the live page also lists Google Docs and Google Slides. Google’s examples are research papers, design documents, presentations, and legal and financial documents. Its strongest published results are on math and computer science papers.