Google Deep Review: what the AI paper checker does

What is Google Deep Review?

Google Deep Review is a free AI tool from Google Research that reads a paper or technical document and points out logic gaps, math errors and claims without evidence. It opened to the public on October 14, 2026, and anyone with a Google account can use it. Behind it are several AI agents (AI systems that plan steps and work toward a goal on their own), each taking one part of the document.

Google Research built it with Google DeepMind and Google Cloud. It started as a tool for computer science papers and now also handles legal and financial documents. Google pitches it both for drafting, where you revise and run it again, and for review, where you need to test someone else's claims.

It is not a grammar checker or a plagiarism detector. Deep Review looks at whether the argument and the math hold up, not at spelling or originality.

It grew out of the Paper Assistant Tool

Deep Review is the public version of the Paper Assistant Tool (PAT), an experiment to stress-test manuscripts before they reach a conference. The pilots ran during the last year at STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027, four flagship venues in the field.

Authors who opted in got private feedback before the submission deadline, and more than 16,000 manuscripts went through it. The feedback had no link to the official review process. Google says PAT kept no data for training and deleted each paper once its program ended.

AI conference submissions roughly tripled in three years

Submissions to ICLR, NeurIPS and ICML combined

2026 is more than 74,000. Source: Google Research blog.

ICLR, NeurIPS and ICML together received about 23,000 submissions in 2023 and more than 74,000 in 2026, roughly three times as many in three years. Google gives this growth as the reason it built Deep Review.

The pool of qualified reviewers has not grown at that pace. The PAT paper notes that checking dense proofs line by line can take a human reviewer days. arXiv has already started limiting how many new computer science papers one person can post (see arXiv caps new paper submissions at two per month), which shows how hard the volume has become to manage.

What Deep Review checks in a paper

Deep Review checks four things: whether the logic holds together, whether the math is correct, whether facts have support, and how the work compares with earlier research. For each problem it finds, it also suggests a fix.

Google frames it as taking on the most tiring parts of review while leaving the big calls to people. Whether a paper is novel or worth accepting is still a human decision.

Logic, math, evidence and prior work

The output goes beyond a list of mistakes and includes suggested rewrites and a comparison with related work. To find that related work, Deep Review may search the web. Google says the document itself stays private while it does so.

For a theory paper, that means flagging a step in a proof that does not follow, or a lemma that quietly depends on an assumption stated nowhere. For an empirical paper, it means asking whether the results actually back the claim in the abstract.

Design docs, slides, and legal or financial documents

Google says the same checks work on any document where claims need evidence, and names three other kinds. "What Deep Review looks for in each document type" lists them.

What Deep Review looks for in each document type

Document	What Deep Review looks for
Design docs	Architecture assumptions nobody wrote down, edge cases, concrete fixes
Slides	Whether one slide leads to the next, whether bullet points match the charts
Legal and financial documents	Clauses and tables that contradict each other, vague terms, claims without backing

So the audience is wider than academics. An engineer writing a design doc or a lawyer reviewing a contract can run the same check.

Deep Review pricing and free reviews

Deep Review access at a glance

Free

Two reviews a week on Gemini 4 Argon, Medium setting

Paid

Your own Gemini API key, billed at Gemini API rates

Cost guide

About $3 to $20 for a 40-page document

Every user gets two free Deep Reviews a week, and anything beyond that is billed to your own Gemini API key. Google adds no separate fee for Deep Review.

Two free reviews a week on Gemini 4 Argon

The free reviews run on Gemini 4 Argon at the Medium setting. Gemini 4 Argon is the model Google made generally available on October 13 (see Gemini 4 Carbon vs Argon: price and availability). Two reviews cover a typical cycle for one paper: one before you submit and one after you revise.

Paying for more reviews with a Gemini API key

To go past two reviews a week, you save a Gemini API key on the Deep Review page. An API key is a credential that lets your own code call Google's models and bills that usage to you. You use the same key to call the Deep Review API.

You pay only for the Gemini API calls the review makes, at standard rates. The Deep Review page puts a 40-page document at about $3 with Gemini 3.8 Flash on Low and $20 with Gemini 4 Argon on Extra High. Google calls those numbers a guide, since the real cost depends on the document.

How to use Google Deep Review

Three ways to run Deep Review

Web

Upload a file on the Google Research Portal

API

Call it from your own pipeline with a Gemini API key

Antigravity

Run a similar review pattern with a slash command

You can run Google Deep Review from the web page, through its API, or inside Antigravity. The web page suits a single paper. The API suits teams that want every draft checked automatically.

Upload a PDF or paste text on the web

The quickest route is the Deep Review page on the Google Research Portal. All you need is a Google account and the file you want checked.

1. Sign in with your Google account.
2. Upload a document, or switch to the text tab and paste it in.
3. Optionally add context, such as the sections you care about or the venue you are submitting to.
4. Submit, then open the result from the "Your documents" table once it is ready.

Call the Deep Review API

With a Gemini API key, you can call the same review from code. Google suggests wiring it into continuous integration (the automated checks that run each time a file changes) and drafting workflows. That way, a review can run whenever a document changes.

A team that keeps design docs in a repository could, for example, review every pull request that touches a spec. The API reference is listed in Google APIs Explorer, according to the Deep Review FAQ.

Run the Teamwork pattern in Antigravity

Google also added a review pattern inspired by Deep Review to Antigravity, its agentic development tool. It is called the Teamwork Deep Document Review Pattern, and it runs on Teamwork, Antigravity's framework for coordinating several agents.

The pattern is meant for research papers, manuscript drafts and technical proposals. In Antigravity you start it with the /teamwork-preview command, without leaving your editor.

Inside the four-stage review pipeline

The four stages of a Deep Review

1. Split
   An agent breaks the document into overlapping sections
2. Budget
   Dense sections get more compute than light ones
3. Check
   Specialist agents review the sections in parallel, each with the full document
4. Combine
   One agent merges the reports and checks them against outside sources

Deep Review does not hand the whole paper to one model in one prompt; it splits the work across four stages. Google says the move to Gemini 4 Argon makes these reviews deeper than the ones PAT produced.

Why a single prompt misses errors in long documents

Google gives three reasons a single prompt falls short on a long, technical document. Each one explains part of why Deep Review splits the job up.

• The model runs out of reasoning budget before it reaches the hard parts.
• It loses track of how a definition on page 3 connects to a proof on page 20.
• It accepts subtle claims without really testing them.

So the first stage is a segmenter agent that cuts the document into logical sections, such as the introduction, methods and experiments. The sections may overlap, and together they must cover every page.

Proofs get more compute, then agents check sections in parallel

Each section gets a compute budget that matches how dense it is, so a block of proofs receives more than an introduction or a related work section. Specialist agents then review their sections at the same time. Each one can see the full document, so it can notice when a symbol defined early is used differently later.

A final synthesis agent merges the section reports, checks them against each other and runs outside checks. In the PAT version, that agent used Google Search to catch citations to papers or theorems that do not exist.

How accurate is Deep Review?

Share of known math and proof errors detected

26 math and computer science papers containing 29 known errors, taken from the SPOT benchmark. Higher is better. The original SPOT figure was graded differently, so it is not a direct comparison. Source: PAT paper.

21.1%Original SPOT best

55.2%Gemini 3.1 Pro, one call

89.7%PAT

0255075%

In Google's test, Deep Review's predecessor found 89.7% of 29 known errors in 26 math and computer science papers, but authors who used it in real pilots rated its feedback as mostly grounded only 56% to 65% of the time. It catches a lot, and it also raises points that turn out to be wrong. Plan to verify each flag yourself.

89.7% of known errors, versus 55.2% for one Gemini 3.1 Pro call

The same Gemini 3.1 Pro model, given each paper in a single call, found 55.2% of the errors. PAT ran on that same model and reached 89.7% by splitting the work, so the gain comes from the pipeline, not a bigger model.

The papers come from SPOT, a benchmark of published papers with confirmed mistakes, filtered to equation and proof errors. The best result reported with SPOT itself was 21.1%, but it was graded another way. According to the PAT paper, single-call models tend to accept complex math claims without really checking them.

What authors said in the STOC and ICML pilots

More than nine in ten authors in both pilots called the feedback very or mostly helpful. PAT ran with STOC in November 2025 and with ICML in January 2026, and reviewed over 4,700 submissions across the two.

The ICML numbers stand out. In that pilot, 35.4% of authors were told about a substantive gap in their theory, and 31% ran new experiments after reading the review.

Only 56% to 65% found the feedback mostly grounded

When asked whether most or all of the feedback was factually grounded, 55.8% of STOC authors and 64.8% of ICML authors said yes. The rest did not say so, which means a real share of flags may rest on a misreading.

The researchers name hallucination, where an AI states something false as if it were fact, as a key weak point of AI reviewing. Their advice is plain: whoever uses a review agent has to check what it says. Treat each Deep Review flag as a lead to confirm against your own text, not a correction to accept.

What happens to the documents you upload

Data and file formats

Training

Not used to train, fine-tune or improve Google models

Storage

Kept for one year so you can view results

Formats

PDF and plain text at launch, multi-file input coming

Google keeps uploaded files for one year and says it does not train its models on them. That matters if you plan to upload an unpublished paper or an internal document.

Not used for training and kept for one year

The announcement says uploads are never used to train, fine-tune or improve Google models, and are stored only so you can see them in the Deep Review interface. The Deep Review page words it a little differently: documents are not used for training without your permission, and you can opt in to let the research team use your data to improve the tool. If your lab or company has rules about sending manuscripts to outside services, check them before you upload.

File formats: PDF and plain text, plus Docs and Slides

At launch, the announcement lists PDF and plain text as the input formats, with multi-file input coming soon. The FAQ on the Deep Review page goes further and lists Google Docs and Google Slides as well.

Until multi-file input arrives, a paper with a separate appendix has to go in as one file. If the main text and the appendix are separate PDFs, combine them into one PDF before you upload.

If you need to combine the main paper and its appendix, try our tool that merges several PDFs into one and lets you drag them into order first. It runs in your browser, so you can prepare the file before you upload it to Deep Review.

Use the free reviews for one paper and the API for a team

Google Deep Review gives you two free reviews a week, which is enough to check a paper before you submit it and again after you revise. If you need more, the API route costs roughly $3 to $20 for a 40-page paper.

Put the two headline numbers side by side and the right way to use it becomes clear. Its predecessor caught 89.7% of known errors, yet only 56% to 65% of authors found the feedback mostly grounded. Deep Review is good at surfacing problems you might have missed, so read each flag against your own text before you change anything.