A document can read smoothly and still contradict itself. Google’s Deep Review is designed to look for those harder problems: a claim that outruns its evidence, a gap in a proof, or two sections that cannot both be right. Google opened the tool publicly on 14 October, extending an experiment in academic-paper feedback to a broader set of technical documents.
Based on 5 reviewed sources · See sources
Two reviews a week, with paid runs beyond that
The live application requires Google sign-in and offers two free reviews a week using Gemini 4 Argon at its Medium setting. Further runs use your own Gemini API key. Google charges for the underlying model calls, with no separate Deep Review fee.
Its pricing FAQ estimates roughly $3 for a 40-page document using Gemini 3.8 Flash on Low, rising to $20 for Argon on Extra High. Those are guides, not fixed document prices; the document and configuration affect the bill.
The launch post specifies PDF and plain-text inputs, with multi-file review still planned. The app’s FAQ also lists Google Docs and Slides, but the signed-out interface we inspected showed PDF upload and pasted text. That does not establish a native review button inside either Workspace app. Google Research introduced the public release in this announcement.
Google announces public Deep Review access after opt-in academic-paper pilots.
View the original post on X ↗What the conference trials actually measured
Deep Review grew out of Google’s Paper Assistant Tool, or PAT. Google says the pilots covered more than 16,000 manuscripts. Its system divides a document into sections, gives reviewing agents the full document for context, then combines their findings and checks supporting sources. This describes the intended process; it is not a measured accuracy rate.
The ICML retrospective gives a more specific view of the earlier tool. Approximately 4,500 papers received feedback; 869 authors answered its survey. Of respondents, 92.1% said they would use it again. Among respondents whose papers had experimental results, 31% said the feedback prompted new experiments. These are voluntary author reports from the pilot, not tests of today’s Argon-powered release.
The retrospective was co-written by Google researchers and ICML programme chairs. It includes this account from Peihan Wu of Shanghai Normal University, describing feedback on a proof:
“PAT discovered that one of my theorem proofs was not rigorous and provided some possible solutions.”
A different retention policy from the pilot
The public app says documents stay for up to a year for users to revisit, with no model training without permission. Its FAQ offers an opt-in to help improve Deep Review. For confidential material, storage and training are separate decisions.
That differs from the NeurIPS pilot, whose published policy scheduled PDFs and feedback for deletion within seven days after feedback delivery and programme completion. The pilot also kept feedback out of official peer review.
NeurIPS warned that PAT could flag correct statements or miss real flaws. That remains the question worth testing in the public product: which criticisms survive a human check, and which send the author chasing a problem that is not there? We have not independently tested the new release; pilot testimonials do not settle that question.
Public access and the earlier NeurIPS pilot
Different programmes, not interchangeable promises.
| Detail | Public Deep Review | NeurIPS 2026 PAT pilot |
|---|---|---|
| Access | Two free reviews per week; own API key for more | Optional, time-limited author voucher programme |
| Document retention | Up to one year for display to the user | Deletion within seven days after feedback delivery and programme completion |
| What the evidence supports | Public availability and stated terms; no independent test by us | Pre-submission assistance, separate from official peer review |
Sources
- Deep Review public launch
Google Research · Announcement ·
- Deep Review application and FAQs
Google Research · Documentation
- Original Deep Review announcement
Google Research on X · Announcement ·
- PAT × ICML 2026 retrospective
ICML / Google researchers and programme chairs · Research ·
- NeurIPS author-assistance pilot and privacy policy
NeurIPS / Google pilot organisers · Documentation ·
Written with AI assistance from the sources above. Analysis and practical implications are our interpretation. Our editorial approach.
