Google opens Deep Review to Docs and slides after 16,000 conference manuscripts
Google released Deep Review publicly on October 14, extending an agentic verifier tested on more than 16,000 manuscripts to Google Docs and presentations. Google says AI-assisted drafting has outrun human capacity to verify.
The Engineer · Build desk
Over 74,000 submissions to ICLR, NeurIPS and ICML this year, up from approximately 45,000 in 2025 and approximately 23,000 in 2023, per the Google Research post.
| Figure | Value | Note | Claim |
|---|---|---|---|
| ICLR, NeurIPS and ICML submissions this year, over | 74,000 submissions | 5 | |
| In 2025, approximately | 45,000 submissions | 5 | |
| In 2023, approximately | 23,000 submissions | 5 |
What happened
- Under its earlier name, the Paper Assistant Tool, it ran as an opt-in pilot with STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027, giving authors private feedback before each submission deadline.
- Pilot feedback was decoupled from official conference review, PAT collected no training data, and papers were deleted promptly after each program ended.
- Google says it checks logical consistency and mathematical statements, suggests improvements, grounds facts and compares with existing work, leaving higher-level judgements to humans.
Why it matters
- exposure Teams that run contracts or financial models through it are leaning on pilot evidence gathered from computer science authors.
- decision Run during drafting, the flags land in the author's queue. Run at review, they land with the reviewers the post calls the bottleneck.
- constraint The pilots never fed the tool's output into official review, so the reviewer-side use the post advertises has no pilot behind it.
The post's evidence for that gap is conference submission counts. More than 74,000 submissions went to ICLR, NeurIPS and ICML this year, against about 23,000 in 2023 [5]. That is a rise of about 3.2 times [10]. Against the roughly 45,000 of 2025 it is about 64% [11]. Each submission needs multiple high-quality reviews, according to the post [8].
The other side of the ratio is an assertion. Jayaram and Cohen-Addad wrote that "the human capacity for verification has remained largely fixed amid this rapid increase in AI-generated content." [13] The counts measure demand for review and say nothing about reviewer supply.
The post sets the system against single-prompt LLM calls. Those, it says, often fail on long documents because they exhaust their token reasoning budget, overlook dependencies between sections, or accept nuanced claims without close scrutiny [12]. Detail on the harness sits in a separate manuscript, "Towards Automating Scientific Review with Google's Paper Assistant Tool" [9].
The quality evidence in the pilot section is sentiment. The post says author reactions were "extremely positive" [14]. Those authors opted in [2], so the sample selected itself. The post also says the Gemini 4 Argon build is more capable than the older PAT iterations it describes further down [15]. The pilot record therefore describes a different system from the shipping build.
Transfer is the open question. For the pilot record to carry over to a contract or a forecast, the errors would need to show up as inconsistency or wrong mathematics. The facts would need an outside source to be checked against. Prior work would need to exist to compare with. All three are properties of the checks Google lists [4]. A published paper meets all of them. We'd expect an internal contract or forecast to meet fewer. The post names legal and finance as domains the system now covers [6].
Even where the checks fit, the output is a list of flagged items, and a person still has to read it.
The 16,000 figure spans STOC 2026 and ICLR 2027 as well as ICML and NeurIPS 2026 [2]. The 74,000 covers three conferences in a single year [5]. Dividing one by the other would compare two different sets.
What to watch
- Whether Google publishes catch rates and false-positive rates for the Gemini 4 Argon build, in the PAT manuscript or elsewhere.
- Whether the no-training and prompt-deletion terms from the conference pilots are stated for Docs and slides.
- Whether any conference moves Deep Review inside its official review process.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption30
- Hype gap+30
- Incentives65
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Google Research announced the public release of Deep Review on October 14, 2026, expanding it from deep mathematical and scientific peer review to knowledge documents including Google Docs and presentations.
- [2]
Under its experimental name, the Paper Assistant Tool (PAT), Deep Review was piloted with STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027; authors of more than 16,000 manuscripts who opted in received private feedback in the weeks before the submission deadline.
- [3]
The pilot feedback was completely decoupled from the official conference review process, PAT did not collect any data for training, and papers were deleted from Google systems promptly after each program concluded.
- [4]
Deep Review checks logical consistency and mathematical statements, suggests improvements, grounds facts, and compares with existing work, while leaving higher-level judgements to humans.
- [5]
This year there were over 74,000 submissions to ICLR, NeurIPS and ICML, up from approximately 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023.
- [6]
Deep Review was initially designed for computer science academic paper analysis; its capabilities have expanded to review documents across many domains, such as legal and finance.
- [7]
Deep Review can be used both as an iterative improvement tool during the drafting process and as a tool for verification of claims and evaluation during the review process.
- [8]
Each of the AI-conference submissions requires multiple high-quality reviews to ensure the integrity of peer review, which can place a larger burden on human experts.
- [9]
Further details on PAT and the pilots are in a manuscript titled Towards Automating Scientific Review with Google's Paper Assistant Tool.
- [10]
Submissions to ICLR, NeurIPS and ICML grew about 3.2 times from 2023 to this year.
- [11]
Submissions to the three conferences rose about 64% over 2025.
- [12]
Single-prompt LLM calls often fail when evaluating long, complex documents because they exhaust their token reasoning budget, overlook long-context cross-section dependencies, or accept nuanced claims without fully scrutinizing them.
- [13]
"the human capacity for verification has remained largely fixed amid this rapid increase in AI-generated content."
- [14]
Google says author reactions to the tool have been extremely positive, reflecting continued improvements to both the underlying models and the PAT harness.
- [15]
With the addition of Gemini 4 Argon support, Deep Review's underlying capabilities have further increased beyond the older iterations of PAT described later in the post.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- research.googleAnalyzing technical documents with Deep Review
1 article · October 14, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- AI-assisted peer reviewFollow
- Agentic AI SystemsFollow
- AI document verificationFollow