Google launches Deep Review: found errors in 35 percent of ICML papers, according to the company
The system has a year of conference pilots behind it, but the error-detection figures come from Google itself – and independent evaluation is still lacking.
Google announced on October 14, 2026 the public release of Deep Review, an agentic AI system that provides thorough feedback and technical verification on complex documents. Behind the launch lies a year of closed pilots under the name Paper Assistant Tool, in which more than 16,000 manuscripts from four major computer science conferences were analyzed at the authors' own initiative. The tool is now being opened to everyone – and it is no longer limited to academic papers.
From conference pilot to public tool
According to the Google Research blog, the company has over the past year collaborated with four flagship computer science conferences – STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027 – to provide private pre-submission feedback on more than 16,000 manuscripts where the authors had opted in. The pilot operated entirely outside the official peer review process: authors could sign up and receive feedback privately in the weeks before submission deadlines, according to secondary coverage of Google's Research@ 2026 event.
In the public version, the scope has been expanded. The system, which was originally designed to analyze academic computer science papers, is claimed by Google to now be able to review documents across many domains, such as legal texts and finance – and it can reportedly be used both during writing and review, including Google Docs and presentations. These are the company's own claims; no independent testing of the new domain capabilities exists in the available evidence base.
What the pilot figures show – and what they don't
The most concrete results from the pilot are self-reported. According to the secondary coverage, 11.6 percent of STOC authors – a cohort from math-heavy theoretical computer science – reported that they had found substantial errors that took more than an hour to fix. Among ICML authors, where the focus is machine learning, the figure rose to 35.4 percent. On satisfaction, the coverage reports figures without cohort breakdown: 94 percent of authors described the feedback as helpful, and 85 percent reported improved clarity in their manuscripts – figures that originate from Google.
All of these figures come from Google and Google's own manuscript about the Paper Assistant Tool. The methodology rests on authors themselves reporting how long a fix took; no independent verification of the errors themselves is described. The difference between STOC and ICML may nonetheless be real and interesting: the more computationally and experimentally heavy the paper, the more errors the figures suggest an agentic system can catch – but that is an interpretation, not a documented finding.
The context: strained conferences and integrity problems
Google frames the launch with a growth curve for the conferences: more than 74,000 submissions to ICLR, NeurIPS and ICML this year, up from roughly 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023 – according to Google itself, with no independent audit in the available evidence base.
The broader integrity backdrop is heavier. The secondary report points to 21 percent AI-generated peer reviews at ICLR 2026, and at ICML 2026 that 398 reviewers were caught using large language models to write reviews, with 497 rejected papers as a consequence. The methodology behind these detection figures is not described in the available material. It is worth noting that Deep Review has not been part of the official peer review in any of the pilots, and nothing in the sources suggests that the conferences themselves have adopted the system.
How to get access
According to Google, all users get two free document reviews per week. Users with large volumes can connect their own Gemini API key for higher limits, and the system is available programmatically via a dedicated Deep Review API. At launch, PDF and plain text are supported.
Google further states that uploaded documents are not used to train the company's models, and that they are stored for one year. These are the company's own privacy terms, not independently verified facts.
Open questions
Several key points remain unresolved. The description of the system's internal pipeline – referred to as a four-stage agent architecture – is only partially verifiable from the primary source, which breaks off mid-review; segmentation and adaptive budgeting are stages that can be confirmed, the rest cannot. There is no independent evaluation of how accurate Deep Review is, whether on academic texts or the new legal and finance domains. Pricing beyond the free tier has not been stated. And the question of whether the conferences themselves will adopt such systems in official review – or whether authors will simply use them privately – remains open.
What does remain after a year of pilots is a concrete shift: verification tools that were previously available only to a small group of conference authors are now generally available – with all the methodological caveats that come with the figures produced so far.
Sources: Google Research blog (37bbb300), secondary coverage of Research@ 2026 (9d286939)