Google is moving a research-paper feedback experiment into a general document-review service that targets the verification bottleneck created by rising document volume. The announcement says the system checks logical consistency and mathematical statements, suggests improvements, grounds facts, and compares work with existing material; it explicitly leaves higher-level judgments to humans.
Observed facts: Google says the predecessor Paper Assistant Tool ran in opt-in pilots with four computer science conferences and reached authors of more than 16,000 manuscripts. Observed facts: the public service offers two free reviews weekly, accepts PDFs and plain text initially, and provides higher-volume API access using a Gemini API key. The product packages repeatable checks into an agentic review workflow, potentially letting experts spend less attention on cognitively demanding first-pass checks and more on interpretation and judgment. Google frames the launch against rising submissions, reporting more than 74,000 submissions to ICLR, NeurIPS, and ICML in 2026, compared with about 45,000 in 2025.
Teams may be able to add machine-generated critique earlier in drafting or review, but the tool should be treated as an assistant rather than an authority. Google says uploaded documents are not used to train or improve its models, yet are retained for one year for display in the UI; that matters for confidential work. These are claims and deployment figures in Google’s announcement, not an independent evaluation of accuracy, reviewer time saved, or reliability across legal and finance documents. Look for independent evaluations showing how often Deep Review catches genuine errors and how often it introduces false alarms across different document types.