Trust and provenance
Google opened an error-checker for technical documents. Peer review is just the first customer.
Deep Review stress-tests a document for logical gaps and math errors. The same problem shows up in every due-diligence memo and regulatory filing.

Efthymios Vogiatzis · · 2 min read
Google Research opened a tool to the public this week called Deep Review. It reads a manuscript or technical document and stress-tests it for logical gaps and mathematical errors.
It splits the document, spends more compute on the dense parts, runs parallel agents that each verify a section with the whole document in view, then grounds the findings against external search.
Google piloted it on more than 16,000 submissions across four computer-science conferences before opening it. Every user now gets two free document reviews a week.
It reads PDFs and plain text only, with multi-file support still coming. It can flag problems that aren't real and occasionally call a correct proof wrong, so the higher-level judgment stays with a human. It's a first reader, not a verdict.
Scientific peer review is the obvious first customer, and it's badly overloaded: those same conferences went from tens of thousands of submissions a few years ago to far more now, reviewed by the same volunteer pool.
But peer review is just the first buyer. Any field that lives or dies on whether a long technical document holds together, a due-diligence memo, a clinical protocol, a regulatory filing, a model's assumptions, has the same problem and no tool for it. Science is where this lands first. It isn't where it stops.
Sources
- Google Researchhttps://research.google/blog/analyzing-technical-documents-with-deep-review/
- GIGAZINEhttps://gigazine.net/gsc_news/en/20261015-google-deep-review/
- Tech Timeshttps://www.techtimes.com/articles/329051/20261015/google-deep-review-caught-errors-35-ml-papers-diagnostic-ai-amie-lands-lancet.htm

Efthymios Vogiatzis
Efthymios Vogiatzis is a pharmacist turned AI and data product leader, and the founder of VE1 Advisory in Vienna. He has led AI, data and analytics work across life sciences, biopharmaceuticals, insurance and industrial technology.
Start a conversationField notes by email.
A new essay about once a month on AI, data integrity and the people side of change.
Keep reading
21 October 2026 · Operations · 2 min read
The quiet half of biotech: scaling it next to a sugar mill
A Brevel and Wave Group plant jumps from a 5,000-litre tank to 300,000 litres, on a feedstock sitting next door. Scale-up is the half that decides whether biology reaches a shelf.
September 2026 · AI drug discovery · LinkedIn
What a pharmacist sees in the AI pipeline
The habits that keep data honest when the model moves faster than the audit trail.
