Inside this story Context & sources

This article links to 2 sources. Source links are provided below so you can inspect the evidence yourself.

  • Google publicly released Deep Review on 14 October.
  • Initial inputs are PDF and plain text; vendor descriptions are not independent accuracy tests.
  • Its PAT conference history is context. Authors still need to check individual suggestions.

A second reader for a technical draft

Google publicly released Deep Review on 14 October, extending its work on AI feedback for technical documents. The company says the system can examine reasoning, check mathematical statements and suggest improvements. At launch, the stated input formats are PDF and plain text, with two complimentary document reviews per user each week. [1]

The useful question is what happens after the feedback arrives. Our view is that a review tool earns its place by making a draft easier to question. A confident report is a starting point for checking an argument, rather than a reason to stop reading it closely.

From conference experiment to public tool

Google describes Deep Review as an evolution of its Paper Assistant Tool, or PAT. There is earlier context on the International Conference on Machine Learning’s blog: a 30 March retrospective describes a PAT programme that ran from 14 to 26 January. It says feedback was kept separate from official peer review and was visible only to the authors. [1] [2]

That retrospective was jointly written by Google researchers and the conference’s programme chairs. It documents the partnership, rather than independently testing the newly released Deep Review. We have not tested the public tool either. The January programme and the October release are different stages of the project, not two new announcements this week. [2]

Turn a comment into a check

Google says Deep Review divides documents into sections, reviews those sections and combines the findings. It also says higher-level judgement remains with people. Those are descriptions of the intended process; they do not establish that any particular criticism or proposed correction will be right. [1]

For a low-stakes trial, we would start with a short, synthetic technical note: a description of a simple app feature, its assumptions and an example. Keep the original beside any suggested changes. For each comment, identify the passage it refers to, the evidence it needs and whether the proposed edit actually follows from that evidence. This is a suggested exercise, not one we have performed.

Keep the author in the loop

For a South African student or small software team considering the idea, our practical test would be a clearer document rather than a longer review report. Did a vague assumption become explicit? Did an example begin to match the explanation? Can another person follow the revised argument without seeing the AI’s comments? These are editorial questions, not findings about local adoption.

The distinction worth carrying forward is between useful criticism and verified truth. A suggestion may point to something worth investigating; the author still needs to resolve it. Deep Review’s public launch makes that drafting approach available to explore, while the quality of a finished document remains something its readers should be able to inspect for themselves.

Sources & further reading

1. Google Research: analysing technical documents with Deep Review — 14 October

Public-launch original; product scope, initial PDF/plain-text formats and two weekly complimentary reviews are vendor statements. No independent accuracy test, native Docs integration or performance guarantee inferred.

2. ICML: retrospective on its PAT paper-assistant programme — 30 March

Jointly authored by Google researchers and ICML programme chairs. Describes the earlier 14–26 January PAT programme and separation from official peer review. Another domain, but a partnership account rather than independent validation of Deep Review.

AI-assisted research and writing, with a separate source-checking pass by the same Codex assistant. Google supplies the public-launch claims; the older ICML post is jointly authored partnership context, not independent validation. We have not tested Deep Review or performed the suggested synthetic-document exercise. No local adoption or accuracy guarantee is claimed.

THE OPEN TABStay curious. Keep reading.