Skip to content

CS146S Week 6: To Make AI Review Useful, Google Deleted 17 Rules First

Aug 16, 2026 1 min
TL;DR Google deployed AutoCommenter to tens of thousands of engineers and published the whole tuning process: suppressing 17 'technically correct but low-value' rules raised the useful ratio from 54% to 66%, with 80% set as the bar for the next rollout stage. Final comment-resolution rate landed around 40%. The bottleneck in AI code review was never detection — it's volume.
Table of Contents
  1. Why the course treats review as high leverage
  2. The one public dataset with real scale
  3. What follows from those numbers
  4. The class it misses
  5. Three rules for the PR workflow
  6. What will go stale
  7. References

🌏 中文版

This is the seventh post in the CS146S series, covering Week 6 of Fall 2026.

Three topics: what AI review catches well and what it misses, review architectures and custom rules, and fitting AI review into a team's PR workflow. The guest is Cognition's Silas Alberti — who also appeared in Fall 2025, then on AI IDEs, now on code review.

That shift is telling. A year ago the talk worth booking was about how to write. Now it's about who checks.

Why the course treats review as high leverage

The matching Fall 2025 session was Week 7, "AI code review" (slides, with Graphite CPO Tomas Reimers as guest). It opens with three sets of numbers, sourced to Coding Horror — itself an assigned reading for the course:

  • Code review has a 55–60% error detection rate, versus 25–45% for various testing modes
  • One study compared defect density without and with review: 4.5 → 0.82 errors per 100 lines
  • An AT&T study found review brought a 14% productivity increase and a 90% decrease in defects

All three are numbers for human review, nothing to do with AI. The course uses them to establish a premise: review itself is enormously valuable, so the question is never whether to review, but who does it and how well.

The course sorts what review should catch into five categories: logic and correctness, readability and maintainability, performance, security, and best practices — including a codebase's own idioms and database access patterns (the slide's example: "use this service rather than a direct DB lookup").

It also gives four conditions for a good review comment: provide specific details, reference specific code or issues, suggest a resolution, and cite evidence or provide an explanation. The counterexample is a comment that says only "This won't work." Those same four work as a rubric for judging an AI reviewer's output.

The one public dataset with real scale

There is no shortage of numbers about AI code review, but nearly all of them come from tool vendors or affiliate-marketing sites with no reproducible methodology. The public data actually worth discussing is Google's AI-Assisted Assessment of Coding Practices in Modern Code Review (AIware '24, assigned in Fall 2025 Week 7) — they deployed an LLM system called AutoCommenter to "tens of thousands of developers" and wrote down every pothole.

All figures below come from that paper.

Comment volume was deliberately suppressed. The system first filters comments landing on unchanged lines, dropping the rate of comments in changed files to 1.3%; switching to beam search (n=4) "tripled the posting frequency to 3.9%." Diversity improved too: the ten most frequent rules fell from 80% of all comments to 41%.

Then they deleted rules. This is the passage worth copying:

First, the rater study identified 17 non-actionable URLs, whose suppression increased the historical useful ratio from 54% to 66% on developer feedback, and from 60% to 74% on rater feedback.

Suppressing five more got them to "our target useful ratio of 80% for the next stage of deployment." In other words — they made an 80% useful ratio the gate for wider rollout, and the way they hit it was turning features off, not adding them.

Their explanation is clear:

Correct but low-value comments: A missing period at the end of a sentence in a code comment is often allowed by human reviewers. While technically correct, asking the author to go back to their IDE and fix the issue may provide net negative value.

Technically correct, net negative. That sentence belongs on the wall of every team about to adopt AI review.

The final numbers are honest too. Using 6,000 snapshot pairs they estimate a comment-resolution rate around 40%. The system covers 330 distinct rules and "covers 68% of historical human comments with a best practice URL" — two-thirds of the best practices human reviewers actually cite, and "many of these are out of scope for traditional static analyses."

The rollout pacing matters as well: an A/B experiment with half of all developers in July 2023, and only after confirming no adverse effects, full deployment in October 2023.

What follows from those numbers

One: the bottleneck is signal-to-noise, not capability. Making the model find more is easy. Making it shut up is hard. A reviewer that's useful 60% of the time gets skipped within two weeks — and the real bugs it finds get skipped along with it.

Two: the rule set must be individually disableable. Google could fix this because they could identify which 17 rules were generating noise and turn those off. If your AI reviewer is a black box with no concept of rules, your only options are all-on and all-off.

Three: AI review complements static analysis, it doesn't replace it. That 68% figure describes coverage of things humans cite that traditional static analysis cannot catch. Anything a linter catches shouldn't be handed to a model — that's the job of Week 5's deterministic validation loops, which are faster, cheaper, and don't drift.

The class it misses

The syllabus explicitly says "what AI review catches well, and what it misses." The Fall 2025 slides give a limitations list that is sharper than the boundary I would have drawn myself:

  • More configuration/setup
  • False positives — "Have to train the system → continuous learning"
  • Can't yet catch the idioms and repo best practices
  • Can't handle complex business logic and architecture decisions — "But that's where humans are still needed"
  • Must be extra cautious with security changes
  • Often misses edge cases

"Can't catch the idioms and repo best practices" is the one to sit with, because it directly collides with the fifth of the course's own five categories — internal conventions and best practices are exactly where a human reviewer adds the most, and exactly where AI is currently weakest.

The course adds two operational rules: explicitly tell it what not to review, and be especially careful with changes touching user input, authentication, file operations, and network requests.

The line between diff-only tools and tools that index the whole codebase falls here too — the former can't see how a change ripples through the rest of the system.

One caution: the "bug catch rate" comparisons circulating online (82% for this tool, 44% for that one) come almost entirely from tool-comparison and affiliate content with no published test set or reproducible method, so this post doesn't cite them. If you're picking a tool, running it over your own last twenty PRs beats any leaderboard.

Three rules for the PR workflow

One: the instance that wrote the code doesn't review it. This appeared in Week 2's RePPIT and is worth repeating: a model defends its initial implementation, "like proofreading your own writing and reading what you meant to type." Switch model families, or at minimum wipe the context.

Two: rank findings, don't dump them. RePPIT's Test step sorts findings into must-fix, should-fix, and nice-to-have. That ranking is the cheapest available defense against Google's "correct but low-value" problem — it doesn't delete comments, it tells people which ones they can skip.

Three: a passing AI review is not permission to merge. The biggest risk of a green checkmark isn't the bugs it missed; it's that human reviewers relax. Google's system ran for over two years and still lands around a 40% resolution rate.

The course puts it harder than I would, on the deck's final line:

Code review is more important now than ever with AI coding systems — You own the code that is merged and shipped, no blaming of the AI

That single sentence answers both accountability and process design.

What will go stale

  • AutoCommenter's numbers come from a 2023–2024 deployment and several model generations ago; they describe the shape of the problem, not today's absolute performance
  • Fall 2026's materials and assignment for this week aren't published
  • Tool capability boundaries move fast; the table above needs periodic re-testing

References