Ankit Jain says AI code review is looking at the wrong artifact, and points at Faros AI numbers showing incidents per pull request up 242.7 percent

Getting somebody to read a diff carefully has become the least reliable step in shipping AI-written code, and Aviator's Ankit Jain has an argument in The New Stack on Tuesday about why an AI reviewer does not fix it. The essay is a sponsored post the site labels as such, and Jain cites numbers from parties with no stake in his product.

The core claim rests on two studies. In 2013, Alberto Bacchelli and Christian Bird classified 570 review comments at Microsoft: 44 percent of developers ranked finding defects as their top reason for reviewing, but only 14 percent of the comments were actually about defects. Caitlin Sadowski and colleagues reached the same conclusion after examining nine million reviewed changes at Google. Jain reads that as evidence that code review's main job has always been building a shared mental model of the system, with bug-catching a smaller part.

Faros AI and DORA numbers on what changed

Jain quotes the Faros AI 2026 engineering report, which covers 22,000 developers across more than 4,000 teams, as showing incidents per pull request up 242.7 percent, bugs per developer up 54 percent, and work restarts up 13.8 percent. Pull requests merged with no review at all, human or AI, rose 31.3 percent. Faros AI is a delivery-metrics vendor, so those figures are its own reading of its own data. Jain also cites the DORA 2025 State of AI-assisted Software Development report, which describes AI adoption as raising delivery throughput and instability at the same time.

Nothing here is Reveneau's own measurement. The point of the piece is that a reviewer who is too tired to read a 4,000-line diff and asks another model to skim it is running "review theater", as Jain calls it, because the human is the only party in the loop not doing what the loop is named after.

The alternative Jain describes

The proposed shape has five layers, and only one of them lives on the pull request. Two AI reviewers, on different models, argue before the pull request is opened, and a record of what they disagreed about is what a person reads at review time. Intent and acceptance criteria, written as the agent works and as the engineer corrects it, replace the up-front specification that goes stale by the second commit. A rolling list of patterns the team keeps correcting, harvested from past review comments, becomes machine-checked invariants. Then a scheduled team debate about the decisions the agents surfaced.

Peter Naur's 1985 essay "Programming as Theory Building" sits underneath the argument. Naur wrote that a program dies when the team holding its theory is gone. Jain adapts that: if reviewing code stops being where a team talks about the system, the shared understanding disappears without a single resignation, because nobody built it in the first place.

The practical piece a reader can pick up this week is the invariant list. Jain says a team's last 1,000 review comments already contain a running registry of what it keeps correcting, and turning the top 20 into automated checks moves work the human should not be doing off the pull request. That leaves the human to argue about intent, which is the part the model cannot do.

Source

• Primary: Kill the code review theater, keep the review, The New Stack, by Ankit Jain, 2026-09-30. Sponsored by Aviator.