AI Code Review Data: Defect Rate Up 54%, Zero-Review Merges Up 31%
Summary
Major data-backed piece. Faros AI study of 22,000 developers: AI-generated code churn up 861%, defect rate rose from 9% to 54%, zero-review merges up 31.3%. GitClear: 4x code output = only ~12% productivity gain. CodeRabbit study: AI PRs have 1.7x more issues than human PRs. Positive: Anthropic internal, substantive PR review rose from 16% to 54% after investing in review culture. Key prescriptions: (1) capture agent reasoning as a decision log; (2) use heterogeneous reviewer tools — 93.4% of findings caught by only one tool; (3) human owns the merge, always.
Why this signal
The single most data-dense piece on Neo's #5 focus area. Every stat is attributable and usable in a team discussion. The 9%→54% defect rate is striking. The 'heterogeneous reviewers' finding suggests a simple, actionable toolchain diversification strategy Thomas Engineering could run cheaply.
Original email
Addy Osmani Elevate · 6/16/2026
Agentic Code Review