Code is unreadable at scale, and the responsibilities of code review remain

In teams with high AI adoption, code review time increased by 441.5% and PRs merged without review increased by 31.3%. This is a structural bottleneck caused by the inability of human verification and approval speeds to keep up with code generation speeds.

Existing code review has served five roles beyond simple bug finding: verification, maintainability checks, knowledge sharing, gatekeeping, and responsibility distribution. In the AI era, these roles must be redistributed between humans and AI.

Analysis of 300 open-source projects revealed that AI agents initiate 55.7% of reviews, with AI focusing on defect detection while humans handle implementation intent and context. The adoption rate of AI-generated fixes is 16.6%, lower than that of humans (56.5%), suggesting AI acts more as a sensor that broadly detects potential defects.

Analysis of 1.02 million PRs showed that applying AI reviewers in bulk may actually degrade review quality. AI adoption is not merely a speed improvement, but a fundamental shift in development processes regarding how to redesign the structure of verification and responsibility.