AI code review agents like GitHub Copilot are fast and helpful, but they miss bugs outside the diff. In an experiment, Copilot approved a billing fix four times despite a guard that failed to catch negative enrollment days, leading to an 11-day overcharge. While Copilot caught 8 of 8 diff bugs with rules, it missed 2 bugs in unchanged code.
The issue highlights a context gap: AI reviewers lack domain knowledge not written in code. A correct diff can still be unsafe for the system. Billing bugs often go unnoticed until customers complain.
The study underscores the need for human oversight and comprehensive testing beyond code diffs.
Source: Towards Data Science · Summarized by HeadlinesBriefing