Models are worse at reviewing their own code

TLDR · 5 hours ago

Greptile researchers tested whether AI code review models catch bugs in code written by other models versus their own code, finding both Claude and GPT detect more bugs in code written by the competing model. Testing on 1,000 PRs with 1,500 verified bugs showed Claude Opus achieved 39% recall on GPT-authored code but only 28% on Claude-authored code, while GPT showed the opposite pattern. The team launched Model Inversion, a feature that routes code reviews to the opposite model based on author detection, leveraging the finding that cross-model reviews outperform same-model reviews.