Study of 5,388 repos finds AI-authored code merges faster at median, but 25% pay real review cost

A scan of 5,388 public repositories covering 444,225 merged pull requests found that the median repo experiences no extra review burden from AI-attributed code, with attributed work merging in roughly half the time of non-attributed work. However, the top quartile of repositories did show a measurable review cost, with time-to-merge running at least 1.18 times longer for AI-attributed pull requests. Attribution was determined strictly from commit-level markers such as co-author tags and agent bot accounts, not inferred from code style or timing. The researcher behind the study disclosed a commercial interest in the scanning tool used, while noting the findings actually work against a simpler sales narrative. Key caveats include that inline AI completions leave no repository trace, meaning detected figures represent a floor, and causation between agent use and review time cannot be established.