AI writes code that breaks main half as often (mergify.com)

A recent analysis of over 200,000 pull requests (PRs) from Mergify's merge queue has revealed that code written with AI assistance breaks main branches approximately half as often compared to code authored without AI—1.9% versus 4.4%. This finding is significant as it directly counters concerns that AI-generated code may introduce subtle errors that could compromise code quality. The research indicates that the risk of a green PR breaking main increases dramatically with team size, highlighting a 16-fold rise in failure rates as engineering teams grow from two to over forty members. Additionally, the study underscores the potential benefits of merging queues and batching PRs, especially for larger teams. While the data shows that 94% of private merges occur one PR at a time, the automated testing that occurs in a merge queue can catch potential breaks before they disrupt development. The report suggests that adopting these strategies can alleviate the strain on teams, particularly in environments where dependency interrelations can lead to higher break rates. Overall, this research highlights the promising role of AI in software development, indicating that its assistance may enhance the integrity of merges, especially in complex team environments.