Half of AI-written code that passes industry test would get rejected by real developers, new study finds

The Decoder
AI Research

About half of the AI code solutions that pass the popular SWE-bench benchmark would get rejected by actual project maintainers, according to a new study by the research organization