METR study finds AI tools made developers 19% slower

The author argues that standard metrics for AI-generated code—such as lines written, acceptance rate, and typing speed—fail to measure long-term code quality. Studies show developers often overestimate AI's productivity benefits while writing less secure code and shipping more defects. To accurately gauge AI's impact, the author recommends tracking delayed metrics like 30-day rework rates and post-release defects. Open article →