SAN FRANCISCO – A new study by the AI research group METR has revealed that advanced AI coding tools may actually slow down experienced software developers, especially when working on familiar projects.
The research involved developers using Cursor, a widely used AI coding assistant, to complete tasks within open-source projects they already knew well. The developers had initially expected a 24% decrease in task time by using AI.
However, the study found the opposite. Developers took 19% more time to finish their tasks when using AI tools. Interestingly, despite the slowdown, many still believed that AI had made them faster—estimating a 20% time reduction.
Lead researchers Joel Becker and Nate Rush expressed surprise at the results. Rush had expected noticeable productivity gains, but the data contradicted the widespread assumption that AI tools always speed up experienced coders.
One key reason for the delay was that developers spent time correcting AI-generated suggestions. While often close to the correct solution, the AI’s output required significant tweaking—leading to wasted time during familiar tasks.
Despite the findings, the researchers noted that AI tools could still benefit junior developers or those unfamiliar with complex codebases. Most participants, including the authors, said they would continue using AI assistants for reducing effort and improving workflow in other scenarios.