Did you know that one study found slower developers with the help of AI? In two thousand twenty-five, METR observes sixteen experienced developers on their projects. The result upsets one idea: having a code assistant doesn’t save everyone time. Tasks are about repositories that participants know, with context to understand. Reading, asking, correcting and checking can take time. Our office illustrates this control work, without inventing the code of a participant, or pretending to measure a duration in the film.
On two hundred and forty-six tasks, access to AI tools results in an average slowdown of nineteen percent. Keep the group and the context: this result does not prove that all developers slow down with each current assistant on any project.
More recent follow-ups discuss measurement difficulties, tool changes, and selection biases. The initial result remains dated. What is important is to measure the effects on the actual task, not to confuse an impression of speed with the time spent.
What we know
The initial 2025 result covers 16 developers and 246 tasks; the updates discuss measurement limits. It does not establish a universal slowdown with all AIs.
To go to the source
References consulted on October 12, 2026.
Signaler une erreur ou apporter une précision
