AI coding tools made developers 19% slower (METR study) (tolearn.blog)

A randomized controlled trial by METR with experienced open-source developers found that popular AI coding assistants (examples: Claude 3.5, Cursor Pro) increased task completion time by 19% even though developers felt 20–30% faster. That perception gap—driven by what researchers call the “activity illusion,” cognitive offloading, and immediate-gratification rewards from instant suggestions—means AI can create the feeling of progress without reducing real work time. The study highlights concrete friction points: frequent context switches to read and validate suggestions, review overhead to fix subtle or unsafe outputs, and time spent reverting or explaining AI-generated changes to teammates. For the AI/ML community this is a wake-up call about evaluation and integration: model-driven speedups must be validated with measured productivity metrics, not just subjective impressions. Practically, the study shows assistants are useful for boilerplate, syntax lookups, code translation and docs, but weak on architecture decisions, novel problem solving, security- or performance-critical code, and cross-file refactors. METR recommends measuring real outcomes, using AI selectively, setting firm boundaries (scaffold, not core logic), mandatory human review for critical systems, and hybrid workflows that reserve human attention for architecture and AI for repetitive tasks. The results suggest the 19% slowdown reflects early-adoption friction and immature tooling—areas where better models, UX and team practices could flip the balance.