Event
Wie misst man die Wirkung von KI beim Programmieren?
Following its early-2025 study, METR again examined completion times for programming tasks with and without AI allowed, beginning in August. On 24 February 2026 the team reported data from 57 developers, 143 repositories and over 800 tasks. After participants selected tasks, each task was randomly assigned to a condition. Research update
METR itself considers the results an unreliable measure of the actual effect: some people or tasks were missing because participants did not want to work without AI; parallel agents also complicated time measurement. The team believed time savings were likely greater in early 2026 than in early 2025, but described the data as very weak evidence of their magnitude and planned changes to the study design. The follow-up therefore does not establish a general productivity figure. Research update
Sources
Offen
Spätere Ergebnisse mit geändertem Studiendesign und unabhängig untersuchte Arbeitsumgebungen prüfen. / Check later results with revised study design and independently studied work settings.