METR survey finds software engineers reporting ~2x AI speedup

METR published a survey of technical workers on their self-reported productivity gains from AI tools, finding a median 1.4–2x increase in the value of their work compared with roughly 1.3x when the same respondents retrospectively estimated their gains a year earlier. On a separate measure of raw speed rather than value, respondents reported a median 3x change in how quickly they completed tasks.

The survey drew 349 responses — including 87 software engineers, 71 researchers, 129 academics or PhD students and 48 founders or managers — recruited through GitHub, institutional directories and professional networks, with most respondents paid for participating and roughly a 2% response rate from email outreach, making it a convenience sample rather than a representative one.

METR itself flagged reasons for scepticism about the headline figures. It cited its own earlier research finding that developers, in retrospective self-report, had overestimated AI’s effect on their task completion time by roughly 40 percentage points compared with what controlled measurement showed, and noted more broadly that survey-based productivity estimates have consistently run higher than those from controlled experiments. The gap between what practitioners believe AI is doing for their output and what has been measured directly under controlled conditions remained unresolved by this survey, which measured belief rather than output.