A new survey from METR, a leading AI evaluation group, asked 349 tech workers to separate AI's speed boost from its actual value gain. The result? Respondents report that AI makes them three times faster, but the median increase in work value is just 1.4x to 2x.
Conducted from February to April 2026, the survey covered software engineers, academics, and founders. On average, respondents had 12 years of coding experience, and half regularly use Claude Code. Despite the modest value gains, they're optimistic about AI's long-term potential — they estimated a 1.3x value bump in March 2025, now see 2x in March 2026, and predict it'll hit 2.5x by March 2027. In fact, the median respondent would give up 29% of their salary to keep their AI tools for a month.
But METR's own team is skeptical. Among all groups surveyed, METR's internal employees gave the lowest value estimates. When researchers examined seven extreme cases where respondents claimed more than 10x value gains, only two had verifiable public outputs — and neither backed up the hype.
METR attributes this bias to a substitution effect: with AI, developers tend to spend time on tasks they can finish quickly — even if those tasks are low-value. For companies evaluating AI investments, the lesson is to look beyond time saved and focus on actual progress in core R&D.