A new survey from METR published this week offers one of the more detailed looks at how technical workers perceive AI tools affecting their productivity—and the numbers are striking, if you trust them. Researchers surveyed 349 software engineers, researchers, academics, and founders between February and April 2026, asking participants to estimate how much more value they produce when AI tools are in the mix versus when they're not.

The Core Numbers

The median respondent reported a 1.4x to 2x increase in the value of work produced due to AI access around March 2026. That's a meaningful jump from the retrospective estimate of roughly 1.3x for March 2025, and respondents are projecting 2.5x by March 2027. The survey also asked about raw speed—how long tasks would take without AI—and found a median self-reported speed change of 3x, which the researchers expected to be higher than value estimates due to people gravitating toward lower-value but easier-to-automate tasks.

Why 'Value' Differs From 'Speed'

METR made a deliberate methodological choice to distinguish between value and speed uplift. Speed measures capture how much faster someone completes tasks with AI; value captures whether those tasks actually matter to their team or organization. The distinction matters because AI might let you spin up an interactive dashboard in minutes, but if that dashboard wasn't strategically important, the value gain is minimal even if the speed gain looks impressive. This framework aligns more closely with what organizations care about—actual output quality and relevance—not just throughput.

Who's Overestimating?

The data reveals some eyebrow-raising patterns among subgroups. METR employees reported notably lower value gains than other respondent categories—a finding researchers attribute to their familiarity with prior work showing gaps between perceived and actual AI-driven productivity uplift. Meanwhile, heavy users of tools like Claude Code and Codex reported higher multipliers, as did startup workers. The survey also found that 48% of respondents are US-based, averaging 12 years of programming experience but only 19 months using AI for coding tasks. Approximately 70% of participants were paid to take the survey, with average compensation around $200 per respondent.

Caveats and Consistency Checks

METR filtered out just 10 anomalous responses from a raw pool of 359 (about 3%), noting that correlations between different value change measures were moderate—suggesting broadly consistent answers. However, researchers openly acknowledged red flags: they reviewed seven respondents claiming at least 10x value uplift and found that in two cases where public work outputs could be examined, the participants appeared to be overstating gains substantially. A prior METR study from early 2025 found people overestimate AI's effect on time spent by roughly 40 percentage points on average—a pattern that complicates interpreting these newer results.

What This Means for Builders

For developers and technical leaders trying to calibrate expectations around AI tooling, the takeaway is nuanced. The survey suggests real productivity gains are happening—probably somewhere in that 1.5x to 2x value range for most teams—but headline figures likely overstate impact when you account for task substitution and measurement bias. METR recommends that organizations tracking AI R&D acceleration should consider surveying managers or dedicated productivity researchers rather than relying solely on individual contributor self-reports, since those groups may give more grounded estimates about actual organizational output.

Key Takeaways

  • Median self-reported value uplift sits at 1.4x to 2x for March 2026, up from 1.3x a year prior
  • Speed gains (3x median) consistently exceed value gains, suggesting task substitution toward lower-priority work
  • METR staff report the lowest multipliers—researchers suggest this reflects awareness of survey overestimation patterns
  • Survey data show internal consistency but qualitative checks on extreme responses raise concerns about magnitude accuracy

The Bottom Line

The numbers coming out of early 2026 look good for AI tooling advocates, maybe too good. Until someone runs controlled experiments at scale that match these self-reports against actual output metrics, treat anything above 2x value uplift with serious skepticism—and definitely don't use a speed measure to justify your next AI budget request.