Did generative AI actually boost productivity growth?

Pooling their February, May and August 2025 RPS survey waves, Bick, Blandin and Deming find that, "[f]actoring in all workers, including nonusers, workers report generative AI time savings equivalent to 1.6% of all work hours." Feeding this into "a standard aggregate production model," they estimate that "generative AI may have increased labor productivity by up to 1.3% since the introduction of ChatGPT."

They tie this estimate to the actual, realized acceleration in US nonfarm business productivity: "productivity increased at an average rate of 1.43% per year from 2015-2019, before the COVID-19 pandemic," versus "2.16% on an annualized basis" from the fourth quarter of 2022 through the second quarter of 2025 — a gap they translate into "excess cumulative productivity growth of 1.89 percentage points since ChatGPT was publicly released" relative to the prepandemic trend. They call their 1.3% survey-based estimate "consistent with" this realized figure, while cautioning the comparison could run either direction: workers banking time savings as "on job leisure" rather than output would make the survey-based estimate overstate what shows up in the data, while unmeasured "firm-side adjustments to capital or work reorganizations" could make it understate it.

It sits beside What is generative AI's actual aggregate productivity gain today?, a note on the same authors' earlier work; this post sets its "up to 1.3%" estimate against realized productivity data the authors read as consistent with that gain, checked loosely against the aggregate record. It also adds an industry-level test: detrended productivity growth correlates with industry time savings at 0.32, with "1 percentage point higher time savings" associated with "2.7 percentage points higher productivity growth relative to their prepandemic trend."

The authors themselves flag the limits of this evidence: "we stress that this correlation cannot be interpreted as causal, and that labor productivity is determined by many factors." The consistency between the survey-based 1.3% estimate and the realized 1.89-point productivity acceleration is also just that — consistency, not attribution; other forces could equally be driving the realized acceleration. The gap between worker-reported time savings and visible returns is also visible in Why do most enterprise AI pilots fail to deliver returns?.