AI productivity gains not showing up in economic data — a heated debate on the timeline
Aug 21, 2025
Key Points
- François Chollet argues that despite nearly 50% of US workers adopting large language models by mid-2025, labor productivity growth has declined versus 2020 levels, undercutting 2023 claims of 10x productivity gains.
- OpenAI's Noam Brown contests the framing, saying the 10x claim was never consensus in 2023 and the actual debate centered on whether LLMs could reason at all.
- The disagreement hinges on what was actually promised and when, leaving unresolved whether the productivity lag stems from the tools, the data, or measurement limitations.
Summary
François Chollet argues that large language model adoption has failed to produce measurable productivity gains. With LLM adoption among US workers approaching 50% as of mid-2025, labor productivity growth has actually declined compared to 2020 levels. This contradicts a 2023 consensus that workers would see 10x productivity improvements from the technology.
Noam Brown from OpenAI disputes the framing. He contends that the 10x productivity claim was never widespread in 2023. The actual conversation at that time centered on skepticism about whether LLMs could reason at all. The real question is not whether AI met an inflated promise but whether the technology has delivered against realistic expectations of capability.
The disagreement is factual. Chollet points to the Stanford/World Bank survey showing 45.9% of US workers 18+ using LLMs as of June/July 2025, paired with productivity growth that lags 2020, as evidence the hype has outpaced reality. Brown offers no specific counter-statistic or timeline but reframes what constitutes a reasonable expectation for an emerging tool.
Neither speaker resolves whether the lag between adoption and measurable productivity stems from the data, the tools, or the measurement itself.