"Why I Think AI Take-Off Is Relatively Slow" is an essay published on 2025-02-23 by economist Tyler Cowen (George Mason University; Marginal Revolution). It argues that AI is real and transformative but that its macroeconomic impact will be gradual, with human and institutional bottlenecks binding faster than AI capabilities advance. Cowen's bottom-line estimate is that AI boosts economic growth by roughly 0.5 percentage points annually — by his account transformative over decades but imperceptible year-to-year.
Summary of argument
Cowen accepts that AI capabilities are advancing rapidly while arguing that aggregate measured growth responds slowly. The essay assembles nine arguments for why capability gains do not translate quickly into measured output, centered on the claim that non-AI constraints, sector composition, and adoption friction govern the pace.
Key claims
The essay's arguments include:
- Cost disease shifts sector share. Citing Baumol's cost disease, Cowen argues that less-productive sectors grow as a share of the economy, so AI-accelerable sectors shrink relative to AI-resistant sectors, dragging the aggregate measured impact.
- Human bottlenecks bind. Even when AI capabilities advance, non-AI constraints become binding — for example, FDA review timelines even if drug discovery accelerates.
- O-ring effects. In complex systems, following Kremer's O-ring theory, the weakest performer sets productivity, and humans become the weak link.
- Traditional growth models underweight integration friction. Cowen argues that Solow and Romer frameworks do not handle the case where, in his phrasing, "Star Trek technology fell into our backyard."
- Historical diffusion was slow. Electricity took roughly 40 years to fully integrate.
- GDP stability. Annual growth hovers near 2% despite large technological changes, which he reads as evidence of structural limits.
Arguments 7 through 9 add further institutional points about complementary inputs and adoption friction.
Relation to other positions
Cowen's quantitative estimate aligns with Daron Acemoglu's in The Simple Macroeconomics of AI (≤0.66% TFP gain over 10 years). The essay sits at the opposite end of the timeline debate from Machines of Loving Grace, which posits a compressed 21st century, and from The Intelligence Age (Altman). It is among the more prominent public statements of the view that AI is real but economically slow.
Relationships
- supports: AI as Normal Technology on the slow-diffusion premise.
- contradicts: The Intelligence Age (Altman), Abundant Intelligence (Altman), Machines of Loving Grace on timelines.
- depends-on: AI Diffusion, Enterprise AI Deployment Gap.
- related: MIT NANDA — The GenAI Divide (State of AI in Business 2025), METR — Measuring Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.