AI as a General Purpose Technology

AI is widely argued to be the most consequential general purpose technology (GPT) in history — one that may automate not just physical or narrow cognitive tasks, but intelligence itself.

What It Is

Economic historians classify technologies as “general purpose” when their transformative effects extend throughout the economy — enabling innovations in many sectors rather than just one. Electricity, the steam engine, semiconductors, and the internet are canonical examples. AI fits this mould, but with a potentially critical difference: previous GPTs replaced humans in physical tasks or narrow cognitive tasks, making human intelligence more valuable in complementary roles. AI threatens to automate the cognitive residual itself — any task a human can do on a computer.

Charles Jones (Stanford GSB, Jan 2026) presents two extreme scenarios: (1) AI accelerates economic growth dramatically, potentially reaching “a country of geniuses in a data center” (Amodei’s phrase) that raises R&D productivity across all fields; (2) AI is “business as usual,” simply being the latest GPT that sustains the US economy’s long-run 2% per year growth trend rather than raising it. Jones argues both scenarios are plausible, and the future lies somewhere between them.

Why It Matters (for the Economy)

The GPT framing carries two important lessons from economic history that temper both excessive optimism and pessimism. First, past GPTs did raise growth rates — but primarily by preventing growth from slowing, not by creating hockey-stick acceleration. Second, diffusion takes decades: the productivity payoff from electricity took 40 years to show up in factory output data. The “AI productivity paradox” — strong AI capabilities but muted macroeconomic impact so far — may simply reflect normal diffusion lag rather than a fundamental flaw. If so, the meaningful question is not whether AI matters, but when its impact will become visible in GDP statistics.

The most radical pathway involves recursive self-improvement: AI models raising the productivity of AI researchers, who build better models, creating a feedback loop that could compress decades of innovation into years. Jones notes the METR time horizon measure (how long it takes AI to complete engineering tasks with 50% success) is doubling every 5–7 months.

Evidence & Examples

  • US GDP per capita has grown at ~2% per year for 150 years despite transformative innovations including electricity, semiconductors, and the internet — suggesting GPTs sustain growth rather than accelerate it (Jones 2026, AIandEconomicFuture.pdf)
  • Epoch AI estimates “effective compute” used to train AI models is rising 10x annually (4x from better chips, 2.5x from better algorithms), with rapid capability improvements as result (AIandEconomicFuture.pdf)
  • Claude Opus 4.5 scored higher than any human candidate ever on Anthropic’s two-hour software engineering take-home exam; the METR time horizon for 50% task success has grown from 19 minutes (18 months ago) to ~5 hours (AIandEconomicFuture.pdf)
  • AlphaFold solved the protein folding problem — demonstrating AI already contributes to scientific discovery at a Nobel-Prize level (AIandEconomicFuture.pdf)
  • David (1990) showed the steam engine and electric motor both took 40+ years for their productivity impact to appear in statistics — supporting diffusion-lag interpretation of current AI productivity data (AIandEconomicFuture.pdf)

Tensions & Open Questions

  • The weak-links problem: Jones uses a “task complementarity” framework — even if AI automates most tasks, the few remaining human-required tasks become bottlenecks. This could limit explosive growth even with very capable AI.
  • Distribution vs. aggregate: Even if GDP grows, the distributional impact could be severe — capital capturing most of the gains while labour loses bargaining power (see AI Labor Displacement and Augmentation).
  • Is this time actually different? Solow’s 1987 productivity paradox observation (“you can see the computer age everywhere but in the productivity statistics”) proved temporary. Whether AI’s current paradox will resolve similarly or persist is genuinely uncertain.
  • Existential risk as tail event: Jones acknowledges scenarios where sufficiently advanced AI poses risks that economics frameworks cannot adequately model — a consideration absent from GPT analyses of electricity or the internet.

AI Productivity Paradox · AI Labor Displacement and Augmentation · Recursive Self-Improvement and AI R&D · AI Capex Economics and Bubble Risk