The O-ring effect suggests that sophisticated AI systems may be constrained by rare but catastrophic failures rather than average performance levels. Harford illustrates this with vivid hypotheticals: a robot masseur that occasionally snaps necks, self-driving cars that rarely mistake pedestrians for trash, or generative AI that produces career-threatening errors after weeks of flawless operation. Unlike traditional productivity improvements where small error rates can be acceptable, these examples involve high-stakes applications where even infrequent failures create unacceptable risks. The Challenger space shuttle analogy is particularly apt—the shuttle represented cutting-edge 1980s technology, yet a simple rubber O-ring’s failure at low temperatures destroyed the entire system and killed the crew. Similarly, AI might achieve superhuman performance on most tasks while remaining unreliable in critical edge cases, preventing deployment in contexts where consistent reliability matters more than average capability. This creates a fundamental barrier: until AI systems solve their weakest-link problems, they cannot be trusted with applications that would drive transformative growth.

The Baumol Effect explains this paradox through spending composition shifts rather than productivity stagnation. When computation becomes dramatically cheaper and more powerful, the economic value generated per computational operation declines even as capability expands—computers become so productive that computational services approach zero marginal cost. Meanwhile, human time and attention remain scarce, making labor-intensive services like healthcare consultations, restaurant meals, childcare, and education relatively more expensive. As people become wealthier through computational productivity gains, they spend growing portions of income on these inherently low-productivity services. Therefore, even as computational capacity doubles repeatedly, aggregate GDP growth reflects the weighted average of high-productivity sectors (whose prices fall and economic share shrinks) and low-productivity sectors (whose share grows). Agricultural productivity provides the historical template: despite revolutionary mechanization and bioengineering, agriculture now represents a tiny fraction of developed economy GDP precisely because its productivity success made it economically marginal. Computational productivity faces the same fate—success paradoxically reduces economic significance.

The 1960s featured multiple self-reinforcing factors that mirror current AI enthusiasm almost exactly. Peak population growth rates meant more potential geniuses generating ideas with spillover benefits to all humanity—analogous to arguments that AI will multiply effective researcher counts. Expanding education created dramatic increases in trained intellectual capacity with potential positive feedback loops—similar to claims about AI augmenting human cognition. Computers were lowering calculation costs and being used to design better computers—precisely the recursive improvement dynamic that optimists expect from AI training AI. The internet’s emergence, sophisticated finance, cheap travel, and extensive libraries all represented infrastructure supporting innovation acceleration. Yet despite this convergence of favorable factors that made growth doubling seem plausible, US growth slumped post-1970 rather than accelerated. This historical parallel suggests that even when conditions appear optimal for explosive growth, structural economic constraints (like those the Baumol Effect describes) prevent technological capability from translating into aggregate prosperity. The comparison warns against assuming current AI advances will prove different.

Readlite provides curated articles with comprehensive analysis including summaries, key points, vocabulary building, and practice questions across 9 different RC question types. Our Ultimate Reading Course offers 365 articles with 2,400+ questions to systematically improve your reading comprehension skills.

This article is rated Advanced level, reflecting its sophisticated economic reasoning despite conversational tone. Harford expects readers to understand exponential growth mathematics (doubling times, compound rates), follow counterfactual historical arguments (what should have happened in the 1960s versus what did), and grasp abstract theoretical concepts like the O-ring effect and Baumol Effect without extensive explanation. The text requires synthesizing multiple threads—growth predictions, historical precedent, theoretical constraints—to understand why optimistic scenarios face structural barriers. Advanced readers must recognize that Harford’s accessibility serves rhetorical purposes rather than simplifying actual complexity; his breezy tone makes sophisticated skepticism more persuasive by avoiding academic pomposity. The essay demands ability to evaluate probabilistic claims about uncertain futures, distinguish between technological capability and economic impact, and understand how sectoral composition affects aggregate statistics. This difficulty level suits readers preparing for graduate economics study or seeking nuanced perspectives on technology-driven growth debates.

This closing observation reframes the entire discussion by questioning whether contemporary stagnation represents temporary policy failure or structural inevitability. Throughout most of the essay, Harford argues that 20 percent growth is implausible because technological miracles don’t automatically become aggregate prosperity. But the final turn goes further: even the modest exponential growth that characterized the 20th century may prove unsustainable long-term once low-hanging fruit (catch-up growth, one-time resource exploitation, demographic dividends) has been exhausted. The phrase ‘exponential growth nonetheless’ emphasizes that 1 percent compounded annually still represents doubling every 70 years—a pattern that cannot continue indefinitely on a finite planet with physical constraints. This suggests the productivity slowdown since 1970 might reflect approaching fundamental limits rather than correctable mistakes. The implication challenges both AI optimists (who assume technology will restore high growth) and conventional economists (who assume 2-3 percent growth is normal and achievable with correct policies), proposing instead that sustaining any positive growth rate may become progressively harder as economies mature and easy improvements become scarce.

The Ultimate Reading Course covers 9 RC question types: Multiple Choice, True/False, Multi-Statement T/F, Text Highlight, Fill in the Blanks, Matching, Sequencing, Error Spotting, and Short Answer. This comprehensive coverage prepares you for any reading comprehension format you might encounter.