The 47% AI Job-Loss Number Was Never a Forecast — Real Risk Is 9%

Only 9% of jobs in developed economies face a genuine risk of full replacement by machines — not the widely cited 47%, according to OECD researchers who studied the task structure of thousands of workers across 21 countries, with national figures ranging from 6% in South Korea to 12% in Austria.

The 47% figure traces back to a 2013 University of Oxford paper, whose own authors stressed it measured the share of U.S. work tasks computers could technically perform — not a layoff forecast. That distinction has been lost in a decade of alarmism, most recently in Igor Lipsitz's new book predicting a global labor-market catastrophe driven by AI, a view the author himself concedes hinges on productivity gains failing to offset job losses.

What does history actually show?

The record argues otherwise. In 1900, 41% of U.S. workers were employed in agriculture; by 2000 the share was 2%. Mechanization displaced them without producing mass unemployment — workers moved into expanding factories and services while food prices fell. The shipping container cut the cost of loading one metric ton of cargo from $5.86 to $0.16 in the 1950s, eliminated dockwork almost entirely, and yet boosted world trade more than all postwar trade agreements combined.

Our World in Data shows global GDP per capita grew roughly 15-fold in the two centuries after 1820, surpassing $18,000, while in early industrializers — Britain, the U.S., Western Europe — it rose 25–30 times, from $2,000–$3,000 to $60,000–$75,000 today. Working hours nearly halved over the same arc; the 10–12 hour, six-day workweek of the 1870s gave way to standard 40-hour weeks, and as low as 35 hours in some countries.

Where AI is measurably working now

The hardest evidence sits at firm level. A Quarterly Journal of Economics study tracked 5,172 employees and 3 million customer-support conversations during a GPT-3-based assistant rollout; successfully resolved inquiries per hour rose 15%, with the tool effectively transferring expert phrasing to novices. In corporate-support software-assistant trials, inexperienced staff reached qualified-specialist levels in two months instead of six, with task resolution up 34%. PwC calculates that productivity growth runs 40% higher in the most AI-exposed firms, whose lead over laggards has tripled since 2022 — with wages and headcount also growing faster.

The macro picture is cooler. A National Bureau of Economic Research study of Denmark's employment register found no statistically significant effect on earnings or hours in ChatGPT's first two years. Stanford's Digital Economy Lab recorded a 19% hiring gap by June 2026 for U.S. workers aged 22–25 in the most automation-exposed occupations — a descriptive, not causal, link, but a troubling career start for young graduates nonetheless.

Forecasts diverge sharply. The OECD's 2024 model projects an extra 0.4–0.9 percentage points of annual U.S. labor-productivity growth within a decade, or 0.25–0.6 points of TFP per year; Daron Acemoglu's baseline caps decade-long TFP growth at 0.66% — under 0.1% a year.

Who captures the gains?

Distribution, not destruction, is the real risk. Acemoglu and Pascual Restrepo's 2022 paper attributed 50–70% of U.S. wage-inequality growth from 1980 to 2016 to technological displacement. An OECD study found that across 24 developed countries from 1995 to 2014, productivity rose about 28% while median wages grew only 15% — a 13-point gap. An MIT study counters that roughly 60% of jobs in today's developed economies did not exist in 1940.

Competition shapes the split. William Nordhaus calculated in 2004 that innovators capture only about 2% of their technologies' total value; consumers take the remaining 98%. Open-source AI is already diluting concentration — Hugging Face downloads of open models hit 2 billion by March 2026, nearly sixfold the 340 million a year earlier, and capable models now run on a $1,500 laptop.

Demographics add urgency. The OECD's Employment Outlook 2025 reports declining working-age populations, with the EU's 20–64 cohort set to fall by more than 10 million by 2030. Denmark's flexicurity system, where about 20% of workers change profession or employer annually without lasting unemployment, offers one policy template: income insurance, retraining, relocation support.

Paul David's 1990 study of electrification supplies the timeline caution: three decades passed between reliable electric motors in the 1880s and the 5% annual productivity surge in U.S. factories — gains arrived only after owners rebuilt plants entirely. AI's workplace-level gains are already measured; the national-level translation, on every estimate cited, remains a question of years, institutions and competition policy rather than of the technology itself.