What Actually Happened to Displaced Workers
The case studies show what changed. This page shows who paid the price — and whether retraining, relocation, or institutional support made any difference.
1. Does Retraining Work? The evidence from 207 studies and 4 major programmes
The conventional policy response to technology displacement is retraining: teach workers new skills, and they'll find new jobs. The evidence says otherwise. Across four major evaluations spanning decades and millions of workers, the pattern is consistent: retraining shows modest effects at best.
The chart below shows the "leaky pipeline" from the US Trade Adjustment Assistance programme — the most comprehensive evaluation of displaced worker retraining ever conducted (Mathematica, 2012). Of every 100 eligible workers, only 18 ultimately found employment in their retrained field. And those who completed the programme earned ~€3,000 less than similar workers who received no training at all. European programmes show similar patterns: the UK Work Programme achieved sustained employment for only 3.6% of 1.81 million referrals — worse than the 5% baseline. Sweden's Job Security Councils perform better (~80% reemployment) but are funded by employers, not government — and cover only unionised sectors.
TAA Retraining Pipeline — Where Do the 100 Workers Go?
Info gap · already in lower-pay work
Financial · family · programme mismatch
Skills mismatch · geography · age
View data as table
The broader evidence confirms this. Card, Kluve, and Weber analysed 207 studies covering 857 programme evaluations worldwide. Their finding: short-run training impacts are "near zero." Modest positive effects appear only after 2+ years, and only for specific subgroups (adult women, long-term unemployed). The gold-standard National JTPA Study found ~€1,100–€1,200 positive effects for adults over 30 months — but zero to negative effects for youth. Japan’s experience mirrors this: despite generous public employment services, displaced manufacturing workers over 50 in the Rust Belt regions (e.g. Kitakyushu) showed similarly poor outcomes.
The Kurzarbeit Contrast: Prevention vs Remediation
Germany's short-time work scheme (Kurzarbeit) saved an estimated 400,000 jobs during the 2008–09 crisis at a cost of ~€5.5 billion. GDP collapsed 7%, but unemployment rose less than 1 percentage point. For comparison, the US spent ~€97 billion on unemployment compensation in 2009 alone. The lesson: prevention is 20× more cost-effective than remediation. But Kurzarbeit was designed for cyclical downturns — structural transformation at AI speed may exceed its design envelope.
The deeper problem is structural, not frictional. Susskind (2020) distinguishes two forms of technological unemployment: frictional (workers have the wrong skills for available work — a transition problem) and structural (not enough demand for human work at all). Retraining addresses frictional TU. It cannot address structural TU. Frey documents that US labour productivity grew 8× faster than hourly compensation since 1979 — a structural divergence, not a skills gap. Acemoglu & Johnson call this the “productivity bandwagon” failure: technology raises average output while reducing the marginal value of each additional worker.
Heckman’s verdict (Nobel laureate, landmark review of retraining evidence): “At best a modest impact. Many programs cannot pass a cost-benefit test. Returns to human capital investment decline steeply with age. Adult retraining cannot remedy skill deficits accumulated over a lifetime.”
2. Thirteen Displaced Worker Cases What actually happened to the people behind the statistics
The macro data tells you jobs were lost. These thirteen cases — spanning the UK, US, Germany, France, China, Japan, and South Korea — show what that meant for real communities: wage collapses, geographic scarring, deaths of despair, and the rare exceptions where institutional response made a difference. Each case follows four questions: How big was the workforce? → Did they recover? → What lasting damage? → What did institutions do?
The pattern across all thirteen: recovery is generational replacement, not worker transition. The handloom weavers didn't become factory workers — their children did. The miners didn't become tech workers — their grandchildren (in Pittsburgh, not Youngstown) did.
3. The Age Gradient Why age 50 is the approximate point of no return
The chart below tells the single most important story in displaced worker research. Green bars show the percentage of displaced workers who find new employment. Red bars show how much less they earn compared to their pre-displacement income. The cliff at age 50 is not gradual — it's a structural break.
Workers displaced at ages 25–34 have a 75% chance of finding new work and lose only ~5% of earnings. By age 50–59, reemployment probability drops to 50% and earnings loss reaches 42% — their household income is effectively halved. Men aged 50–61 are 39% less likely to find work each month compared to 25–34 year-olds (Urban Institute). After age 60, the system effectively collapses: most never return to equivalent employment.
Reemployment Rate (green) vs Earnings Loss (red) by Age
View data as table
Why the cliff exists
Three forces compound at age 50: (1) employer age discrimination reduces callback rates regardless of qualifications, (2) shorter remaining career makes retraining economically irrational — the return on investment is approximately half that of younger workers (Jacobson et al.), and (3) disability pathways open — SSDI becomes an alternative to job search, and in trade-exposed US regions, disability payments were 30–40× larger than retraining assistance.
What Protects You: Frey’s Three Engineering Bottlenecks
Frey & Osborne identified three capabilities that remain hardest to automate: (1) Perception and manipulation in unstructured environments (surgical nurses, physical therapists), (2) Creative intelligence producing genuine novelty (not just recombination), (3) Social-emotional intelligence requiring complex interpersonal interaction (coaches, negotiators, therapists). Workers whose roles require one or more of these bottlenecks are durably protected. The income gradient is stark: 83% of workers earning under €18/hour are at high automation risk; only 4% earning over €36/hour.
4. Geographic Scarring Cities survive bombs better than economic shocks
Glaeser's research shows that cities are resilient to physical destruction (bombs, earthquakes, fires) but highly vulnerable to economic shocks. The reason: physical destruction doesn't eliminate human capital; economic shocks do.
Geographic scarring — thirteen cases across six continents
Light shock (<25% / <50pp)Moderate (25–50%)Severe (>50% / >80pp)Recovery (Pittsburgh)Numbers on pins key to the table below.
| # | Case | Shock magnitude | Metric | Status | Source |
|---|---|---|---|---|---|
| 1 | Detroit · N America | −60% | Population (from 1.85M peak) | Bankruptcy 2013; still scarred | — |
| 2 | Pittsburgh · N America | Recovery | Pivot to healthcare + tech | 31% degree holders · Carnegie Mellon · UPitt | — |
| 3 | UK coalfields · Europe | −22% | Jobs per 100 working-age (40yr post) | Need 80k extra residents in work to close gap | — |
| 4 | Nord-Pas-de-Calais · Europe | Poorest | Department rank (France) | Lens, Valenciennes in bottom quintile | — |
| 5 | Gelsenkirchen (Ruhr) · Europe | +100pp | Unemployment vs national average | Scarred despite €90B+ in structural funds | — |
| 6 | Kitakyushu · E Asia (Japan) | −40% | Population (since 1970s) | Scarred despite relocation programmes | — |
| 7 | Mumbai Girangaon · S Asia (India) | −95% | Textile mill employment (post-1982 strike) | Mills razed; land converted to luxury real estate | pending |
| 8 | São Paulo ABC · S America (Brazil) | −48% | Industrial jobs: 363,333→187,759 (1989–1999) | Auto-cluster decline; partial logistics pivot | Ramalho, Rodrigues & Conceição 2009, RCCS no. 85 (RAIS-CAGED) |
| 9 | Copperbelt (Kitwe) · Africa (Zambia) | −66% | ZCCM mine employment: 56.6k→19.1k (1991–2001) | Privatisation 1997–2000; Copperbelt unemployment 22% vs 6% nat’l (2004) | Fraser & Lungu 2007 (CSTNZ/CCJDP, Ch. Mines Zambia data) |
| 10 | Abadan · Middle East (Iran) | −98% | Population: 294,068→6 (1976→1986 census) | Iran–Iraq war + revolution; partial rebuild only | Iran nat’l census (SCI); see BL-07-verification.md |
| 11 | Latrobe Valley · Oceania (Australia) | +0.7pp | Local SA4 unemployment effect after coal closures (~0.7pp controlled; pooled AU stations) | Hazelwood 2017 largest closure in sample; effects persist past 6mo | Burke, Best & Jotzo 2018, CCEP WP1809 (ABS SA4) |
| 12 | Shenyang / Liaoning · E Asia (China NE) | +35pp | Effective unemployment vs national avg (SOE reform) | 35M+ SOE laid off 1995–2002; <50% reemployed formally | — |
| 13 | Geoje (Korea) · E Asia (S Korea) | −40% | Shipbuilding employment (2015–2018) | “Industry crisis special area”; subcontractors hit hardest | — |
Strongest predictor of recovery: pre-existing human capital. Pittsburgh had it. Detroit, Gelsenkirchen, and Kitakyushu didn't.
Intergenerational Transmission
Children of displaced workers earn 9% less as adults, concentrated when displacement occurs at ages 10–14 (Oreopoulos, Page, Stevens 2008). Geographic variation in upward mobility maps closely onto areas of concentrated displacement (Chetty). Former industrial areas are “social mobility cold spots” where young people are half as likely to attend university (UK Social Mobility Commission 2025).
The Marienthal Warning: Work Provides More Than Income
Jahoda’s 1933 study of Marienthal, Austria — where factory closure left 75% of families workless — found that unemployment destroyed temporal structure, social participation, and sense of purpose, not just income. Library borrowing halved. Athletic club membership fell 52%. Anonymous denunciations tripled. Work provides five latent functions beyond wages: time structure, social contact, collective purpose, social identity, and regular activity. Unemployment benefits replace income but none of these. This is why geographic scarring persists even where benefits are generous — and why Susskind argues structural technological unemployment requires “leisure policy,” not just welfare policy.
5. What Happens Next? Ten predictions grounded in the historical record
The ten predictions for AI displacement — base rates from 580 years of data — live on the Findings & Predictions page →