Does retraining work for displaced workers?

Evidence from 207 studies (Card, Kluve & Weber) shows short-run training impacts are near zero. The US Trade Adjustment Assistance programme's comprehensive Mathematica 2012 evaluation found that of every 100 eligible workers, only 18 ultimately found employment in their retrained field. Those who completed training earned about €3,000 less than similar workers who received no training at all. Modest positive effects appear only after 2+ years, and only for adult women and the long-term unemployed.

How long does geographic scarring from industrial decline last?

Evidence from thirteen industrial cities across six continents shows scarring typically persists 30–50+ years after the shock. The UK coalfields still run 22% below baseline on jobs per 100 working-age residents 40 years later. Gelsenkirchen's unemployment is still double the German national average despite €90+ billion in structural funds over 60 years. Only Pittsburgh has clearly recovered, and the strongest predictor was pre-existing human capital (31% degree holders, Carnegie Mellon, University of Pittsburgh).

At what age does reemployment probability collapse after displacement?

Age 50 is the cliff. Reemployment probability drops sharply past that threshold, and when workers do find new jobs, earnings are roughly halved. This pattern holds across the US, Germany, Japan, and Korea. Children of displaced workers earn 9% less as adults, concentrated when displacement occurs at ages 10–14 (Oreopoulos, Page & Stevens 2008).

Is prevention cheaper than remediation for technology displacement?

Yes, by roughly 20×. Germany's Kurzarbeit short-time work scheme saved an estimated 400,000 jobs in 2008–09 at a cost of €5.5 billion. GDP collapsed 7%, but unemployment rose less than 1 percentage point. The US spent approximately €97 billion on unemployment compensation in 2009 alone. Kurzarbeit was designed for cyclical downturns, however, and may not scale to structural transformation at AI speed.

What 580 Years Actually Shows

Ten findings from the historical record and ten predictions for AI, grounded in 52 primary sources and 20 case studies.

By Philipp Maul · Nexalps · Early 2026 · Part 3 of 7 in the European AI Labour Market suite

Based on 52 primary sources including Acemoglu & Restrepo, Dittmar, Feigenbaum & Gross, David and Autor.

1. Ten Findings from 580 Years What the historical record actually shows

These findings emerge from the full dataset: 10 general-purpose technologies and 11 occupation-level disruptions. Each one is grounded in specific case evidence.

2. Ten Predictions for AI Displacement What the historical record suggests will happen next

These aren't forecasts. They're the base rates from 580 years of data. The burden of proof is on anyone claiming AI will be different.

3. Spreadsheets Counterfactual The one case where massive task exposure produced more of the occupation, not less

In a sample of 20 disruptions, accountants are the single counterexample to the displacement pattern: US accountants grew from ~1M (1985) to ~1.4M (2020) after Excel substituted the core calculation task. Three structural conditions had to hold at once: elastic output demand, an intact skill ladder, and complement-not-substitute for the core task. The same scorecard can be applied per occupation to AI exposure.

Read full analysis →

4. Two Eras of Institutional Response Why institutional response lag cannot be measured for pre-1930 cases

The 20 cases partition into a pre-welfare-state era (1440–1880), where institutional response apparatus did not yet exist, and a post-welfare-state era (1930–present), where institutional response lag (the years between first displacement and the first major policy response) can be measured. The current freelance/platform/cross-border-remote coverage gap is the second structural gap of this kind. Historical precedent suggests it closes through institutional invention, not extension.

Read full analysis →

5. Methodological Appendix Cross-case metric feasibility matrix

Per-case data confidence per metric: where numbers are observed (H), derived via triangulation (M), heavily estimated (L), or unavailable (U). Full narrative + scope decision on analysis.html. Read full analysis →

Author’s Note

The most uncomfortable finding in 580 years of disruption data isn't that workers suffer. Everyone expects that. It's that recovery is the exception, not the rule. Of the thirteen cases studied, only telephone operators experienced anything close to a “good news” outcome, and that required a 57-year transition timeline plus a rapidly growing economy absorbing displaced women into new roles.

The actual adjustment mechanism, again and again, is not retraining or relocation. It's generational replacement. The handloom weavers didn't become factory workers; their children did. The miners didn't become tech workers; their grandchildren (in Pittsburgh, not in Youngstown) did. The question for AI isn't whether the economy will eventually adjust. It will. The question is whether we're willing to accept that “eventually” means a generation of workers who bear the full cost.

Philipp Maul

Go deeper

Five deep-dive analyses extend these findings. Or jump to the 20 case studies and worker outcomes.

Deep Analysis → 20 Case Studies → Worker Outcomes →