Randomized evidence shows typical job-training programs yield only modest labor-market gains—about a 1.7 percentage-point employment uptick and ~$800 extra per year—while a few employer-led 'sector' programs produce tenfold earnings gains but rely on heavy selection and close employer relationships that resist rapid scaling.

Main Finding

Across 56 randomized trials in the US (146 impact estimates), subsidized job training has on average small positive impacts: employment rises about 1.7 percentage points in years 3–5 after assignment and annual pre‑tax earnings rise roughly $800 per person offered training. At typical program costs (~$13,600 per participant) these programs are roughly break‑even under plausible (but debatable) assumptions about persistence of effects. A distinct subset—“sector programs” that screen applicants, co‑design curriculum with employers, and directly place trainees—produce much larger, often sustained gains (order $5–10k/year), but they are selective, privately run, and hard to replicate at scale.

Key Points

• Evidence base
  • Meta‑analysis: 56 randomized US trials (1973–present), 146 outcome estimates; supplemented by a few high‑quality European and judge‑randomization studies.
  • Follow‑up windows are typically short (most studies stop by year 6).
• Average impacts (training-primary interventions)
  • Employment: +2.8 pp in year 2; +1.7 pp in years 3–5 (per person offered training).
  • Earnings: ~$1,139 (year 2) and ~$791 (years 3–5) in 2025 dollars (per person offered treatment).
• Costs and benefit accounting
  • Average program cost (67% of interventions that report costs): ~$13,598 per treatment‑group member.
  • Estimated NPV of earnings impact (projecting benefits to retirement): ~$14,146; rises to ~$19,525 including employer‑paid taxes and fringe benefits. Result: modestly better than break‑even under the report’s projection assumptions.
  • From the government’s perspective, higher taxes and lower benefits recoup ~75% of program cost, leaving a long‑run fiscal cost of roughly 24% of upfront spending.
• Compliance and interpretation
  • Non‑compliance is common: treated individuals are 66 percentage points more likely to receive targeted training; only 28 percentage points more likely to receive any training versus control.
  • Estimates are reported per offer; impacts per actual trainee are ~2–4× larger, but that is an upper‑bound estimate and depends on higher compliance scenarios.
• Large federal programs
  • JTPA (late 1980s): modest positive impacts (employment +~2.3 pp; earnings +~$1,100).
  • Job Corps: essentially null long‑run impacts (minor transitory effects years 3–5).
  • Workforce Investment Act evaluations: essentially zero impacts for adults and dislocated workers (low uptake differential reduced estimated treatment effects).
• Sector programs
  • Examples: Year Up, Per Scholas, Wisconsin Regional Training Partnership.
  • Typical traits: intensive applicant screening (often rejecting >80% / accepting ~1-in-5), strong employer engagement in curriculum and placement, short‑to‑medium duration skills training plus soft skills and placement services, privately run.
  • Costs comparable or lower than average training programs ($11,602 vs $13,598 average) but much larger present value earnings impacts (~$60,319), implying high benefit‑cost ratios (report cites ~7× for the sector cluster).
  • Replication and scaling have been difficult; success appears to rest on employer relationships, selectivity, and local labor‑market intelligence.
• Dislocated workers / trade‑affected groups
  • Evidence is sparser and mixed. Some natural‑experiment / judge‑randomization studies (e.g., Denmark) show classroom training can help occupational change; US TAA evidence suggests positive labor market adjustments but conflates training with extended UI.
• Limitations and external validity
  • Most trials target low‑income adults and youth and short trainings (weeks–months). The evidence is therefore imperfect for predicting retraining needs of displaced mid‑career or highly skilled white‑collar workers who may need multi‑year reskilling.
  • Persistence of benefits is uncertain; projecting benefits to retirement drives the fiscal break‑even result.

Data & Methods

• Inclusion criteria
  • High evidentiary standard: randomized controlled trials prioritized; credible natural experiments (regression discontinuity, judge randomization) included.
  • Aggressive literature search including government reports, archive scans, author contacts; code and data published at github.com/droodman/job-training-meta-analysis; web interface at droodman.github.io/job-training-meta-analysis.
• Meta‑analysis approach
  • 56 randomized US trials form the core of the meta‑analysis (146 outcome observations).
  • Outcomes extracted at the finest temporal resolution available (monthly/quarterly/yearly).
  • Main estimates reported per person offered training (intent‑to‑treat); compliance and take‑up details are recorded and used to compute per‑trainee bounds.
  • Random‑effects means used for average impacts; cost data available for ~67% of training‑primary interventions.
• Complementary narrative review
  • Deep dives into national scale‑up evaluations (JTPA, WIA, Job Corps), sector program evaluations, and studies of dislocated‑worker supports.
  • Attention to implementation details (screening, employer engagement, placement mechanisms) and scalability challenges.

Implications for AI Economics

• Realistic expectations for retraining as an AI policy
  • Retraining is popular and plausible as part of an AI disruption response, but the historical evidence suggests modest average returns for the kinds of short‑duration programs that currently dominate policy portfolios.
  • Large, rapid displacement of white‑collar workers would pose different challenges than the evidence base addresses (these workers often require longer, credentialed retraining).
• Value of sectoral, employer‑engaged models
  • Programs that tightly align with employer demand and screen entrants can yield large returns and positive fiscal payback—important for short‑horizon, high‑demand reallocation following automation in specific sectors.
  • However, these models are selective and not well suited for universal mass retraining unless scalability and equity tradeoffs are addressed.
• Fiscal and budgetary takeaways
  • Average training programs can be close to fiscally neutral under optimistic persistence assumptions; government recoups a large share of upfront costs via taxes and reduced benefits, lowering long‑run net fiscal burden.
  • Policymakers should be explicit about persistence assumptions when budgeting for mass retraining; short follow‑up windows mask uncertainty about long‑term payback.
• Design priorities if retraining is scaled for AI shocks
  • Prioritize programs with employer partnerships, fast cycles of curriculum updating, and placement guarantees where possible—these features are correlated with much higher impacts.
  • Build selective pilot expansions (randomized scale‑ups) of promising sectoral programs to test scalability before large unconditional rollouts.
  • For displaced, higher‑skilled workers, invest in evaluating compressed degree or mid‑career credential programs (masters, bootcamps adapted for professionals), since the existing evidence mostly concerns lower‑wage, short‑training cohorts.
  • Account for compliance/take‑up dynamics in planning: per‑offer impacts understate per‑trainee gains; but real‑world take‑up will vary and affects cost‑effectiveness.
• Research and evaluation agenda for AI preparedness
  • Rapid “fire‑drill” randomized expansions of leading sectoral programs to learn about scalability, cost structures, and employer pipelines in an urgent response context.
  • RCTs and quasi‑experiments focused on retraining for white‑collar occupations (compressed degrees, industry partnerships for professionals).
  • Measurement of persistence: invest in longer follow‑ups and administrative linkages to credibly project lifetime benefits.
  • Investigate complementarities between retraining and other policies (enhanced UI, wage subsidies, job search assistance, relocation support) for displaced workers in different skill strata.
• Bottom line for AI policy economists
  • Retraining can be part of an AI response toolkit, but past evidence cautions against treating existing mass retraining infrastructure as a turnkey solution for large, rapid labor‑market displacement. High‑impact models (sector programs) exist and merit prioritized, rigorously evaluated scale‑ups; otherwise, expect modest average returns and nontrivial implementation and equity tradeoffs.