Anthropic has released a comprehensive meta-analysis finding that worker retraining programs deliver statistically significant but modest economic returns—results that arrive as policymakers worldwide grapple with how to address potential AI-driven labor displacement, and as projections from organizations including Goldman Sachs and the World Economic Forum estimate that generative AI could expose hundreds of millions of jobs to significant change within the coming decade.

The report, co-authored by independent researcher David Roodman and Anthropic's Maxim Massenkoff, employs an AI-accelerated methodology in which Anthropic's Claude model extracted the majority of underlying data and generated all analytical code. The study draws on 56 randomized controlled trials conducted in the United States since the 1970s, supplemented by experimental evidence from Europe, constituting one of the most systematic evaluations to date of government and nonprofit training initiatives. The full report is available in PDF form.

Measured but Limited Outcomes

On average, the programs examined produced positive yet constrained effects. For each individual offered a training slot, employment increased by approximately two to three percentage points, while annual earnings rose by roughly $1,000. These gains were measured against an average per-participant cost of about $13,000. From a fiscal standpoint, governments recovered more than half of that expenditure through additional tax revenue and reduced public benefit payments, meaning the interventions approximately broke even overall. The United States has historically spent less on active labor market programs as a share of GDP than most peer OECD nations, a gap that places the modest returns documented here in a broader context of comparatively underfunded workforce infrastructure, primarily administered through the Workforce Innovation and Opportunity Act.

The findings form part of Anthropic's broader Economic Research agenda, which monitors AI adoption across occupations and industries. Although retraining consistently ranks as the most popular policy response to technological unemployment in both public and expert surveys, the authors caution that historical performance indicates current program designs would likely fall short if advanced automation were to displace workers at substantial scale.

Sector Programs and the Replication Problem

A notable exception to the modest averages comes from a small group of "sector programs"—initiatives that work closely with employers in high-demand industries to screen applicants, develop occupation-specific curricula, and place graduates directly into roles. Programs such as Year Up and Per Scholas have raised participant earnings by several multiples above the mean, operating effectively as labor-market intermediaries that connect overlooked talent with employers committed to sustained hiring pipelines. Year Up's results were independently corroborated in a multi-year randomized controlled trial conducted by the research firm MDRC, one of the largest evaluations of its kind in workforce development.

However, the report emphasizes that replicating these successes has proven exceptionally difficult. High-fidelity copies of leading models have frequently failed at new locations, indicating that their effectiveness relies on delicate, context-specific conditions: deep local relationships, highly selective admissions processes that screen out more than 80 percent of applicants, and organizational capabilities developed over years of operation.

Mismatch with AI Displacement Scenarios

Because most programs studied targeted low-income or marginally employed individuals rather than mid-career white-collar professionals who might require extended reskilling periods, the authors conclude that existing retraining infrastructure is poorly aligned with a scenario of widespread AI-driven displacement. This mismatch is particularly salient given that large language models are increasingly demonstrating capabilities in tasks such as writing, coding, legal analysis, and customer support—occupations that traditional workforce programs were never designed to address.

Their central recommendation is to invest now in demonstrating, evaluating, and scaling the most promising program models before any large-scale crisis emerges. Specifically, they propose rapidly expanding a leading sector program for a well-defined worker cohort while rigorously tracking employment and earnings outcomes. Anthropic's Economic Futures Research Fund is positioned to support such investigations, reflecting an emerging consensus that understanding the limits of current retraining approaches is essential to preparing for an era of rapid automation.