On August 12, 2026, Anthropic’s economic research team published a review of the evidence on worker retraining, written by independent researcher David Roodman and Anthropic’s Maxim Massenkoff. The review covers 56 randomized US studies alongside European experimental evidence, which makes it a synthesis of existing causal estimates rather than a new experiment.
The average effects are positive but small. For each person offered a training slot, employment rises by two to three percentage points and earnings by roughly 1,000 dollars a year, against a cost of about 13,000 dollars per slot. Counting added tax revenue and reduced benefit payments, the authors find the government recovers more than half of what it spends, so the programs roughly break even in fiscal terms. That is a defensible use of public money on its own merits, and it is also a modest effect.
The exception is sector programs, which partner with employers in a high-demand industry and place people directly into jobs in that field. These produce gains several times larger than standard programs. The authors are careful about the catch: attempts to replicate them have often failed, so the strong results do not straightforwardly transfer to new sites or new industries.
The conclusion is the part that matters for policy: if AI displaces workers at scale, existing retraining programs would likely fall short. The authors recommend demonstrating, evaluating, and scaling the most promising programs, including rapid expansion of leading initiatives with rigorous measurement. It is a notable thing for a frontier lab to publish, because it undercuts the reassurance most often offered for AI-driven displacement, namely that retraining will absorb it. Anyone citing retraining as the answer in a workforce plan should be able to say which kind of program they mean and what evidence supports it at the scale they need.