Anthropic warns existing worker-retraining programmes are too small for AI-driven job losses

Anthropic published a review of the evidence on government-funded worker-retraining programmes, drawing on 56 randomised US studies plus additional experimental evidence from Europe, and concluded that current programmes are unlikely to be sufficient if AI displaces workers at scale.

The meta-analysis found that offering someone a training slot raises their probability of employment by two to three percentage points and their annual earnings by roughly $1,000, at an average cost of about $13,000 per participant. Anthropic calculated that the government recovers more than half of what it spends through the resulting gains, mainly via higher tax receipts and lower benefit payouts — a positive but modest return, and one the review said falls well short of what would be needed to offset large-scale displacement.

The review flagged one exception: “sector programs,” which partner directly with employers in high-demand fields, produced gains several times larger than the general-purpose average. But it also noted these results have frequently failed to replicate when tested at new sites or larger scale, making it hard to know how much of the sector-programme effect would survive expansion.

Anthropic’s stated conclusion was not that retraining is ineffective, but that the scale and design of existing programmes were built for a labour market disrupted by trade and automation shocks over years, not one potentially disrupted by AI over a much shorter horizon. It recommended investing now in testing and scaling the most promising programme designs, with rigorous measurement built in from the start, rather than waiting for evidence of AI-driven displacement to accumulate before acting. The review is one of several efforts by Anthropic in 2026 to quantify labour-market effects of AI ahead of, rather than after, disruption becomes visible in employment data.