Research
AI Minute Newsroom
2026-08-13
Anthropic asked whether retraining works. Across 56 trials it buys two or three points of employment.
A review published by Anthropic on 12 August, written by the independent researcher David Roodman with Anthropic's Maxim Massenkoff, pools 56 randomised American studies of job training. The result is small. Offering someone a training place raises their chance of being employed by two to three percentage points and their earnings by roughly $1,000 a year, at a cost of about $13,000 per place. Government recovers more than half of that through extra tax and lower benefit payments, so the programmes roughly break even for the public purse without transforming anyone's prospects. One exception stands out: 'sector programmes', which partner with employers in an industry that is actively hiring and place people directly into those jobs, do several times better — but attempts to copy them elsewhere have repeatedly failed. The authors conclude that if AI displaces workers at scale, the retraining system as it currently exists would not cope, and argue for testing and scaling the few approaches that work before that happens.
Why it mattersRetraining is the answer every politician and every AI executive reaches for when asked about jobs, and it is usually offered as though it were a solved problem. It is not. This is the most careful summary of the evidence anyone has put in front of the public, and it comes from a company with an obvious interest in the answer being reassuring — which makes the finding harder to wave away. If you work somewhere that talks about reskilling, the useful question after reading this is which kind of programme is meant: the generic sort that buys two points, or the employer-partnered sort that works and is hard to replicate.
✓ Verified · 2 sources
Read in the app — free, in 9 languages
Related stories
OpenAI says its new models block nearly all reasoning-theft attacks.
2026-11-22Lean's own community says the AI-assisted Riemann proof is correct.
2026-11-22Researchers say an open AI model improved by editing its own weights.
2026-11-21A US watchdog found that no AI therapy chatbot has FDA authorization.
2026-11-21Outside mathematicians say the AI Riemann proof passed first checks.
2026-11-21