Let China copy American AI

Anthropic and OpenAI forbid “distillation” (training a weaker model on a stronger model’s outputs) in their terms, have complained to the US government about Chinese rivals, and Anthropic wants penalties; a Senate bill would sanction distillers and bar their models from America. Experts put Chinese models three to six months behind the US frontier, largely thanks to distillation. In this leader The Economist argues against a ban: competition from mostly free Chinese models helps consumers and pushes US labs to compete in poorer countries; dependence on copying leaves China ill-prepared for a self-improving AI take-off and possibly in permanent second place (the Soviet chip-copying analogy); cutting it off would set China back only months, so that option is better kept for a crisis; US safeguards (e.g. on bioweapons) carry over into distilled models; and the labs, built on scraped web content they defend as fair use, can hardly claim copying hurts frontier investment, while machine output can’t be copyrighted. It accepts the labs’ concern is “understandable” and that they may restrict distillation under their own terms. The companion Science & technology piece (“China increasingly depends on copying American AI”, 25 October, not counted separately) supplies the evidence: US frontier labs now spend more than half their compute on reinforcement learning, up from a twentieth in early 2024; a small DeepSeek model trained on 800,000 examples rose from 15% to 46% on a maths test; Anthropic found five Chinese firms collected over 180m messages in five months; its countermeasures cut harvested data by 90-99% but may cost 5-7% of revenue; Saif Khan estimates the gap could nearly double without distillation, and ByteDance’s Zhang Yiming has told staff it won’t rely on it.