US weighs AI distillation crackdown as labs cite $6 billion annual loss

Unauthorized AI model distillation by foreign labs costs American AI companies up to $6 billion a year, the US government estimates, according to reporting published July 13. Distillation here means querying US frontier models at scale to cheaply train rivals.

The estimate lands as Anthropic and OpenAI press Washington for sanctions and tighter export controls, escalating a months-old fight over how much US labs lose when their models are copied through the interface.

The pressure follows a letter Anthropic sent the Senate Banking Committee in June, which accused Alibaba’s Qwen lab of running about 28.8 million exchanges with its Claude models through roughly 25,000 fraudulent accounts between April and June. Anthropic has called for penalties or sanctions on firms that use illicit distillation, alongside stronger export controls.

The White House Office of Science and Technology Policy is reportedly weighing executive action that could, within roughly six months, restrict open-weight models above capability levels comparable to GPT-5.5, Claude Opus 4.8 or GLM-5.2.

The framing is contested. Nathan Lambert, a machine-learning researcher, wrote on July 12 that Anthropic’s distillation campaign is partly a competitive and regulatory-capture play, pointing to a separate incident in which unauthorized users reached Anthropic’s Mythos model through Discord as evidence that interface security, not only open-weight distillation, is a comparable risk.

The $6 billion figure is a government estimate that has not been publicly detailed and could not be independently confirmed, and Anthropic has not released the underlying data behind its account-fraud tally.