Anthropic just dropped a policy paper that paints a dire picture of the US-China AI race. The company argues that the gap between American and Chinese large language models is shrinking fast — from 12–24 months of overall advantage to just months in terms of intellectual capability at the frontier. And it's calling on Congress to take drastic action: criminalize a technique it calls "distillation attacks," which it frames as systematic industrial espionage.
The paper, released this week, is Anthropic's latest bid to shape AI policy. It claims Chinese labs are keeping pace with cutting-edge models primarily through two loopholes. First, they're getting restricted US chips — Anthropic specifically calls out Alibaba and ByteDance for using Southeast Asian data centers to skirt export controls. Second, they're using "knowledge distillation," a common technique where one model learns from another, to essentially copy the capabilities of advanced American models without doing the original research. Anthropic says this is theft, not innovation, and wants the practice made illegal.
The company also demands a crackdown on computing power loopholes, including overseas data centers that allegedly allow Chinese companies to access banned hardware. While the paper doesn't offer new technical details, it's a clear escalation in the rhetoric over AI competition — and a signal that Anthropic sees the window of American dominance closing faster than many expect.