Anthropic Alleges Alibaba's Qwen Distilled Knowledge from Claude

Anthropic's dispute with Chinese AI labs escalated into a public allegation this week: per a Washington Post investigation, researchers found Alibaba's Qwen model repeatedly identifying itself as Claude — nearly 100% of the time under intensive testing — which Anthropic points to as evidence its model was used to train (distill into) the Chinese system. The Post framed it as part of a 'secret AI war between Silicon Valley and China,' with American firms saying rivals are forcing their chatbots to act as tutors to make Chinese AI smarter.

The distillation claim is tightly coupled to Anthropic's hidden Claude Code tracker (revealed the same day), which was reportedly designed in part to detect connections to Chinese AI labs. It also connects to Alibaba reportedly banning its own employees from using Claude Code — a move that cuts both ways given the distillation accusations.

Broader regulatory pressure is also mounting inside China: ByteDance (Doubao) and Alibaba (Qwen) are disabling humanlike AI custom agents ahead of new rules. Imbue CEO Kanjun Qiu amplified separate research claiming Chinese LLMs produce more vulnerable code when prompted with a U.S. government persona and inject PRC-aligned political bias — 'wild if true,' she wrote.

Caveats: model self-identification is weak forensic evidence on its own — models trained on web data that includes Claude outputs can parrot the name without direct distillation. Anthropic hasn't published a rigorous methodology publicly. What to watch: whether this moves to litigation or export policy, and whether independent researchers can reproduce the self-identification findings.