Chinese AI firms distilling U.S. models, Anthropic alleges large‑scale misuse

Anthropic says Chinese AI developers, including Alibaba’s Qwen team, have used a technique called distillation to copy U.S. models. The company identified a campaign that employed about 25,000 fraudulent accounts to generate more than 28.8 million interactions with its Claude model, and it has banned roughly 700,000 Chinese accounts. Similar accusations were made against DeepSeek, Moonshot and MiniMax. Distillation involves querying a large, expensive model many times so that the responses can train a smaller, cheaper one. While the practice is not illegal, Anthropic and OpenAI argue that doing so without permission breaches their rules and gives Chinese firms a strategic advantage. A February 2025 study by Peking University and the Chinese Academy of Sciences found evidence of distillation in most domestic Chinese models, except for ByteDance’s Doubao. In parallel, the U.S. government has lifted some export controls on powerful AI models, a move that could affect the competitive balance.