How Chinese AI Labs Copied Claude: Anthropic's Distillation Report, Broken Down
Anthropic caught Alibaba pulling 151 million answers out of Claude in three months to train its own models. What distillation is, what the labs were really after, the numbers from Anthropic's September report, and what it means if you use Claude
Training a top AI model from scratch costs a fortune. Copying one costs a fraction of that, and Anthropic's September threat report shows exactly how some labs tried it with Claude.
The technique is called distillation: ask the expensive model millions of questions, then train your own model on its answers until it starts answering the same way.
What Anthropic found
- Nearly 200 million exchanges with Claude across five distillation campaigns, all attributed to China-based AI companies.
- The biggest by far was Alibaba's: 151 million exchanges between May and July 2026, which Anthropic calls the largest wholesale distillation effort it has ever seen.
- At its peak it ran nearly 3 million exchanges a day, from about 3,500 fraudulent accounts.
- The answers went into training Alibaba's own Qwen models: Qwen 3.5, 3.6 and 3.7.
Keep reading
Free, for an email.
Unlocks 3 more sections, 1 copy-paste prompt, the document version, and every other guide on the site, for free. Enter your email once to keep reading.
19,000+ follow where these guides come from.No spam. Unsubscribe in one click.