TL;DR

Anthropic has published claims that three Chinese AI companies — DeepSeek, MiniMax, and Moonshot — ran “industrial-scale” distillation campaigns to extract capabilities from its Claude model, using 24,000 fraudulent accounts and generating over 16 million exchanges. The dispute touches on AI safety, US-China tech rivalry, and the legal grey area around model training practices.

What Anthropic alleges

According to Anthropic’s blog post, DeepSeek’s operation involved more than 150,000 exchanges, with techniques including asking Claude to spell out its internal reasoning step by step — effectively generating chain-of-thought training data. MiniMax allegedly accounted for over 13 million exchanges and Moonshot for 3.4 million. All three campaigns targeted Claude’s “most differentiated capabilities: agentic reasoning, tool use and coding.”

The distillation debate

Distillation — training smaller models on outputs from larger ones — is a standard industry technique. Companies routinely use their own large models as “teachers” for smaller, cheaper “student” models. The practice drew widespread attention last year when OpenAI accused DeepSeek of using it to build efficient competitors based on open-source systems from Meta and Alibaba.

What Anthropic characterises as different is the alleged scale and deception: thousands of fake accounts systematically extracting proprietary capabilities rather than using published APIs within their terms.

Counter-accusations fly

The claims drew immediate pushback. Elon Musk, whose xAI operates Grok, wrote on X: “Anthropic is guilty of stealing training data at massive scale and has had to pay multibillion-dollar settlements for their theft.” The exchange highlights the broader tension: US AI companies that built their models on vast quantities of scraped data are now objecting to others doing something similar to their outputs.

Looking forward

Anthropic warned that distilled models are unlikely to retain the safety guardrails built into frontier systems, raising proliferation risks. But with no laws governing the practice and limited legal recourse beyond terms-of-service enforcement, the industry faces an unresolved question about where legitimate model training ends and intellectual property theft begins.