Anthropic Accuses Alibaba, Moonshot AI and DeepSeek of Running Distillation Campaigns Against Claude

Anthropic published a report Thursday accusing three China-based AI developers — Alibaba, Moonshot AI and DeepSeek — of running distillation campaigns against its Claude models, alleging the activity has grown more aggressive in recent months. The company says the operations amount to a sustained effort to extract capabilities from its systems and fold them into rival products, according to the report. The disclosure marks one of the most direct public accusations a major U.S. lab has leveled at named Chinese competitors over model-training practices.

Distillation, in this context, refers to training a model on another system's outputs — queries and responses — rather than building capability from scratch. It is cheaper and faster than independent pretraining, and it sits in a gray zone: some forms are common industry practice, while doing it at scale against a competitor's API typically violates terms of service. Anthropic's claim is that the activity it documented goes well beyond incidental use, describing it as persistent and coordinated rather than the work of isolated users.

The three named companies are not fringe players. Alibaba operates the Qwen family of models and one of China's largest cloud businesses; DeepSeek drew global attention with low-cost models that traded blows with leading U.S. systems; Moonshot AI, behind the Kimi assistant, has raised significant venture funding and pushed long-context models. All three compete directly with Anthropic and other U.S. labs for enterprise customers and developer mindshare, which gives the allegation a commercial edge as much as a technical one.

Anthropic did not disclose in the summary how it identified the campaigns or what volume of traffic it attributes to each company, leaving key evidentiary questions unanswered. The report arrives amid rising tension between Washington and Beijing over AI, with export controls already restricting Chinese access to advanced chips and policy debate in the U.S. focused on how frontier labs should police misuse of their APIs. Accusations like these can feed directly into that debate, and they carry weight because they come from a company whose models are widely used by businesses.

Distillation disputes are difficult to prove and harder to litigate, since API traffic can be routed through intermediaries and terms-of-service violations are usually handled contractually rather than in court. What Anthropic does next matters: it could tighten rate limits and identity checks, share technical details with regulators, or press for policy action. The open question is whether other U.S. labs corroborate the pattern, and whether Chinese developers named respond with their own account of how their models were trained.