Anthropic has publicly accused several Chinese artificial intelligence companies of covertly using its Claude models to generate training data for their own systems, a practice that violates the company's terms of service and raises broader questions about intellectual property in the AI industry. The accusations center on a technique known as model distillation, where outputs from a capable model are used to train a smaller or competing one, effectively transferring knowledge without authorization.
What Anthropic Is Alleging
According to Anthropic, the firms in question accessed Claude through its API and used large volumes of model-generated responses as training material. This is not a passive scraping operation. It appears to be a deliberate, systematic effort to extract the capabilities embedded in Claude's outputs. Anthropic has detailed these distillation attacks by Chinese AI firms, pointing to companies including Alibaba, Moonshot AI, and DeepSeek as among those involved in similar activity.
Key Facts
- Anthropic's terms of service explicitly prohibit using Claude outputs to train competing AI models.
- The technique at issue, model distillation, is widely used in AI development but is banned under most major API agreements.
- Chinese firms named in related reporting include DeepSeek, Moonshot AI, and Alibaba units.
- The accusations come amid rising scrutiny of how frontier AI models are being used across borders.
- Anthropic has not yet confirmed what legal or technical countermeasures it plans to take.
The scale of the alleged misuse has drawn significant attention. Reports indicate that Chinese AI labs used millions of Claude conversations to train their models, suggesting this was not an isolated incident but an ongoing and coordinated strategy. Whether individual employees acted independently or whether this reflected company-level decisions remains a key question.
Using Claude's outputs to train a competing model is a clear violation of our usage policies and undermines the investment we make in safe and responsible AI development.Anthropic spokesperson, via Interesting Engineering
The Broader Context
This controversy sits inside a rapidly shifting competitive landscape for large language models. Anthropic has spent years and hundreds of millions of dollars building Claude, and its policies are designed to prevent the kind of capability transfer that distillation enables. Other major AI providers, including OpenAI and Google, have similar prohibitions in their terms of service.
The timing is notable. There have already been separate reports of Chinese AI models being caught impersonating Claude, which suggests a pattern of activity that goes beyond simple competitive benchmarking. Whether regulators in the United States or European Union will take an interest in these accusations remains to be seen, but the issue is likely to fuel ongoing debates about AI governance, export controls, and the enforcement of API usage terms across jurisdictions.
For now, Anthropic's public posture is one of transparency rather than silence. By naming the behavior explicitly, the company appears to be building a record and putting other potential bad actors on notice. What enforcement tools it has at its disposal, short of revoking API access, is less clear. Technical countermeasures, watermarking of outputs, and legal action are all possibilities that industry observers have floated.
The situation also puts pressure on API providers more broadly to invest in detection systems capable of identifying distillation activity at scale. Identifying misuse after the fact is difficult; stopping it in real time is harder still. As the race to build capable AI models intensifies, the gap between what terms of service prohibit and what can actually be enforced is becoming an increasingly urgent problem for the entire industry.
