Anthropic claims Chinese AI firms used Claude to train rival models

Anthropic claims that Chinese AI companies used its Claude model to train competing models, according to a report by Interesting Engineering. The allegation points to a long-standing gray practice: calling a frontier model's interface to obtain high-quality outputs, then feeding them into training data for a homegrown model.

If the claim holds, it touches the already blurry boundary between terms of service and model distillation. For the provider, model outputs are a core asset; for the user, high-quality outputs are a shortcut far cheaper than training from scratch, and that tension is where the dispute lives.

The report does not present a complete list of named companies, and it reflects a claim Anthropic is making rather than a third-party-verified finding. As it stands, this reads closer to a public position than a settled conclusion.

The dispute is sensitive because it touches both commercial interest and geopolitical competition at once. Distillation can be framed as reasonable technical reuse or as a way to sidestep licensing and compute barriers, leaving wide room for narrative.

For the industry, the episode puts the long-tolerated question of whether model outputs may be used to train rivals back on the table. Platforms may tighten terms of service, detect unusual usage patterns, and consider adding more restrictions at the output layer.

What to watch next is whether Anthropic publishes more specific evidence or a timeline, whether the companies involved respond, and whether other frontier developers adopt similar constraints. Those answers will determine whether the claim stays at the level of rhetoric or becomes an enforceable industry norm.