Beijing Says US AI Firms Trained on Chinese Models. Washington Says China Did It First.

Beijing's commerce ministry publicly accused US AI firms of training their AI models on the outputs of Chinese AI systems (MOFCOM statement), a direct counter to last week's US allegation that Chinese startup Moonshot AI covertly copied Anthropic's most advanced model to build its Kimi K3, released July 16.

Distillation is a standard training technique: one model learns to mimic another's outputs. It is widely used across the industry, not unique to China or to illicit copying, though Washington treats unauthorized distillation as a form of theft (CSET analysis).

MOFCOM called US behavior "AI hegemonism" and said Washington "lacks factual basis and legal grounds" while "adopting double standards in practice." Beijing warned of "all necessary measures to firmly safeguard its legitimate and lawful rights and interests." The ministry noted that Chinese AI model release dates are now "very close" to US frontier models and that some Chinese systems have already taken the lead on certain capabilities.

The exchange lands as Treasury Secretary Scott Bessent threatened sanctions against China last week, and the US weighs a ban or curbs on foreign-made open-source models, many built through distillation. The concrete next step is the US decision on those open-source curbs and what they cover.