China's Commerce Ministry turns the AI distillation charge back on US labs

Beijing's Ministry of Commerce answered the United States' AI theft charge on Monday with one of its own. A ministry spokesperson told a regular press briefing that "many American AI companies have distilled Chinese models in their research and training," flipping an allegation the White House had aimed at Chinese labs days earlier.

Distillation is a machine-learning training technique in which a smaller model is taught to mimic a larger model's outputs. It is a standard tool in the AI engineer's kit, and the one at the center of the current dispute. The US charge, leveled by a White House official, accused Chinese AI firms of distilling Anthropic's Claude. MOFCOM's counter is that the flow runs in both directions, and that the timeline makes one direction implausible.

The ministry's defensive case rests on two pillars. First, release dates: Chinese frontier models, including the K3 model the White House specifically named last week, hit public release within weeks of the US frontier models they are now accused of copying. That window, in MOFCOM's telling, is too short for the kind of systematic output sampling that distillation requires. Second, position: on at least some axes, the Chinese models already lead. "Some Chinese models are already in a leading position," the spokesperson said, in language The Register reported verbatim from the briefing.

Beijing paired the counter-allegation with a warning. Asked what China would do if Washington moved to sanction Chinese AI firms, the spokesperson said China would take "all necessary measures" without naming specific retaliation vectors. The briefing's framing vocabulary leaned on a phrase that translates as "AI hegemonism" (人工智能霸权主义), and accused the United States of politicizing and instrumentalizing technology trade.

The Chinese case leans on a US audience the White House does not control. Inside the MOFCOM statement, the ministry cited "nearly 200 US startups" that recently asked the US government not to cut their access to Chinese open-source models, and unnamed "large US multinationals" that told Beijing distillation is a widely used technique and one US firms should be free to apply to Chinese open-weight releases. Both claims come from MOFCOM's own characterization of those exchanges and have not been independently verified.

A separate industry signal runs in the same direction. In the days before MOFCOM's briefing, 77 technology companies signed an open letter to US lawmakers defending open-weight AI development. In earlier public remarks, the White House official called distillation a legitimate technique when used to build "smaller, more efficient models." The US complaint about Chinese distillation, in other words, sits on a foundation that parts of US industry and the White House's own science office have described as ordinary engineering.

China has built a diplomatic scaffolding of its own. In recent weeks Beijing convened the 29-member World Artificial Intelligence Cooperation Organization (WAICO) to promote an AI governance pitch that includes open-sourcing models as a public good. The MOFCOM statement closed with an invitation to joint AI governance dialogue, aimed at the same audience the White House is trying to mobilize against Chinese models.

Neither side has proven who distilled whom. From here, the US government will have a hard time calling distillation theft when its own science office, large parts of its industry, and a foreign ministry have put on the public record that the practice is ordinary. Each new accusation will land on a public that has now read the rebuttal.