China accused U.S. artificial intelligence companies on Monday of training their models on Chinese systems, days after Washington threatened sanctions over the alleged theft of American technology.

The timing casts Beijing’s move as a deliberate counter to U.S. pressure rather than a standalone complaint.

“It is understood that many American artificial intelligence enterprises have distilled Chinese models during their research, development, and training processes,” China’s commerce ministry said in a statement. Distillation trains a smaller model to mimic the output of a larger one. The method is common across the industry, but U.S. officials say Chinese firms have used it at scale to copy proprietary American systems.

The exchange follows a specific trigger. Last week a White House official accused Chinese startup Moonshot AI of covertly copying Anthropic’s most advanced model to build its Kimi K3. Its release unsettled the U.S. industry and reopened debate over the AI rivalry with China. Treasury Secretary Scott Bessent then threatened sanctions, while Washington weighs curbs on foreign open-source models often built through distillation.

Beijing rejected the U.S. case and turned it back on Washington. The American accusations “lack factual basis and legal grounds, adopt double standards in practice, and constitute typical acts of artificial intelligence hegemonism,” the Chinese commercial ministry said. It pledged “all necessary measures” to defend China’s interests. By flipping the distillation charge and branding U.S. conduct hegemonic, China aims to unsettle claims of American technological primacy and court a wider audience skeptical of Washington’s dominance.

That contest points to a deeper strategic logic: frontier AI is redrawing the balance of power among states while making that balance far harder to see. Neither Washington nor Beijing can reliably measure what the other has built, or how, so each reads the other’s advances as theft and answers in kind. Distillation charges, sanctions threats, and open-source bans are less about proving a specific case than about contesting an advantage that neither side can fully observe. In such a contest, advantage may rest less with the state that builds the most capable model than with the one that can see what its rival already holds. (EIR)