Bessent says U.S. could sanction China over AI model 'theft'

U.S. Treasury Secretary Scott Bessent announced that the current administration is investigating whether Chinese artificial intelligence models are using technology distilled from American software. This accusation centers on distillation, a training method where smaller models are created by mimicking the outputs of more powerful systems. Bessent claims the government has identified watermarks from U.S. large language models within various Chinese offerings.

Chinese startups are gaining ground against U.S. leaders like OpenAI and Anthropic. One recent example is the Kimi K3 model from Moonshot AI, which outperformed competitors in several benchmarks. Anthropic recently informed the Senate Banking Committee that it suspects the Chinese firm Alibaba of executing a major distillation attack against its own systems. The government views this as a threat to the U.S. competitive advantage in technology.

Economic tensions are increasing as officials prepare for bilateral talks on artificial intelligence in September. Bessent stated that the U.S. holds the authority to issue sanctions if the evidence proves that intellectual property was stolen. He emphasized that the theft of proprietary model data is unacceptable and indicated that the administration will review this evidence in the coming days.

Legal disputes regarding AI training data are not limited to international conflicts. Domestic firms face similar scrutiny, with Anthropic settling a massive copyright lawsuit last year and ongoing litigation between The New York Times and Microsoft over model training practices. The global race for artificial intelligence dominance remains a focal point for both domestic policy and international trade strategy.