U.S. Treasury Secretary Signals Possible Sanctions on Chinese AI Firms Over Intellectual Property Theft

The U.S. government has launched an investigation into whether Chinese open-source artificial intelligence (AI) models have infringed upon intellectual property rights held by U.S. companies. Treasury Secretary Scott Bessent disclosed in a July 21 interview with Fox Business that the U.S. is reviewing allegations that Chinese AI models may have used American technology without authorization.

Bessent clarified that the U.S. government supports open-source AI initiatives but draws a firm line against intellectual property infringement. He stated, "We support open-source models, but we do not support intellectual property infringement."

He further warned that if evidence confirms that foreign AI companies have stolen U.S. technology, the government possesses the authority to impose sanctions on those entities.

This announcement comes amid rapid advancements in the performance and global dissemination of Chinese AI models. A notable example is "Kimi K3," developed by China's Moonshot AI, which has raised concerns in the U.S. about the potential disruption of domestic AI companies’ business models, including leaders like OpenAI and Anthropic. There is apprehension that Chinese open-source AI could undermine these companies’ fundraising capabilities for next-generation AI development.

Reports indicate that the previous U.S. administration considered a complete ban on Chinese open-source AI models within the United States. While the current administration’s stance remains uncertain, Bessent’s remarks suggest a possible regulatory shift from focusing solely on AI chips and data center equipment to scrutinizing AI models themselves.

U.S. AI firms have repeatedly expressed concerns that foreign competitors might exploit their models’ outputs and underlying technology to create and distribute similar open-source models. The White House announced in April its intention to collaborate with U.S. AI companies to address such technology theft.

A central technical issue in this dispute is "model distillation," a process that transfers the knowledge and performance of a large AI model to a smaller, more efficient one. Although widely used in AI development, the industry remains divided on whether distillation that leverages outputs from competitor models constitutes intellectual property infringement.

Microsoft CEO Satya Nadella publicly challenged this distinction earlier in July, highlighting a contradiction between allowing training on public data under fair use and imposing restrictions on distillation using trained models.

Meanwhile, U.S. AI companies themselves face scrutiny over training data and copyright compliance. For instance, Anthropic recently gained court approval to initiate compensation payments to authors under a $1.5 billion settlement after being found to have illegally downloaded and stored millions of copyrighted books for AI training purposes.

Countering claims that Chinese AI firms’ rapid progress is solely due to model distillation, Hugging Face CEO Clem Delangue stated in a recent podcast that distillation is just one part of AI development and is employed by most AI companies globally, including those in the U.S. He argued that if distillation alone enabled superior models, more high-performance open-source models would already exist worldwide.

Delangue also noted that China’s competitiveness in AI is bolstered by a large pool of talented researchers and a more open, collaborative research culture compared to U.S. companies.

Should the U.S. government proceed with sanctions against Chinese AI firms, the AI competition between the two countries is expected to extend beyond semiconductor supply chains and export controls into complex disputes over AI model training methodologies and intellectual property rights. How the U.S. distinguishes between legitimate model distillation and unlawful technology theft will be a pivotal issue moving forward.