The U.S. government is extending its restrictions on China’s AI sector from chips and equipment to the model level. On July 21, U.S. Treasury Secretary Bessent stated that the U.S. will review Chinese open-source AI models for signs of intellectual property theft; if such activities are confirmed, the U.S. has the authority to impose sanctions on Chinese AI companies.
On that day, Bessen stated on Fox Business that the U.S. government supports the development of open-source models but does not tolerate intellectual property theft, particularly if overseas models are found to have "stolen"成果 from U.S. companies, in which case the U.S. could impose sanctions accordingly. Bloomberg subsequently broke the story.
The U.S. White House continues to expand its pressure on China.
This statement emerges against the backdrop of the United States continuously tightening its AI policies toward China. Previously, Washington had restricted China’s access to advanced AI chips and steadily strengthened export controls. Now, signals from the U.S. suggest that future targets may extend beyond hardware and supply chains to include AI models themselves.
Axios previously reported that the Trump administration is considering a complete ban on Chinese open-source models, though this claim has been questioned by some. Although no such measures have been formally implemented, recent statements indicate that the U.S. government is indeed discussing tougher options.
The focus of the dispute lies in model distillation.
One central aspect of the discussion surrounding so-called "theft" is model distillation—a technique commonly used to transfer part of a large model’s capabilities to smaller, more deployable systems. In recent months, several U.S.-based AI companies have repeatedly warned that foreign actors may replicate their technology and re-release it as open source.
However, the industry does not uniformly agree that “distillation equals theft.” Microsoft CEO Nadella recently questioned this claim, pointing out that it is contradictory for large model companies to assert their right to train models on publicly available data while imposing strict restrictions on distillation.
U.S. AI companies also face copyright risks.
Intellectual property disputes are not limited to Chinese companies. U.S. AI firms themselves are also facing ongoing legal pressure over their training practices. This week, Anthropic was granted approval to begin payouts under its $1.5 billion settlement, after a judge ruled that the company had illegally downloaded and stored millions of copyrighted books to train its AI.
Some industry insiders also believe that China’s progress in AI relative to the U.S. cannot be simply attributed to distillation. Hugging Face CEO Clem Delangue recently stated that distillation is only a minor factor, as American companies themselves are also using this approach. He believes that China’s faster advancement is more significantly due to its stronger research teams and a more open, collaborative approach to AI development.
If the United States ultimately formally applies sanctions to Chinese AI models, the focus of U.S.-China technological competition will further expand from chips to algorithms and model ecosystems, and the compliance and business pressures on related companies may also rise accordingly.