US Agencies Accuse Six Chinese AI Firms of Distilling American Models

Model distillation—training a smaller system on outputs from a more capable one—is widely used to cut costs, but extracting a rival’s capabilities at scale through deceptive access may breach service terms and weaken export controls. On Feb. 23, 2026, Anthropic said DeepSeek, Moonshot AI and MiniMax used about 24,000 fraudulent accounts to generate more than 16 million exchanges with Claude, allegations that sharpened Washington’s concerns over intellectual property and national security.

U.S. Treasury Secretary Scott Bessent said on July 21 that the Trump administration would examine Chinese models over the coming days or weeks and could sanction firms if intellectual-property theft is established. On July 22, White House Office of Science and Technology Policy Director Michael Kratsios accused Moonshot AI of distilling Anthropic’s Fable to develop K3. Officials are also weighing disclosure requirements for U.S. companies using Chinese AI models; no penalty amounts or final rules have been announced.

Chinese startup DeepSeek uses “knowledge distillation” to train smaller models on the outputs of large models, reducing computing and development costs and placing it at the center of disputes over US-China AI leadership and intellectual property. DeepSeek-V3 was released in December 2024, with its technical report estimating pretraining compute costs at about $5.576 million, drawing scrutiny from US officials.

On April 23, 2026, White House Office of Science and Technology Policy Director Michael Kratsios issued a memorandum accusing Chinese entities of using proxy accounts and jailbreaking techniques to distill US models. Anthropic said in February that DeepSeek, Moonshot AI and MiniMax had used about 24,000 fraudulent accounts to conduct 16 million interactions. The State Department followed with a worldwide cable on April 24 calling for efforts to detect and defend against such activity and hold those responsible accountable.

Model distillation, a common technique, uses the outputs of powerful models to train smaller ones. The controversy arises when competitors use fake accounts to extract capabilities from closed models at scale, potentially violating terms of service. On February 23, 2026, Anthropic accused DeepSeek, Moonshot AI and MiniMax of using about 24,000 fake accounts to interact with Claude more than 16 million times, making model capabilities a new front in the US-China technology and national security rivalry.

Bill Huizenga and John Moolenaar introduced H.R. 8283 on April 15, 2026. The US House Foreign Affairs Committee approved the bill on April 22 and sent it to the full House. The legislation would establish a public list and could place companies including DeepSeek, MiniMax and Alibaba on the Commerce Department's Entity List. It would also authorize asset freezes under the International Emergency Economic Powers Act of 1977. The Congressional Budget Office estimates that the measure would require $35 million in appropriations from 2026 through 2031.

Model distillation is intended to let smaller models learn from larger “teacher models,” reducing training costs. The controversy arises when third parties harvest outputs at scale without authorization, potentially replicating proprietary capabilities and removing safety guardrails. OpenAI, Anthropic, Google and Microsoft established the Frontier Model Forum on July 26, 2023. Their cooperation is important to protecting the results of hundreds of billions of dollars invested in data centers and research and development.

On April 6, 2026, OpenAI, Anthropic and Google shared intelligence through the Frontier Model Forum to identify adversarial distillation that violated their terms of service. In February, Anthropic accused DeepSeek, Moonshot and MiniMax of using 24,000 fake accounts to conduct 16 million interactions. U.S. estimates indicate that such activity costs the industry billions of dollars in lost profits each year.