The AI theft technique that has Washington alarmed

The US-China gap in top AI model performance has "effectively closed", according to a 2025 report from Stanford University's Institute for Human-Centred AI, and Washington is now asking how much of that progress was earned versus extracted.

In a memo released on Thursday, President Trump’s Chief Science and Technology Adviser Michael Kratsios blamed foreign actors principally based in China for launching campaigns designed to extract capabilities of top American AI systems. Distillation is accomplished by interrogating the targeted system extensively and then training an alternate model to imitate its functions.

Kratsios noted that the administration will collaborate with American AI firms to detect distillation operations, develop defensive tools to prevent them, and apply penalties to those behind them. This is the first time that a high-ranking official in the White House has addressed distillation as a problem of statecraft.

Both companies made their positions clear before Washington acted. In a February letter to US lawmakers, OpenAI accused China of advancing autocratic AI by appropriating and repackaging American innovation. Anthropic went further, naming DeepSeek and two other Chinese AI labs by name, accusing them of campaigns to "illicitly extract Claude's capabilities".

Acknowledging the potential legitimacy of distillation as a training method, Anthropic stated that the issue lies in the fact that "in a fraction of the time, and at a fraction of the cost, that it would take to develop them independently". Former AI and Crypto Advisor to Donald Trump, David Sacks, also made the same claim against DeepSeek's breakthrough AI model, R1.

The United States House of Representatives Foreign Affairs Committee recently voted unanimously to pass a bill creating an official process to identify foreign entities that exploit critical technical elements through the theft of closed-source US AI models, with possible sanctions. 

Republican bill sponsor Rep. Bill Huizenga of Michigan noted that model extraction attacks represent "the latest frontier of Chinese economic coercion".

China's embassy in Washington rejected the framing. Spokesperson Liu Pengyu said China "attaches great importance to the protection of intellectual property rights" and opposed what it called "unjustified suppression of Chinese companies".

A fellow at the Brookings Institution and an expert on Chinese technology development, Kyle Chan, said detecting unauthorised distillation will be like "looking for needles in an enormous haystack", given the volume of legitimate API requests that legitimate developers make every day. 

He said coordination among US AI labs, facilitated by the federal government, would be more effective than unilateral action.