White House draws new AI line on China

The Trump administration this week staked out a new position that could reshape its approach to AI policy, accusing Chinese companies of stealing U.S. technology even as it defends open AI development.

Why it matters: That distinction could become the foundation for U.S. action against Chinese AI companies without broadly restricting the open-weight models that businesses, researchers and startups increasingly rely on.

Driving the news: The Trump administration is backing open-weight models while accusing China's Moonshot of using distillation — the process of training a smaller AI model using the outputs of a larger model — to copy Anthropic's Fable model.

• "Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem," Michael Kratsios, chief technology adviser to President Trump, wrote on X.
• However, "large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable," he added.
• Treasury Secretary Scott Bessent, U.S. Trade Representative Jamieson Greer and State under secretary Jacob Helberg all made similar claims of intellectual property theft.

What they're saying: OpenAI co-founder Greg Brockman told Bloomberg Wednesday it was too early to tell if recent Chinese models had been distilled off GPT systems.

• In a briefing with reporters, Brockman told Axios he had not been involved in any conversation with members of the administration about banning Chinese open models.

Between the lines: The administration is effectively defining two categories of distillation.

• Small-scale or authorized distillation that produces cheaper, specialized models is being treated as a normal part of AI development.
• Industrial-scale distillation conducted through fraudulent accounts, evasive access techniques or violations of U.S. companies' terms is being framed as theft.
• When asked whether distillation is a technical or policy problem, Brockman said it's "definitely a technical issue."

The big picture: The advent of Moonshot's Kimi has riled U.S. policymakers and some industry players. The model is cheaper, and U.S. companies are increasingly opting for these alternatives.

• The Trump administration could be laying the groundwork for action against potential AI IP theft via distillation.

Zoom in: Distillation is a common AI practice.

• Smaller models that act as students can be distilled from larger models that act as teachers, replicating the teacher model's behavior without needing to go through the same training.
• That compact model can then run more efficiently while producing outputs similar to the larger model, giving the user access to similar performance that's quicker and cheaper.

The intrigue: The U.K. Artificial Intelligence Security Institute and the U.S. Center for AI Standards and Innovation on Thursday released a joint evaluation of Kimi K3's cyber capabilities, determining it performs significantly below other frontier models.

What we're watching: The administration's comments suggest the White House wants to preserve open AI development while taking a hard line on China as major AI labs warn of misuse and open source players emphasize the need for competition.

• The administration has previously considered adding Chinese AI labs to the Commerce Department's Entity List to address risks from open-source models, as Axios first reported, a move that would effectively cut off U.S. access without a license.
• This week Bessent made clear in a post on X the administration is willing to flex those muscles: "When PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table."