The Trump administration has carved out a new distinction in artificial intelligence policy this week, one that could allow it to target Chinese companies for IP theft without broadly cracking down on the open-weight models that startups and researchers depend on.
The flash point is distillation, a standard technique where smaller AI models learn from larger ones. Michael Kratsios, Trump's chief technology adviser, said the practice is vital to innovation when done at small scale. But when China's Moonshot deploys it at industrial scale to copy Anthropic's Fable model, that crosses into theft and becomes fair game for retaliation.
"Large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable," Kratsios wrote on X. Treasury Secretary Scott Bessent backed him up, signaling the administration is prepared to use sanctions and Entity List designations if Chinese firms engage in what it views as coordinated IP attacks.
The move reflects a growing frustration in Washington and Silicon Valley over Moonshot's Kimi model, which offers cheaper performance than U.S. alternatives and is winning over customers. But it also reveals the administration's effort to thread a needle: punish China without stifling the open-source innovation ecosystem that has become central to AI development across the industry.
OpenAI co-founder Greg Brockman offered a cautious note, telling reporters it remains unclear whether recent Chinese models were actually distilled from GPT systems. He also said he has not been involved in conversations with the administration about banning Chinese open models. When asked if distillation is a technical or policy problem, Brockman called it "definitely a technical issue," suggesting the answer lies in better protective measures rather than regulation.
The administration's framing sets up two categories of distillation. One is legitimate: smaller, specialized models created through authorized or transparent means, cheaper and faster than their sources. The other is industrial espionage: conducted through fraudulent accounts, unauthorized access techniques or violations of terms of service. That distinction gives the White House room to act against suspected Chinese theft without banning the open-model development that major AI labs and smaller competitors view as essential.
Treasury Secretary Bessent made clear in a separate post that the administration has tools ready. "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."
The timing matters. The U.K. Artificial Intelligence Security Institute and the U.S. Center for AI Standards and Innovation released a joint evaluation of Moonshot's Kimi K3 on Thursday, finding it performs significantly below other frontier models on cybersecurity tasks. That technical assessment may provide cover for any enforcement action the administration pursues.
What emerges is a strategy to contain China without dismantling open-source competition. The White House is preserving room for smaller startups and researchers to build on public models while preparing to move against what it characterizes as coordinated Chinese IP theft. Whether that distinction holds under pressure from U.S. companies seeking tighter restrictions on Chinese access to U.S. technology remains an open question.
Author James Rodriguez: "The administration is essentially betting it can isolate China without killing the open innovation that Silicon Valley claims to need, but that line will blur fast once enforcement begins."
Comments