U.S. Threatens Sanctions Over Chinese AI Models as IP Fight Escalates

The Trump administration is warning that Chinese AI models could face U.S. sanctions if investigators determine they were built using stolen intellectual property, a move that would escalate Washington’s fight with Beijing from chips and exports to the models themselves. Treasury Secretary Scott Bessent said the U.S. will scrutinize open source systems from China and act if it finds evidence of theft, a signal that could reshape the global AI competition.

The comments arrive as Chinese open source models grow more capable and more widely used, raising pressure on leading American AI firms such as OpenAI and Anthropic. They also come amid a broader policy shift in Washington, where officials are looking for new ways to protect U.S. companies’ competitive edge in frontier AI.

What the Treasury Secretary said

Treasury Secretary Scott Bessent said on Fox Business that the administration supports open source AI in principle, but will not tolerate what it views as copying that crosses the line into intellectual property theft.

“We’ve seen a lot of talk about open source models coming and threatening the large language models in the US,” Bessent said. “This administration supports open source models, but what we do not support is IP theft. If we see, especially, that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft.”

Bloomberg first reported the remarks. The message was notable not only for its substance, but for its potential to broaden U.S. enforcement beyond traditional export controls and into the behavior of AI systems and the companies that build them.

Why this matters for the AI race

The threat matters because it suggests Washington may be preparing to treat AI model development as a national security issue on par with semiconductors. For years, the U.S. has limited China’s access to advanced chips and tightened export restrictions aimed at slowing Chinese progress in frontier computing. Sanctions tied to model theft would be a more direct attempt to police how AI capabilities are built, copied and distributed.

That would be especially significant if the government concludes that Chinese firms have used distillation, data extraction or other techniques to reproduce the performance of U.S. systems. The policy could affect not just individual vendors, but also the broader open source ecosystem that has helped make AI more accessible worldwide.

How are Chinese models changing the competitive landscape?

Chinese open source models are becoming stronger and more popular, and that is putting pressure on the commercial outlook for top U.S. labs. One of the latest examples is Moonshot AI’s Kimi K3, which has drawn attention for its capabilities and helped fuel the sense that Chinese companies are closing the gap.

For American AI leaders, the concern is not limited to benchmark performance. If high-quality open source models become easier to access and adapt, they could reduce the scarcity that supports premium pricing, enterprise contracts and repeated fundraising rounds for frontier research.

• More capable Chinese models can erode the perceived moat around U.S. labs.
• Open source distribution makes those models easier to adopt and modify.
• Price pressure could complicate fundraising for expensive frontier development.

What is model distillation, and why is it controversial?

Model distillation is a technique that transfers behavior from a larger AI system into a smaller one that is cheaper and easier to run. In practice, it can make advanced capabilities more widely available, which is one reason the method is so attractive to developers.

But distillation is also at the center of a growing dispute. Some U.S. companies argue that if competitors use outputs from proprietary systems to train near-copycat models, that amounts to theft. Others say distillation is a legitimate engineering practice and an established part of the AI research toolkit.

The disagreement is important because any sanctions decision would likely require the government to decide where legitimate learning ends and unlawful appropriation begins. That is a difficult line to draw, particularly when AI systems are trained on massive data sets and their behavior often reflects a mix of public information, synthetic outputs and iterative refinement.

Why some industry leaders disagree

Not everyone in the AI sector believes distillation is the main explanation for China’s progress. Some argue that the technology is overblown as a competitive factor and that Chinese labs are advancing for deeper reasons, including research quality and organizational culture.

Hugging Face chief executive Clem Delangue said distillation is only a minor part of the story and argued that China’s AI teams are strong because they are highly capable and work in a more open, collaborative environment.

That view pushes back against the idea that Chinese firms are simply reverse-engineering American breakthroughs. It also suggests that sanctions alone may not slow the broader pace of Chinese AI development if the underlying research capacity continues to improve.

How U.S. policy is shifting

The threat of sanctions is part of a larger policy pattern in Washington: keep the U.S. ahead in AI by limiting foreign access to key inputs and by scrutinizing the methods used to create powerful models. The strategy now appears to be moving from hardware bottlenecks to software and model governance.

In April, the White House said it would work closely with AI companies to address model theft concerns. Since then, those warnings have been reinforced by public comments from officials and executives who see proprietary model imitation as both a business risk and a security problem.

Axios reported on Monday that the Trump administration is considering a broad ban on Chinese open source models, although that claim has been disputed by others. Bessent’s remarks, however, show that the debate is active at the highest levels of government.

Legal risks are already mounting for AI companies

The industry is already facing major legal exposure over training data, copyright and storage practices. This week, Anthropic received approval to begin issuing payments to authors as part of a $1.5 billion settlement after a judge found it had illegally downloaded and stored millions of books for AI training.

That case underscores how quickly legal disputes can turn expensive when courts conclude that training practices crossed a line. It also gives added context to the government’s willingness to scrutinize how model capabilities are assembled and whether those capabilities are derived from unauthorized material.

In that environment, a sanction-based response to alleged IP theft would add another layer of pressure. Rather than leaving disputes to civil litigation alone, the U.S. could move to isolate offending companies through financial and trade measures.

What would sanctions actually change?

Sanctions would not just be symbolic. If imposed, they could restrict access to U.S. financial systems, bar certain business relationships and create a chilling effect for investors, cloud providers and distribution partners that touch the targeted companies.

For Chinese AI firms, even the threat of sanctions could complicate global expansion plans. It could also make open source releases more politically sensitive, since the act of publishing a model for public use might be interpreted in Washington through a national security lens rather than a purely technical one.

At the same time, enforcement would likely be difficult. Determining whether one model was trained by copying another is technically complex, and companies often release partial details about their pipelines. Any sanctions case would probably need evidence that holds up under scrutiny, not just broad allegations of similarity.

Why proving theft is so hard

AI systems can look alike for many reasons. They may be trained on similar public data, optimized using comparable architectures or fine-tuned for similar tasks. Strong performance by itself does not prove copying.

That means regulators would need to distinguish between:

1. shared access to public training material,
2. lawful use of open source code and models,
3. distillation from outputs, and
4. direct misappropriation of protected material or proprietary systems.

That distinction is exactly why the issue has become so contentious. The more capable and widely distributed open source models become, the harder it becomes to separate inspiration from appropriation.

What this means for OpenAI, Anthropic and rivals

For OpenAI, Anthropic and other U.S. leaders, the administration’s stance is a sign that policymakers are listening to their warnings about foreign imitation. These companies have spent heavily to build frontier models and are seeking the capital needed to keep pace with rising compute costs.

If cheaper Chinese open source systems continue gaining traction, that could undercut the argument that only a few U.S. companies can justify the enormous spending required to stay ahead. It could also change enterprise buying behavior, especially if customers decide that powerful open source alternatives are “good enough.”

The stakes are not merely commercial. In Washington’s view, frontier AI leadership carries strategic value, from economic influence to military and intelligence applications. That helps explain why the government is now speaking more forcefully about model theft than it did just months ago.

Timeline of the escalating dispute

The current clash did not emerge overnight. It has built over months of rising model capability, legal fights and political pressure. The following timeline captures the key developments that frame the latest warning.

Could this become a broader crackdown?

Yes, and that possibility is what makes the latest remarks so consequential. If the administration concludes that Chinese open source models are routinely derived from U.S. systems in ways it considers improper, the result could be a wider campaign involving sanctions, import restrictions, cloud controls or new disclosure requirements.

That would align with Washington’s broader effort to preserve a lead in AI by shaping the environment around development, deployment and access. It would also deepen the already sharp divide between a U.S. industry built around proprietary frontier labs and a Chinese ecosystem that has increasingly embraced open distribution.

Still, the policy path is far from settled. Public comments can signal intent without guaranteeing action, and any real crackdown would likely face technical, legal and diplomatic complications. What is clear is that the AI competition is entering a more confrontational phase.

Bottom line

The U.S. is now openly considering sanctions against Chinese AI companies if their models are found to involve intellectual property theft, marking a new escalation in the technology rivalry with China. The move could reshape how governments regulate AI competition and how companies define the line between open collaboration and unlawful copying.

For the global AI industry, the message is unmistakable: model development is no longer just a business story. It is becoming a geopolitical one.