Treasury's Bessent warns of sanctions over Chinese AI models built on alleged U.S. IP theft
The Trump administration is reviewing evidence that Chinese AI systems were trained through distillation of American models, with Anthropic accusing Alibaba of the largest known such attack.
What matters
- Treasury Secretary Bessent said the U.S. could sanction Chinese AI models if IP theft through distillation is confirmed.
- Anthropic accused Alibaba of carrying out the largest known distillation attack against its models.
- U.S. officials claim to have found digital watermarks from American LLMs in several Chinese AI systems.
- The U.S. and China are expected to hold AI-related talks in September.
- U.S. AI firms including Anthropic and OpenAI also face their own copyright and IP lawsuits over training practices.
What happened
Treasury Secretary Scott Bessent said the Trump administration is examining whether Chinese artificial intelligence models were developed by improperly using American AI systems, warning that sanctions could follow if intellectual property theft is confirmed.
According to CNBC, Bessent said the government opposes the unauthorized use of U.S. technology and is reviewing evidence that Chinese models may have been created through AI "distillation"—a process that trains smaller models using outputs from more advanced ones. Bessent also claimed U.S. officials have identified digital watermarks from American large language models embedded in several Chinese AI systems.
The scrutiny follows a recent accusation by Anthropic that Chinese technology company Alibaba carried out what Anthropic described as the largest known distillation attack against its models. The claim adds a concrete corporate flashpoint to what has been a broader geopolitical concern about whether Chinese AI developers are free-riding on the costly training work of U.S. frontier labs.
The United States and China are expected to hold AI-related talks in September, according to the report.
Why it matters
Distillation is a well-known and technically legitimate technique in machine learning, but it becomes contentious when a model's terms of service prohibit using its outputs to train competing models. If the U.S. government determines that Chinese AI systems were built on improperly obtained American model outputs, sanctions could target not just specific companies but the broader ecosystem of Chinese open-weight models that have been gaining global adoption.
The stakes are high for the open-source AI community. Chinese open-weight models have become widely used by developers worldwide precisely because they are freely available. Sanctions or export restrictions could disrupt downstream users, cloud providers, and integrations that have come to depend on these models.
There is also an irony worth noting: leading U.S. AI companies including Anthropic and OpenAI have themselves faced copyright and intellectual property lawsuits over their own model training practices. The administration's framing of model-output reuse as IP theft could complicate the legal landscape for all AI developers, not just Chinese ones.
What to watch
- Whether the September U.S.–China AI talks produce any framework or agreement on model-output usage and distillation.
- Whether sanctions, if imposed, target specific companies like Alibaba or extend to entire categories of Chinese open-weight models.
- How frontier labs like Anthropic and OpenAI enforce their terms of service against distillation at scale, and whether technical watermarks become a standard enforcement tool.
- Whether U.S. IP lawsuits against domestic AI companies influence the government's credibility in pressing the distillation-theft case internationally.
What to do next
Developers
Audit any Chinese open-weight models in your stack for licensing and provenance risk, and document where model outputs or weights originate.
If sanctions target specific Chinese models, downstream users may face compliance exposure or sudden access cutoffs.
Founders
Review your AI provider agreements and terms of service to ensure your use of third-party model outputs for training or fine-tuning is explicitly permitted.
The distillation debate is intensifying, and ambiguous licensing could become a legal liability as enforcement tightens.
PMs
Map which features in your product depend on Chinese open-weight models and identify fallback providers in case of sanctions or access restrictions.
Supply-chain resilience matters when geopolitical policy can abruptly restrict model availability.
Investors
Assess portfolio exposure to Chinese open-weight model dependencies and to companies whose competitive moat relies on output-reuse terms being enforceable.
Sanctions or new IP frameworks could reshape the competitive landscape for both Chinese and U.S. AI companies.
Operators
Inventory all AI services and model endpoints your organization consumes, flagging any Chinese-sourced models for legal and compliance review.
Regulatory action could convert a procurement choice into a compliance issue with little warning.
Testing notes
Caveats
- This is a policy and geopolitical story, not a product or model release. There is no software, API, or tool to test. The situation may evolve depending on the outcome of the September U.S.–China AI talks and any subsequent sanctions decisions.