Bessent Raises Sanctions Prospect Over Chinese AI Models

Bessent said the administration supports open-source AI but does not support intellectual-property theft. He said US authorities were finding "watermarks" from American large language models in Chinese systems and would examine the matter in the coming days or weeks. He also raised the possibility that companies might have to disclose their use of Chinese AI models.

The remarks did not name a company or model as a sanctions target. They also did not disclose the alleged watermarking method, technical evidence, legal standard, investigating agency or timetable for a formal decision. No sanctions action was announced. The claim should therefore be treated as an allegation and policy warning, not as an established technical finding.

Model distillation is a broad family of techniques in which one system's outputs help train or improve another. Whether a particular practice violates intellectual-property law, contractual terms or access controls depends on the facts. Evidence would need to distinguish ordinary use, permitted synthetic-data generation and benchmarking from coordinated extraction or prohibited reconstruction.

That distinction is especially important for open-weight models. Their availability does not by itself reveal how their training data was obtained or whether a developer violated another provider's rules.

If the administration develops a formal enforcement policy, reproducible evidence will matter more than broad similarity claims. Model providers should retain API abuse telemetry, account-linkage evidence, rate-limit events, watermark test methods and versioned model outputs. Developers using third-party models should preserve licenses, vendor terms, data-lineage records and evaluation inputs.

Those records can support a technical review without assuming the outcome. For now, the material development is that a senior US official placed sanctions on the list of possible responses; the evidentiary and legal process remains unresolved.