The White House has accused Chinese AI labs of systematically stealing U.S. Technology through distillation attacks, according to an internal memo reviewed by the BBC.
The memo, circulated earlier this year, alleges that three Chinese firms — DeepSeek, Moonshot and MiniMax — used model distillation techniques to replicate proprietary AI systems developed by Anthropic, a leading U.S. Artificial intelligence company. Distillation involves training a smaller model to mimic the behavior of a larger, more powerful one, often by querying the target model repeatedly and using its outputs as training data.
How distillation enables AI model copying
Anthropic told the BBC it detected coordinated efforts by the three labs to extract knowledge from its models via application programming interfaces, then rebuild comparable systems without direct access to the original architecture or training data. This method allows competitors to bypass expensive development cycles by leveraging the front-runner’s investments in research and compute.
The technique is not illegal per se, but becomes problematic when it violates terms of service or involves unauthorized access to proprietary systems. Anthropic said it observed patterns consistent with coordinated scraping and querying designed to evade detection.
Why the White House is involved
The memo elevates what had been a private dispute between companies into a matter of national economic security, reflecting growing U.S. Concern over China’s rapid advances in AI capabilities. Officials worry that illicit technology transfer could erode America’s competitive edge in foundational AI models, which underpin everything from chatbots to autonomous systems.
This echoes concerns raised during the semiconductor trade disputes of the early 2020s, when the U.S. Imposed export controls on advanced chips to slow China’s AI progress — measures that are now being expanded to cover AI model weights and training data.
What happens next
The memo does not announce sanctions or legal action, but signals that the administration is reviewing whether existing export controls or cybersecurity laws can be applied to AI distillation campaigns. Experts say proving intent and attribution remains difficult, especially when activity is routed through third-party servers or cloud providers.
For now, the allegation adds pressure on U.S. AI firms to strengthen access controls and monitor API usage for anomalous patterns — a costly but necessary defense in an era where model weights are as valuable as source code.
What is model distillation in AI?
Model distillation is a technique where a smaller AI system is trained to imitate the outputs of a larger, more complex model, allowing it to achieve similar performance with far less computational cost.
Is distillation illegal?
Distillation itself is not illegal, but using it to copy a model without permission may violate the target company’s terms of service or constitute unauthorized access to proprietary systems.
