White House Criticizes China's Industrial-Scale AI Model Distillation, Pledges Information Sharing

The White House's OSTP has characterized China's AI model distillation as 'industrial-scale' theft and announced it will strengthen a system to share information on foreign attempts together with OpenAI, Anthropic, and Google. This message came ahead of the Trump-Xi Jinping summit scheduled for May 14, elevating AI technology protection to a security agenda item.

Distillation is a method of mass-querying frontier models and using the responses as training data, without stealing model weights. According to the article, OpenAI informed Congress in February that DeepSeek bypassed restrictions to extract outputs, and Anthropic disclosed that DeepSeek, MiniMax, and Moonshot AI interacted with Claude more than 16 million times using approximately 24,000 fake accounts.

Congress is also taking action. The Deterring American AI Model Theft Act (H.R. 8283) has been introduced, and the House Select Committee on China held a hearing on campaigns to steal China's AI advantage. However, since this memo is a policy statement rather than an executive order or regulation, actual sanctions will depend on follow-up measures and enforcement tools.

The White House's judgment is that hardware controls alone are not sufficient.

• Chip export controls have already been tightened multiple times, but chip smuggling and China's development of domestic replacement chips have continued.
• Models with open weights, like open-source Llama, are difficult to control and could be retrained for military and intelligence purposes.
• Ultimately, the U.S. is shifting to a multi-layered defense that combines chips, models, and tracking together.