US accuses China of industrial-scale AI model distillation, will share Intel
5 hours ago
- #AI Security
- #US-China Relations
- #Intellectual Property
- The White House OSTP accused China of 'industrial-scale' distillation of US AI models, pledging intelligence sharing with US companies and accountability measures.
- OpenAI and Anthropic provided evidence: OpenAI accused DeepSeek of circumventing restrictions, while Anthropic named DeepSeek, MiniMax, and Moonshot AI for creating fraudulent accounts and generating millions of exchanges with Claude.
- The Deterring American AI Model Theft Act was introduced to impose sanctions on entities using improper query-and-copy techniques, with bipartisan support in Congress.
- Distillation is a technique that uses queries to train rival models, raising unresolved legal questions about whether extracted outputs qualify as trade secrets.
- US policy shifts from hardware controls (like chip restrictions) to model-level protections due to smuggling issues, emphasizing defense in depth through chip and model controls.
- Open-source models, such as Meta's Llama, complicate security as they enable military applications by adversaries despite use policy restrictions, highlighting policy tensions.
- The upcoming Trump-Xi summit will test the OSTP memo's impact, with AI protection treated as national security amid wide gaps between US and Chinese positions.