Overview
- The White House, in a memo Thursday from science adviser Michael Kratsios, alleged that entities based mainly in China used tens of thousands of proxy or fake accounts and jailbreak tricks to siphon capabilities from frontier U.S. AI models.
- Distillation, the method at issue, trains a smaller and cheaper system on answers from a larger model, which is common in AI but is alleged here to reproduce proprietary performance without permission.
- The administration said it will share threat intelligence with U.S. AI firms, coordinate defenses across companies, and develop ways to detect and deter large-scale distillation while exploring options to hold foreign actors accountable.
- China’s embassy and foreign ministry rejected the accusations as slander and said the country values intellectual property protection and scientific cooperation.
- Congress has moved in parallel as the House Foreign Affairs Committee advanced bills that direct the administration to consider placing groups that conduct such distillation on the U.S. Entity List, a trade blacklist that makes American tech sales far more difficult.