Overview
- The White House, in a memo Thursday, accused entities based mainly in China of mounting large campaigns to copy capabilities from U.S. AI models.
- Officials said the effort used tens of thousands of proxy accounts and jailbreak tricks to pull proprietary answers, a tactic used to train copycat systems through a process called distillation.
- U.S. labs have reported similar activity, with Anthropic saying DeepSeek, Moonshot AI, and MiniMax made over 24,000 fake accounts for 16 million prompts, and OpenAI alleging in February that DeepSeek covertly copied its models.
- The administration said it will share intelligence with U.S. companies and will explore ways to hold foreign actors to account.
- Pressure on policy grew as Commerce Secretary Howard Lutnick said no Nvidia AI chips have shipped to China after January approvals, while DeepSeek’s new V4 model launched Friday with a claimed one‑million‑token context and AI companies stepped up lobbying to shape U.S. and EU rules.