China rejects U.S. accusations of large-scale AI technology theft

What This is

• The FBI, National Security Agency, and Cybersecurity and Infrastructure Security Agency issued a joint advisory accusing China-based AI firms of large-scale “distillation” of U.S. AI models.
• The advisory and related reporting named Chinese companies including DeepSeek, Alibaba, Moonshot AI, Z.ai, MiniMax, and StepFun as targets of the allegations.
• U.S. officials say the firms pulled huge volumes of model outputs—described as billions of tokens—since late 2024 to help train or improve their own systems.
• U.S. officials describe the method as “knowledge distillation,” where a model is improved by learning from another model’s outputs rather than copying its code or weights.
• Anthropic says several China-based labs and other foreign actors used illicit distillation against its Claude models, including millions of fraudulent exchanges.
• U.S. officials warned the alleged copying could reduce development costs for Chinese firms and raise national-security concerns, including possible military or cyber uses.
• U.S. officials signaled possible penalties, including sanctions or adding firms to trade restriction lists, while also urging defensive steps and information-sharing.

Why This Matters

• The dispute could accelerate a U.S.–China tech split, with tighter limits on data, chips, and access to top AI systems.
• If “distillation” becomes a common way to copy model behavior, AI companies may change how they sell access, monitor users, and protect their systems.
• New sanctions or trade restrictions could reshape global AI markets and raise costs for firms that rely on cross-border tools and customers.
• The fight also tests how governments define “theft” in AI, where copying can happen through outputs rather than stealing source code.