China Uses U.S. AI Models to Build Defense Systems

The Federal government of the United States and leading AI firms have accused Chinese military and commercial entities of using model distillation to unauthorizedly extract capabilities from proprietary artificial intelligence models. This technique involves using a large teacher model to train a smaller student model, allowing China to replicate advanced reasoning and software engineering capabilities while bypassing U.S. export controls on high-end chips.

A review of over 80 academic papers and patents indicates that researchers linked to the People's Liberation Army use outputs from OpenAI and Anthropic to develop specialized systems for cyber warfare, surveillance, and tactical decision-making. Specific applications include using GPT-3.5 to process military source code and Claude 3 Haiku for social media monitoring, as well as enhancing drone navigation and maritime target recognition.

Anthropic specifically accused Chinese firms DeepSeek, Moonshot, and MiniMax of running large-scale campaigns to harvest reasoning traces from its Claude models. Moonshot denied that its Kimi K3 model was built using these methods. The Government of China has rejected all allegations of intellectual property infringement, characterizing Washington's accusations as AI hegemonism. These disputes emerge as a primary point of contention ahead of scheduled U.S.-China talks on AI governance and safety.