China's Military Distills U.S. AI Models for Defense Systems

A Reuters investigation of more than 80 Chinese academic papers and patent filings found that researchers linked to the People’s Liberation Army have been using a technique called model distillation to train domestic artificial‑intelligence systems with the outputs of frontier U.S. models from OpenAI and Anthropic. By feeding the responses of large “teacher” models such as OpenAI’s GPT‑3.5 into smaller “student” models, Chinese teams can run advanced AI on modest hardware for applications that include surveillance, cyber‑operations, drone navigation, target identification and tactical decision‑making. The review cites specific instances, such as PLA unit 96941 employing GPT‑3.5 to process military code and the Defense University deploying distilled vision models on unmanned aircraft that can operate offline. U.S. officials argue the practice may evade export‑control rules and infringe intellectual‑property rights, while Beijing rejects the accusations and criticises U.S. “AI hegemony.” Recent Chinese releases, including Alibaba’s Qwen 3.8 and Moonshot AI’s Kimi K3, demonstrate the country’s growing frontier‑level AI capabilities.

The findings highlight an accelerating AI competition between the United States and China, raising concerns over technology transfer, export‑control effectiveness and the militarisation of advanced machine‑learning tools.