Chinese military researchers distilled US AI models to build defense, surveillance tools
Researchers linked to China's PLA are extracting OpenAI and Anthropic model outputs to build defense and surveillance tools, per a Reuters review of 80+ academic papers and patents. One team distilled Claude into a social media monitoring classifier. What to watch: upcoming US-Ch
Researchers tied to China's People's Liberation Army are using outputs from US AI models to build their own defense and surveillance systems, according to a Reuters review of more than 80 Chinese academic papers and patents, corroborated by research from the Jamestown Foundation published July 30. The technique, known as distillation, trains smaller domestic models on the outputs of larger US systems like OpenAI's GPT-4o and Anthropic's Claude, letting Chinese labs approach frontier performance without the computing cost of training from scratch. Researchers at North University of China, an institution rooted in the ordnance industry, distilled Claude into a classifier proposed for social media monitoring and content moderation. Researchers at Nanjing University of Science and Technology, one of China's military-linked "Seven Sons of National Defense," tested hiding prompt-injection commands inside images of tanks and warships to see if GPT-4o and Claude would obey them. A separate Xidian University team working with a People's Armed Police counterterrorism lab used OpenAI's InstructGPT to generate and manually correct training data. US Treasury Secretary Scott Bessent has said the US government found watermarks from American AI models embedded in Chinese models. The issue is now a flashpoint ahead of planned US-China talks on AI governance and safety.