China PLA distilling US frontier AI models for military applications
PLA-linked researchers built defense AI by distilling outputs of US frontier models, per a Reuters review of 80+ Chinese papers and patents. Uses include drone targeting and processing sensitive military code. Distillation needs only outputs, not model weights - hard to block wit
A Reuters review of Chinese research has found that military-linked researchers in China have used AI distillation techniques to build defense systems from outputs of US frontier AI models, without needing direct access to the underlying model weights. More than 80 papers and patents point to work by researchers affiliated with the People's Liberation Army, China's defense industry and the Chinese Academy of Sciences. Documented applications range from monitoring social media content to drone targeting and maritime target recognition on tactical hardware. One example: a 2024 paper from the PLA's National University of Defense Technology described distilling an image-processing model small enough to run on drones, letting them analyze video and support targeting even when communications are cut. Separately, PLA Unit 96941, the military's main cyberwarfare group, distilled OpenAI's GPT-3.5 to build a lightweight model for processing sensitive military source code on Chinese networks. The White House Office of Science and Technology Policy alleged in April that China was running industrial-scale campaigns to distill US frontier systems. Because distillation only requires querying a model's outputs rather than accessing its weights, it sidesteps the export controls and access restrictions the US has used to limit China's access to frontier AI capability.