China’s military tapped U.S. AI to build battlefield systems: report

Chinese Military Researchers Use Distillation Technique on U.S. AI Models, Review Finds

Chinese military researchers have applied a technique known as model distillation to adapt outputs from U.S. artificial intelligence systems for domestic military applications, according to a Reuters review of academic papers compiled by the Washington-based Jamestown Foundation. The review examined more than 60 studies and identified additional examples through independent verification.

Model distillation involves training smaller AI systems to replicate selected capabilities of larger models. The method is standard in commercial AI development for creating efficient versions that require less computing power. Reuters reported that the papers linked researchers affiliated with China’s People’s Liberation Army and defense institutions to the use of this approach with models from OpenAI and Anthropic.

One study from PLA Unit 96941 described feeding military software code into OpenAI’s GPT-3.5 model and then transferring the resulting analysis into a Chinese-developed model intended to function on domestic military networks without external connections. Other papers detailed the use of Anthropic’s Claude model to produce training data for systems focused on social media monitoring and content moderation.

Additional research outlined applications for battlefield systems. A paper from the PLA’s National University of Defense Technology explained the use of distillation to reduce the size of an image-recognition model so it could analyze live video on unmanned aerial vehicles without continuous external communication. Researchers at the Academy of Military Sciences reported testing distilled models in simulated operations involving drones, warships, and unmanned submarines.

Sunny Cheung, a Jamestown Foundation fellow who conducted the initial analysis, stated that the papers indicate efforts to transfer reasoning capabilities from Western models into locally controlled systems. Reuters confirmed the institutional affiliations and located further examples beyond the original compilation.

Anthropic stated that it does not provide commercial access to Claude in China or to entities controlled by the Chinese government and maintains detection systems for policy violations. The company noted that distilled models may retain capabilities while losing some safety measures originally present in the source model. OpenAI did not respond to Reuters’ request for comment on the specific papers.

The reviewed studies occur against the backdrop of U.S. export controls on advanced semiconductors and Chinese government investment in compact AI models suitable for devices with limited processing capacity, such as drones and satellites. Trevor Koverko, co-founder of AI data company Sapien, said distillation transfers selected functions into lower-cost systems but does not replicate the full performance of frontier models.

The White House, the Pentagon, China’s Foreign Ministry, and the People’s Liberation Army did not respond to Reuters’ inquiries. The Jamestown Foundation is a research organization focused on China-related security issues; its compilation formed the starting point for the Reuters examination, which added dozens of independently identified papers.

The practice of distillation itself remains legal and common in the broader AI sector. U.S. officials have raised concerns that the technique could circumvent export restrictions on advanced AI capabilities, while Chinese authorities have described U.S. export controls as attempts to preserve technological leadership. The reviewed papers do not indicate direct commercial access to the U.S. models by the Chinese military; they describe the use of model outputs obtained through various means to train separate systems.