Chinese Military Advances Tech Using Distilled US AI Models

According to a report by Reuters on July 31 (local time), Chinese military and security researchers are reportedly advancing the development of domestic AI systems through 'model distillation,' a technique that utilizes the output results of advanced models from U.S. AI firms. Model distillation is a technology that trains smaller-scale AI models using the output results of high-performance large AI models as training data. This allows for the construction of systems optimized for specific purposes without the need to directly develop advanced AI models that require massive computational resources.

Reuters reported, citing an analysis by Zhang Kunyang, a researcher at the Washington-based Jamestown Foundation and an analysis of approximately 80 Chinese academic papers and patent documents, that Chinese military research institutions are developing AI systems for defense and military purposes—such as surveillance, cyber warfare, and tactical decision-making—using this method.

The practice of extracting model performance without the original developer's permission has been highlighted as a problematic issue during this process. The U.S. Treasury Department indicated that such practices could face sanctions and export controls if they involve intellectual property infringement.

To date, the U.S. White House, the Department of Defense, China's Ministry of Foreign Affairs, and the People's Liberation Army have not released official statements in response to Reuters' request for comment.