On the eve of important security negotiations regarding artificial intelligence scheduled for September, new data has emerged regarding the methods China uses to develop its military technologies. The agency conducted a comprehensive analysis of over 80 Chinese scientific papers and patents, revealing the systematic use of advanced American AI models from OpenAI and Anthropic as a foundation for creating its own defense systems.
The "Distillation" Method as the Key to Technology
The primary tool used by the Chinese military and security-related institutions is the "model distillation" method. The essence of this technology lies in using the results of a powerful, resource-intensive AI system to train more compact and lightweight models. This allows for the deployment of advanced algorithms locally, without the need for the massive computing power required to create such systems from scratch.
Document analysis shows that the PRC's defense departments view American developments not merely as competitors, but as a source of technical knowledge that allows them to reduce the technological gap.
From Code to Drones: Examples of Application
Researchers have identified numerous specific cases of using this methodology in various fields, ranging from cybersecurity to the control of combat robots.
One striking example is a 2025 article published by researchers from Unit 96941 of the Chinese People's Liberation Army (PLA). This Beijing-based unit deals with military intelligence and cyber warfare. The work describes the use of OpenAI's GPT-3.5 model to process confidential military data. The researchers acknowledged that third-party models cannot directly process secret information. To bypass this limitation, they used GPT-3.5 to generate brief descriptions of software code, which were then used to train a domestic model designed to operate exclusively within China's isolated military networks.
Another example concerns information monitoring. At the North China University, which is closely linked to the arms industry, researchers applied the Claude 3 Haiku library from Anthropic. It was used to generate synthetic training data for a text classification model used to monitor social media and moderate content.
Particularly interesting is the application of distillation in the field of unmanned systems. A 2024 article published by the PLA's National University of Defense Technology describes using the method to reduce the size of image processing models. This allows drones to analyze video feeds in real-time, maintaining navigation and target acquisition even when communication with the operator is completely severed.
Similar approaches were recorded in a 2026 study by the Chinese Academy of Military Sciences. It describes the deployment of target recognition models on tactical equipment during simulation operations involving drones, ships, and unmanned submarines.
Geopolitical Tensions and Accusations
These findings have become a cause for serious deterioration in relations between Beijing and Washington. As reported by Reuters, the issue of the Chinese military using American technology has become one of the main flashpoints ahead of the September negotiations.
US officials have already leveled accusations against several Chinese companies, asserting that the distillation method is being used to extract intellectual capabilities from American models. According to Washington, this undermines export controls and violates intellectual property rights.
China categorically rejects these claims. In Beijing, they state that Washington is attempting to pursue a policy of "hegemony" in the field of artificial intelligence by restricting other countries' access to technology. Furthermore, Chinese officials claim that American companies themselves practiced similar methods in the past.
