Chinese military researchers have used outputs from leading artificial intelligence models developed by US companies OpenAI and Anthropic to train domestic AI systems aimed at strengthening China’s defence capabilities, according to a Reuters review of more than 80 Chinese academic papers and patents.
The previously unreported findings, based on research compiled by the Washington-based Jamestown Foundation and shared exclusively with Reuters, provide a rare insight into how military and security-linked institutions in China are using advanced US AI models to accelerate the development of specialised domestic systems. According to Reuters, the approach allows researchers to bypass some of the significant computing resources normally required to build frontier AI models from the ground up.
The documents reviewed by Reuters show widespread use of a technique known as “model distillation”, in which outputs from a powerful AI model are used to train smaller, specialised systems capable of running locally without the enormous computational demands of the original models. Reuters reported that the practice is widely employed by researchers associated with the People’s Liberation Army (PLA) and other Chinese military institutions.
The papers reviewed by Reuters suggest that Chinese defence organisations regard leading US AI models both as valuable sources of technical knowledge and as a means of narrowing the technological gap with their American counterparts. While model distillation itself is a widely used industry practice, Reuters noted that the dispute centres on the alleged unauthorised extraction of capabilities from proprietary AI systems.
The issue has emerged as a significant point of contention ahead of US-China discussions on artificial intelligence governance and safety. According to Reuters, US officials have accused some Chinese entities of using model distillation to extract capabilities from American AI models in ways that could undermine export controls and infringe intellectual property rights. China has rejected those allegations, accusing Washington of pursuing AI “hegemonism” and arguing that US companies have also employed similar techniques.
Chinese AI developers have likewise disputed claims that their technological advances depend on foreign models. Reuters reported that AI startup Moonshot last week denied allegations by the Trump administration that its Kimi K3 model had been built using distillation, stating instead that the model was based on proprietary innovations.
Sunny Cheung, a Jamestown Foundation fellow who analysed more than 60 of the papers, told Reuters that Chinese military scientists are systematically attempting to capture the reasoning processes of Western AI models for applications including surveillance, cyber warfare and tactical decision-making. He said that while teaching a model the correct answer is relatively straightforward, transferring the reasoning behind those answers is considerably more difficult, adding that the papers indicate efforts to adapt that proprietary reasoning into smaller systems that can be deployed under Chinese control.
Reuters independently verified the academic literature and identified an additional two dozen military-linked case studies. One paper published last year by researchers in PLA Unit 96941, a military intelligence and cyber-warfare unit in Beijing, described using OpenAI’s GPT-3.5 to process sensitive military source code. According to Reuters, the researchers concluded that third-party models were unsuitable for handling classified information and therefore used GPT-3.5 to summarise software code before training a domestic model on those summaries, allowing it to operate entirely within Chinese military networks.
Reuters reported that requests for comment sent to the White House, the Pentagon, China’s foreign ministry, the PLA and OpenAI did not receive responses.
The Reuters review also found that Chinese researchers have applied model distillation across a broad range of projects extending beyond defence. At the North University of China, which has close links to the country’s weapons industry, researchers used Anthropic’s Claude 3 Haiku to generate synthetic training data for a text classification model intended for social media monitoring and content moderation.
Anthropic told Reuters that it does not provide commercial access to Claude in China or to Beijing-controlled organisations and operates monitoring systems to detect violations of its policies. The company also warned that distilled models may lose the original systems’ built-in safety safeguards, potentially allowing sensitive capabilities to be transferred into models beyond its control.
Additional studies reviewed by Reuters illustrate how the technology is being adapted for military operations. A 2024 paper from the PLA’s National University of Defense Technology described using model distillation to reduce the size of an image-processing model for deployment on unmanned aerial vehicles, enabling drones to analyse live video and support navigation and targeting decisions even when communications are disrupted. Another study published earlier this year showed researchers at China’s Academy of Military Sciences using the technique to operate a target-recognition model on tactical hardware during simulated maritime operations involving drones, ships and unmanned submarines.
Reuters reported that China has embraced model distillation as it seeks to compete with the United States in advanced artificial intelligence while facing restrictions on access to high-end semiconductors under US export controls. Central and local governments have promoted model lightweighting and edge computing, directing funding towards technologies capable of running AI models on drones, satellites and other devices with limited processing power.
Despite these advantages, experts told Reuters that model distillation has important limitations. Distilled systems inherit only selected capabilities and cannot fully reproduce the broad intelligence of frontier AI models. Trevor Koverko, co-founder of AI data company Sapien, said distilled models remain less capable than the original systems, describing the process as a way of transferring selected capabilities into lower-cost, locally controlled models rather than achieving complete technological independence from frontier AI.
