Photo: Getty Images
Chinese military researchers used data generated by leading US artificial intelligence models developed by OpenAI and Anthropic to train domestic AI systems aimed at strengthening China's defense capabilities, according to a Reuters investigation based on more than 80 Chinese academic papers and patents.
The previously unreported documents provide a rare look at how Chinese military and security institutions are using cutting-edge US AI models to accelerate the development of their own specialized systems, despite Washington's efforts to restrict Beijing's access to advanced chips and other strategic technologies.
The documents show extensive use of a technique known as model distillation, in which the outputs of large AI models are used to train smaller, specialized systems that can run locally without the massive computing power required to build frontier AI models from scratch.
Reuters, citing research by the Jamestown Foundation, found that model distillation is widely used by researchers affiliated with the People's Liberation Army (PLA) and other Chinese military institutions.
According to the documents, Chinese defense organizations view leading US AI models both as a source of technical expertise and as a shortcut to narrowing the technological gap with American competitors.
The dispute centers on the unauthorized extraction of model outputs rather than the distillation process itself, which is a common industry practice. The issue has become a key point of contention in US-China talks on AI governance and security. US officials have accused some Chinese organizations of using distillation to replicate the capabilities of American AI models, potentially undermining export controls and violating intellectual property rights. China has rejected those allegations, accusing Washington of seeking AI "hegemony" and arguing that US companies engage in similar practices.
Chinese AI developers have also denied claims that their latest advances rely on foreign models. Last week, AI startup Moonshot rejected accusations from the Trump administration that its Kimi K3 model had been built using model distillation, insisting it was based on its own innovations.
Jamestown Foundation researcher Sunny Cheng, who analyzed more than 60 Chinese papers, said Chinese military scientists are systematically capturing the reasoning processes of Western AI models to adapt them for surveillance, cyber warfare and tactical decision-making.
Reuters also identified dozens of military-related studies describing the use of model distillation. One paper published last year by researchers from PLA Unit 96941, a military intelligence and cyber warfare unit based in Beijing, described using OpenAI's GPT-3.5 to process sensitive military source code.
Because external AI models were considered unsuitable for handling classified information, researchers used GPT-3.5 to summarize the code and then trained a domestic AI model on those summaries for deployment within China's secure military networks.
The investigation found additional military applications. Researchers at China's North University, which has close ties to the defense industry, used Anthropic's Claude 3 Haiku model to generate synthetic training data for an AI system designed to monitor social media and moderate online content.
Anthropic said it does not provide commercial access to Claude in China or to companies controlled by Beijing and has monitoring systems in place to detect policy violations. The company also warned that distilled models may lose the safety protections built into the original systems, potentially transferring sensitive capabilities to models outside its control.
A 2024 paper from the PLA's National University of Defense Technology described using model distillation to shrink an image-processing AI model so it could be deployed on drones, allowing UAVs to analyze video in real time and continue navigation and targeting even if communications are disrupted.
Another study published earlier this year by researchers at the Chinese Academy of Military Science detailed the use of distilled AI models for target recognition during simulated naval operations involving drones, surface vessels and autonomous underwater vehicles.
China has increasingly embraced model distillation as it seeks to compete with the United States in advanced AI while facing restrictions on access to cutting-edge computing hardware due to US export controls on advanced semiconductors. Both central and local governments have invested in "lightweight" AI models and edge computing technologies that allow AI systems to run on drones, satellites and other devices with limited computing power.
However, experts caution that model distillation has significant limitations. Chinese military researchers have also identified it as a potential security risk. In January, researchers at the PLA Engineering University published a paper examining so-called "data-free distillation" — a reverse-engineering technique capable of reproducing model capabilities without direct access to the original parameters.
To counter that threat, they proposed protective mechanisms designed to obscure the hidden reasoning patterns revealed in public AI outputs, while acknowledging that distilled models inherit only selected capabilities and cannot fully replicate the broad intelligence of state-of-the-art AI systems.
