American export controls were meant to blunt China’s artificial intelligence ambitions, but a Reuters investigation has found that Chinese military-linked researchers are quietly using U.S. frontier models as a training shortcut. The practice, known as model distillation, is helping Beijing build specialized defense AI systems without massive computing resources.
China’s defense establishment has latched onto a technique called model distillation, in which data generated by a powerful existing AI system is used to train smaller, task-specific models. These distilled models can run locally and do not require the enormous computing infrastructure needed to develop cutting-edge models from scratch. According to Reuters, documents show the method is widespread among researchers connected to the People’s Liberation Army and other state security institutions.
The findings land at a sensitive point in U.S.–China relations. Both governments view access to AI technology as strategically crucial, and Washington has been tightening controls on advanced chips and other technologies to curb Beijing’s military modernization. Yet U.S. officials now accuse Chinese entities of squeezing capabilities out of American large language models, potentially violating intellectual property rights and undermining export restrictions.
Chinese military researchers and American AI models
One concrete example comes from North University of China, an institution with close ties to the country’s arms industry. Researchers there used Anthropic’s Claude 3 Haiku software to generate synthetic training data for a text classification model designed for social media monitoring and content moderation. In another case, OpenAI’s GPT-3.5 tool was reportedly used to process sensitive military source code.
Anthropic has said it does not provide commercial access to Claude in China or to Beijing-controlled companies, adding that it uses monitoring systems to detect policy violations. For Chinese defense institutions, American AI models appear to work as a shortcut to technical knowledge, helping them narrow the gap without building pioneering systems from zero.
Beijing rejects accusations of technological dependence
China has strongly rejected U.S. allegations. Beijing says Washington is pursuing hegemony in artificial intelligence and argues that American companies have used similar practices themselves. Chinese developers also push back on the idea that their progress rests on foreign models.
One notable case is startup Moonshot AI, which last week denied accusations from the Trump administration that its Kimi K3 model was built through distillation. The company insists the model is powered by proprietary innovation, not borrowed U.S. technology. The dispute is expected to feature in U.S.–China talks on AI governance and security.
Źródło: WNP.PL, Fot. EPA/AARON SCHWARTZ / POOL / PAP
