Jakarta, ThedailyID — Chinese military researchers have used outputs from artificial intelligence models developed by OpenAI and Anthropic to strengthen domestic AI systems, according to a Reuters investigation.
Reuters reviewed more than 80 academic papers and patents from China. The investigation shows how military and security institutions relied on leading U.S. AI models to speed up domestic development.
Researchers mainly used a technique called model distillation. They trained smaller AI models with outputs from more advanced systems. This method requires far less computing power than building frontier models from scratch.
The strategy comes as the United States continues to restrict China’s access to advanced chips and other critical technologies.
Reuters and the Jamestown Foundation found that military-linked researchers widely adopted the technique. The papers show Chinese defense institutions view top U.S. AI models as technical references. They also see them as tools to narrow the technology gap with American rivals.
The dispute centers on unauthorized extraction of AI outputs. Distillation itself remains a common practice across the AI industry.
The issue has become a key topic ahead of U.S.-China talks on AI governance and security. U.S. officials accuse some Chinese organizations of using distillation to capture the capabilities of American AI models. They argue the practice could weaken export controls and violate intellectual property rights.
China denies those allegations. Officials say Washington is pursuing AI “hegemony.” They also argue U.S. companies use similar training methods.
Reuters verified 24 military-related studies. One paper came from PLA Unit 96941, a military intelligence and cyber warfare unit in Beijing. Researchers used OpenAI’s GPT-3.5 to summarize sensitive military software code. They then trained a domestic model that could operate entirely inside China’s military network.
Other studies described additional military applications. Researchers used Anthropic’s Claude 3 Haiku to generate synthetic training data for content monitoring systems. Another study adapted distilled AI models for drones. The system enabled real-time video analysis, navigation, and targeting.
Anthropic said it does not provide commercial access to Claude in China or to companies controlled by Beijing. The company also warned that distilled models may lose important built-in safety protections.
