Chinese Military Researchers Used U.S. AI Outputs to Train Defense Models
Researchers linked to Chinese military and security institutions used outputs from OpenAI and Anthropic models to train specialized domestic systems, according to a Reuters review of papers and patents.
Why it matters: The report shows that AI controls must account for model access and output leakage, not only chip exports.
Chinese military-linked researchers have used outputs from U.S. artificial-intelligence models to train specialized domestic systems, according to a Reuters review of academic papers and patents.
The reviewed documents involved researchers associated with the People's Liberation Army and other Chinese military or security institutions. Applications described in the documents included surveillance, cyber operations, drone navigation and tactical decision support.
The technique is known as model distillation. A larger teacher model produces answers, examples or reasoning traces that are then used to train a smaller model for a narrower task.
Distillation is common in AI development and is not automatically improper. The controversy is whether organizations are extracting capabilities from proprietary models without authorization and applying them to restricted military or security programs.
The findings complicate U.S. efforts to slow China's AI development through semiconductor export controls. Restricting advanced chips may limit frontier-model training, but access to model outputs can still transfer knowledge and task-specific behavior.
AI competition is therefore expanding beyond control of processors. Model access, API monitoring, usage enforcement and contractual restrictions are becoming national-security issues.