Chinese military used OpenAI, Anthropic AI outputs to train their defence systems: Report

Summary

Chinese military researchers affiliated with the People’s Liberation Army have drawn on outputs from OpenAI and Anthropic models to build smaller, specialised domestic systems through model distillation. Reuters examined more than eighty academic papers and patents, supplemented by material from the Jamestown Foundation, and found repeated use of this technique to transfer capabilities from frontier models into systems that can run on constrained hardware without relying on restricted high-end chips.

The papers describe distillation as a way to capture not only answers but also the underlying reasoning processes. One study from PLA Unit 96941 used GPT-3.5 to summarise sensitive military software code, then trained a local model on those summaries so the final system could operate entirely inside Chinese networks. Researchers at the PLA National University of Defense Technology applied similar methods to compress an image-processing model for unmanned aerial vehicles, enabling onboard video analysis for navigation and targeting when communications are lost. At the Academy of Military Sciences, distillation supported target-recognition models deployed on tactical hardware during simulated maritime operations involving drones, surface vessels and unmanned submarines.

Additional work at the North University of China used Claude 3 Haiku to generate synthetic data for social-media monitoring and content-moderation systems. Across these cases, distillation is presented as a practical response to US export controls on advanced semiconductors, allowing edge deployment on drones, satellites and other low-power platforms. At the same time, some Chinese researchers have begun studying distillation itself as a potential security risk, examining methods such as data-free distillation that could expose model behaviour through public outputs.

The findings coincide with separate allegations that large-scale queries were used to extract capabilities from Anthropic’s Claude models, underscoring ongoing concerns over unauthorised transfer of proprietary reasoning into systems beyond the original providers’ control.

Why it matters

Directly addresses military AI development, dual-use distillation techniques, and export-control circumvention relevant to NATO defense professionals tracking adversary capabilities and IP risks.