Chinese military researchers tap US AI models to train defence systems

Reuters  |  Eduardo Baptista

Chinese military researchers affiliated with the People's Liberation Army have leveraged outputs from leading American artificial intelligence models developed by OpenAI and Anthropic to train domestic defence systems, according to a review of over 80 academic papers and patents. The practice relies heavily on model distillation, a process where smaller, specialized neural networks capture the complex reasoning pathways of advanced Western systems for localized deployment in surveillance, cyber operations, and tactical decision-making without requiring massive compute infrastructure.

By extracting proprietary technical capabilities, military-linked institutions seek to bypass Washington's strict export controls and hardware restrictions to narrow the technological gap with the United States. In response, American officials contend this unauthorized extraction violates intellectual property rights and undermines strategic export controls, while Beijing rejects the allegations as American technological hegemonism. The dispute over model distillation has subsequently emerged as a major friction point ahead of bilateral United States-China discussions on global artificial intelligence governance and safety standards.

Comment

Distillation techniques allow military software developers to convert the complex latent reasoning of frontier models into highly efficient, domain-specific algorithms. This architectural transfer bypasses physical compute bottlenecks established by the Bureau of Industry and Security's semiconductor export restrictions. Consequently, tactical edge devices deployed in bandwidth-constrained environments gain sophisticated decision-support capabilities without relying on high-power data centres. The trend demonstrates that software optimisation strategies can partially offset foreign hardware embargoes in state-level military modernisation.

To read in details click here