Chinese military uses US AI models to train domestic defense systems

Researchers linked to the Chinese People's Liberation Army (PLA) and other military institutions are leveraging frontier AI models from US companies to train domestic defense systems. This practice, known as model distillation, allows China to transfer advanced reasoning capabilities from American software into smaller, locally controlled systems.

According to a Reuters investigation based on more than 80 academic papers and patents, the PLA is using outputs from OpenAI and Anthropic to build specialized AI. Specifically, PLA Unit 96941, which focuses on military intelligence and cyber-warfare, utilized OpenAI's GPT-3.5 to summarize sensitive military source code. This process was used to train a domestic model designed to operate entirely within secure military networks.

Other military institutions are applying these techniques to autonomous hardware. The PLA's National University of Defense Technology used distillation to shrink image-processing models for drones, enabling real-time video analysis and targeting without the need for external communications. Similarly, researchers at the Academy of Military Sciences applied the process to target-recognition models for maritime operations involving ships, drones, and unmanned submarines. Beyond combat systems, the North University of China used Anthropic's Claude 3 Haiku to generate synthetic training data for content moderation and social media monitoring.

Circumventing Hardware Constraints

This strategy emerges as a direct response to strict US export controls on high-end AI chips, which were designed to hinder China's military AI progress. By promoting "model lightweighting" and edge computing, China aims to run sophisticated AI on hardware with limited processing power, such as satellites and drones.

Sunny Cheung, a fellow at Jamestown, noted that these papers demonstrate an effort by military-linked researchers to transfer "expensive, proprietary reasoning from Western models into smaller systems they can control and deploy locally." However, Trevor Koverko, co-founder of Sapien, clarified that the practice is best understood as transferring specific capabilities into cheaper systems rather than achieving total independence from frontier AI.

Security Implications

The use of distillation creates a security paradox for the US. While hardware restrictions limit the raw computing power available to the PLA, the "intelligence" of US software is being used to bridge that gap. Furthermore, the distillation process effectively strips away the AI safety safeguards implemented by US developers, potentially accelerating the development of autonomous weapons and cyber-warfare tools using American intellectual property.

Future Outlook

As China continues to refine its ability to shrink frontier models, the effectiveness of chip-based export controls may be partially undermined. Observers will be watching for further evidence of this "software-to-hardware" bypass and whether US AI labs can implement more robust protections to prevent their models from being used as training engines for foreign military systems.