Chinese military researchers tap US AI models to train defence systems

• Chinese military units, including PLA Unit 96941, have used model distillation to adapt U.S. AI capabilities for surveillance, cyber warfare, tactical decision-making, and drone-based target recognition while overcoming limitations on handling classified information.
• Applications range from processing sensitive military source code to deploying AI on unmanned aerial vehicles for real-time video analysis and navigation when communications are cut, as documented in papers from institutions like the National University of Defense Technology.
• While distillation helps China compete amid U.S. chip export restrictions, experts note it cannot replace the massive computing power needed for frontier AI development and distilled models remain less capable than their original 'teacher' systems.

Summary: Chinese Military Leverages US AI Models for Defense Systems

Chinese military researchers have systematically used outputs from leading US AI models, including OpenAI's GPT-3.5 and Anthropic's Claude 3, to train domestic defense systems, according to a Reuters review of over 80 Chinese academic papers and patents.

Key Technique:

The practice, known as "model distillation," involves using outputs from powerful AI systems to train smaller, specialized models that can operate locally without massive computing requirements. This allows Chinese institutions to bypass US export controls on advanced chips while developing military AI capabilities.

Military Applications:

• PLA Unit 96941 used GPT-3.5 to process sensitive military source code and train domestic models within Chinese military networks
• National University of Defense Technology developed image-processing models for unmanned aerial vehicles (UAVs) to support real-time navigation and targeting
• Academy of Military Sciences deployed target-recognition models for tactical operations involving drones, ships, and submarines
• North University of China used Claude 3 Haiku for social media monitoring and content moderation

Strategic Context:

The revelations emerge as a major flashpoint in US-China AI competition ahead of bilateral talks on AI governance. US officials accuse Chinese entities of unauthorized extraction that undermines export controls and infringes intellectual property. China denies the allegations, calling it US "AI hegemonism," with companies like Moonshot rejecting claims their models rely on foreign technology.

Limitations:

Experts note distillation cannot replace the massive computing power needed to develop frontier AI from scratch. Distilled models inherit only selected capabilities and remain less sophisticated than their source systems, functioning primarily as a cost-effective shortcut rather than achieving AI independence.