- Chinese military researchers have used outputs from U.S. AI models like OpenAI's GPT-3.5 and Anthropic's Claude to train domestic defense AI systems via a technique called model distillation.
- A Reuters review of over 80 Chinese academic papers and patents reveals widespread use of distillation by PLA-linked institutions to develop specialized AI for surveillance, cyber warfare, and tactical decision-making.
- Model distillation involves using outputs from a powerful AI system to train smaller, specialized models, reducing the computing requirements compared to building frontier AI from scratch.
- The practice is disputed, with U.S. officials accusing Chinese entities of unauthorized extraction that undermines export controls and infringes IP, while China denies reliance and accuses the U.S. of 'hegemonism'.
- One paper from PLA Unit 96941 described using GPT-3.5 to summarize sensitive military source code and training a domestic model on those summaries for use within military networks.
- Other examples include using Claude 3 Haiku to generate training data for social media monitoring, distilling image-processing models for drone deployment, and target recognition for maritime operations.
- China promotes model lightweighting and edge computing to enable AI on drones and satellites, but experts note distilled models have limitations and cannot fully replicate frontier AI's broad capabilities.
- Chinese researchers are also studying distillation as a security risk, proposing defenses against 'data-free distillation' that could reverse-engineer models.