An investigative report by Reuters has pointed towards increasing use of US artificial intelligence (AI) models by Chinese military researchers to train their domestic AI models, through distillation techniques, in order to boost China’s military capabilities.

According to the Reuters report, a review of over 80 academic papers and patents from China based institutes, with close ties to People’s Liberation Army (PLA), show use of model distillation on US AI models of OpenAI and Anthropic.

Distillation refers to a technique in which outputs from a powerful AI system are used to train smaller, specialized models without the massive computing cost required otherwise. Incidentally, a recent joint letter addressed to the Congress by over 200 American AI and Tech companies including OpenAI, Google, META, NVIDIA has advocated for open-weight AI models and distillation as a widely accepted technique for learning.

The Reuters report involve research done by the Washington-based Jamestown Foundation, which pointed to wide use of American models by researchers from China based institutes that have close ties to the PLA.

The report cites research studies that mention use of OpenAI’s GPT 3.5 to analyze and summarize military software code and ultimately train an indigenous AI model based on those summaries. Similarly, another research paper from North University of China indicates use of Claude Haiku model for machine learning training purposes for monitoring social media.

Also Read: China Accuses US of “AI Hegemonism” as Distillation Row Deepens