Chinese military researchers have used outputs from leading U.S. artificial intelligence models built by OpenAI and Anthropic to train domestic AI systems, according to an exclusive Reuters report by Eduardo Baptista published from Beijing on July 31.
Reuters says its findings are drawn from papers and patents, as summarized by the aggregator Techmeme: Chinese military researchers "distilled" OpenAI and Anthropic models to train domestic AI systems and advance China's defense capabilities.
Distillation, in plain terms, means using one AI model's answers as teaching material for another. Rather than building a powerful model from scratch — an enormously expensive undertaking — a team can query an existing top-tier system and use its responses to train a smaller, specialized model of their own. One MSN summary of the Reuters story frames the appeal exactly that way: the technique allows researchers to train specialized domestic AI systems efficiently.
The report landed alongside a broader push. According to an MSN item on the People's Liberation Army's 99th anniversary, Chinese President Xi Jinping directed the PLA to accelerate the military application of artificial intelligence, with a focus on building an "intelligent military system."
The Reuters exclusive was picked up widely, including by The Economic Times, Türkiye Today, The News International, Technology Org and Internazionale, a sign of how closely the AI-and-defense beat is being watched internationally.
It matters because export controls have focused on keeping advanced chips out of Chinese hands, but a model's outputs are not a physical good — and if they can be turned into military-grade training data, the hardest part of the U.S. lead may be far easier to copy than policymakers assumed.