Report claims Chinese entities are distilling U.S. frontier AI models to power military AI applications, escalating geopolitical AI competition.

A Reuters investigation, conducted in collaboration with the Jamestown Foundation, has uncovered evidence that Chinese artificial intelligence firms are leveraging the outputs of American frontier models developed by OpenAI Group PBC and Anthropic PBC to train their own AI systems for defense applications. The review examined more than 80 academic papers and patent applications published by Chinese researchers and found that multiple military-linked institutions have been harvesting outputs from U.S. models to circumvent Washington's aggressive export controls, which have prevented Chinese startups from accessing the advanced chips needed to build sophisticated AI models independently.

The report accuses Chinese researchers of misusing "model distillation"—a technique that extracts outputs, reasoning steps, and data from powerful "teacher models" to train smaller, specialized "student models." This process allows the creation of more efficient models capable of running on less powerful hardware. By employing distillation, China has developed numerous models matching the performance of OpenAI's and Anthropic's most advanced frontier models while consuming far fewer computational resources.

While distillation has legitimate applications in optimizing models for specific workloads, significant controversy surrounds the technique, especially as Chinese models continue to demonstrate frontier-level capabilities. Last month, Alibaba Group Holding Ltd. released a preview of Qwen3.8 with benchmarks suggesting performance second only to Anthropic's Claude Fable 5, surpassing OpenAI's GPT-5.6. This followed Moonshot AI's release of Kimi K3, claimed to be the world's largest open-weights model with comparable performance claims.

In response, U.S. officials and AI startups have accused Chinese firms of distillation practices. U.S. Treasury Secretary Scott Bessent has threatened sanctions against Chinese AI companies deemed guilty of distilling American models, though some industry figures, including Microsoft Chief Executive Satya Nadella, have highlighted the irony given that Anthropic's models were trained on massive amounts of internet data without seeking permission.

Reuters documented specific instances of Chinese model distillation. PLA Unit 96941, the People's Liberation Army's main cyberwarfare group, distilled OpenAI's GPT-3.5 to process and summarize sensitive military source code, developing a lightweight model specialized for secure data handling on Chinese military networks. Researchers at the North University of China, believed to have close ties to weapons manufacturing, used Anthropic's Claude 3 Haiku to create synthetic training data for a text classification system for social media monitoring. At the National University of Defense Technology, researchers distilled a U.S. image processing model for deployment on unmanned aerial vehicles, enabling drones to analyze live video feeds in real time for navigation and weapons targeting.

The revelations are expected to intensify tensions ahead of upcoming talks between U.S. and Chinese officials on AI governance and safety. The White House contends that distillation undermines export controls and infringes intellectual property rights, while Beijing accuses the U.S. of pursuing "AI hegemonism."

SapienX Inc. co-founder Trevor Koverko characterized concerns about China surpassing the U.S. in AI through distillation as likely unfounded. "It is best understood as transferring selected capabilities into a cheaper, locally controlled system," he explained. "It doesn't mean achieving independence from frontier AI." While Chinese researchers may devise systems matching U.S. model capabilities in narrow applications, they will still need to develop original breakthroughs to outperform their U.S. rivals.