Chinese researchers tied to military institutions have reportedly been leveraging outputs from leading U.S. AI models to build advanced defense applications, according to an exclusive Reuters investigation. This strategy allows China to circumvent stringent U.S. export controls that limit access to the latest AI hardware required for developing cutting-edge models from the ground up.

The report, developed in collaboration with the Washington D.C.-based Jamestown Foundation, analyzed over 80 academic papers and patent filings by Chinese entities. It found these groups extensively applied a technique called “model distillation,” which extracts knowledge from powerful “teacher” models—such as those created by American startups like OpenAI and Anthropic—and uses this information to train smaller, highly efficient “student” models. This method enables China to produce sophisticated AI systems capable of matching the performance of U.S. frontier models while relying on less advanced computational resources.

Model distillation has valid uses across the AI industry, such as optimizing systems to run on limited hardware. However, contention arises from evidence that Chinese AI development has surged through distillation of proprietary U.S. technology to gain military advantages. For instance, Alibaba recently unveiled Qwen3.8, a model reportedly rivaling the capabilities of Anthropic’s Claude Fable 5 and surpassing OpenAI’s GPT-5.6 in benchmarking tests. Similarly, Moonshot AI introduced Kimi K3, the world’s largest open-weight model, which also claims top-tier performance.

These breakthroughs prompted strong reactions from U.S. officials and AI developers, who accuse Chinese firms of appropriating American AI breakthroughs without authorization. The U.S. Treasury Secretary has warned of potential sanctions against companies found guilty of this distillation practice. Conversely, some industry leaders, including Microsoft CEO Satya Nadella, have highlighted the nuanced debate around intellectual property in AI, noting that many leading models, including those by Anthropic, were trained on vast internet datasets without explicit consent.

The Reuters investigation cites multiple examples illustrating how Chinese military-linked organizations have systematically incorporated outputs from U.S. frontier AI systems into their own technology pipelines, accelerating their capacity to deploy highly capable AI applications within defense sectors.