Chinese military researchers have used outputs from top American artificial intelligence models — including those built by OpenAI and Anthropic — to help train domestic AI systems, according to a Reuters review reported by MSN and echoed by RealClearDefense, which described the Chinese military as tapping U.S. AI models to train defense systems.
The technique at the heart of the story is straightforward, even if the geopolitics are not. Rather than needing direct access to a rival lab's proprietary model, researchers can feed prompts into a leading system, collect the answers, and use that pile of high-quality text to train a separate, homegrown model. The stronger model effectively becomes a teacher. It is a way of capturing some of the capability of a system you do not own and cannot legally copy.
That is what makes the finding awkward for U.S. AI companies. Both OpenAI and Anthropic restrict military use and generally prohibit using their outputs to build competing models, but those are contractual rules, not technical locks. Once a model is reachable — directly or through intermediaries — its outputs are hard to fence in.
The available reporting is thin on specifics: the source items summarize a Reuters review without detailing which models, which Chinese institutions, or what military systems were involved, and neither company's response is included here.
It matters because export controls have focused on chips and hardware, and this suggests the harder thing to control may be the knowledge those chips produce.