Chinese military researchers have been using American artificial intelligence models — including systems from OpenAI and Anthropic — to help develop and train defense technology, according to a Reuters review reported by outlets including Tekedia and SOFX.
The finding was carried widely on Friday under near-identical headlines across US and international news organizations, among them the Honolulu Star-Advertiser, the Rapid City Journal, the Independent Record and Malaysia's Awani International, all reporting that "Chinese military researchers tap US AI models to train defense systems."
The specific mechanism described in the coverage is worth noting: SOFX framed the story as China training military AI on outputs from OpenAI and Anthropic models — not on the underlying models themselves. That distinction matters. The most capable American models are not handed over wholesale; they are typically reached through interfaces that generate text, code and analysis on demand. Those generated outputs can then be collected and used to teach a separate, locally controlled system to behave in similar ways — a practice sometimes called distillation. Beyond the Reuters characterization reported by these outlets, the sources here do not specify which systems were built, by whom, or over what period.
Both OpenAI and Anthropic publicly prohibit military weapons applications of their products under their usage policies, which makes any documented military research use a test of whether such rules can actually be enforced once a model is answering prompts from anywhere in the world.
It matters because it suggests the export controls and chip restrictions the US has built to slow China's military AI progress may not cover the most portable ingredient of all — the knowledge these models produce every time someone asks them a question.