American AI companies may be unintentionally tutoring China's defense sector.
According to a report summarized by Business Today, more than 80 Chinese academic papers and patents show that researchers linked to the Chinese military have used outputs generated by US AI models — including those from OpenAI and Anthropic — to train military systems.
Separately, Forbes reports that a Chinese AI firm siphoned American AI knowledge from Anthropic's Claude by running millions of prompts through it.
The technique at the center of both reports is straightforward in concept. Instead of stealing code or model weights, you simply ask a leading AI system a very large number of questions and collect its answers. Those answers become training material for a new model, which absorbs some of the original's capabilities without any of the cost of building it. The industry calls this distillation, and it is one of the reasons frontier labs guard their systems with usage rules and account restrictions rather than just firewalls.
What makes the reporting notable is the destination. Business Today describes the outputs feeding military systems, not commercial chatbots — meaning the same models that answer homework questions in the US may be indirectly shaping defense tools abroad.
Neither source item here details how the companies detected the activity, what the affected accounts were, or how the firms responded. The Forbes and Business Today items are the extent of the claims available in this brief.
Why it matters: US export controls were built to stop advanced chips from crossing borders, but a model's knowledge can leave through an ordinary API connection — and no one has figured out how to put a tariff on a conversation.