- Distillation sidesteps the logic of US export controls, which restrict the chips needed to train frontier models but cannot restrict the text those models produce
- Review of more than 80 Chinese papers and patents found PLA-linked researchers distilling outputs from OpenAI and Anthropic models into small systems they can run locally
- The Trump administration is increasingly critical of the approach and has claimed Chinese research lab-created Kimi K3 distilled Anthropic's Fable model
A new investigation has claimed Chinese military researchers have used outputs from American AI models built by OpenAI and Anthropic to train domestic systems intended to advance the country's defense capabilities.
The findings are based on a Reuters review of more than 80 Chinese academic papers and patents, incorporating research compiled by the Washington-based Jamestown Foundation shared exclusively with the news agency.
Reuters says it independently verified the literature and found a further two dozen military-linked case studies of its own.
A distillation problem in an AI race that continues to heat up
The mechanism at issue is model distillation, in which the outputs of a large, capable system are used as training data for a smaller one. The smaller model inherits selected behaviors at a fraction of the compute cost and, crucially, can run on modest local hardware.
That is the strategic point. Washington's export controls are built to deny China the chips needed to train frontier models. Distillation does not need those chips, because the expensive part has already been paid for by someone else. It sits alongside Beijing's other route around the controls: substituting domestic silicon for American designs, a shift that carries its own long-term threat to Nvidia and AMD.
Sunny Cheung, the Jamestown fellow who analyzed more than 60 of the papers, frames the value as reasoning rather than answers. Getting a model to produce the right output is comparatively easy, he told Reuters, but "teaching it the reasoning behind the answer is much harder."
The papers, on his reading, show Chinese military-linked researchers trying to move that expensive proprietary reasoning into small systems they can control and deploy themselves.
There is an irony in the pattern. Beijing has been moving to restrict consumer access to Western AI models on security grounds, and has pushed its own labs away from US-designed chips, even as its military researchers mine the outputs of American systems.
The asymmetry points at what is wrong with the US instrument. Export controls regulate objects: chips, tools, hardware that crosses a border and can be counted. The thing being transferred here is text a model produced- which crosses no border in any customs sense and cannot be enumerated.
The White House has registered the problem at both the chip and distillation end, but registering it is not the same as having a lever. For now, the debate over whether distillation is theft or standard practice may matter less than the fact that neither answer offers a workable control.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.
