Reuters: Chinese Military Researchers Distilled OpenAI and Anthropic Models for Defense Systems
More than 80 papers and patents show PLA-linked researchers training smaller domestic models on Claude and GPT output, then applying them to code analysis, drone navigation, and target recognition.
Chinese military researchers built defense AI by distilling US models. Reuters found the paper trail.
A Reuters investigation published July 31, 2026 reviewed more than 80 Chinese academic papers and patents and found military-linked researchers distilling OpenAI and Anthropic models, training smaller domestic systems on their outputs, to build defense-relevant AI.
What the papers actually show
Named institutions include PLA Unit 96941, North University of China, the PLA National University of Defense Technology, and the Academy of Military Sciences. Examples Reuters and independent researcher Sunny Cheung of the Jamestown Foundation documented: North University of China used Claude 3 Haiku to generate synthetic training data for content moderation, PLA Unit 96941 used GPT-3.5 to analyze military software code and train an internally deployable model, and separate teams applied distillation to UAV navigation under disrupted communications and to target recognition for simulated drone and submarine operations.
How distillation actually works
Distillation trains a smaller model to imitate a larger one's outputs, and in the more advanced cases documented here, its reasoning steps, not just its final answers. It is a cheaper shortcut to specialized capability than training a comparable model from scratch, and it does not require the export-controlled chips that training a frontier model from the ground up would need. Reuters and outside researchers caveat that a distilled model inherits only the specific capability it was trained to imitate, not the full range of what the original frontier model can do.
How this differs from the last version of this story
This site previously covered US Treasury Secretary Scott Bessent saying regulators were finding watermarks of US models inside Chinese ones, a stated intent to investigate with no lawsuit or sanction filed. This Reuters investigation is the more concrete version of that same underlying concern: actual papers and patents, not just an accusation.
Why a build studio cares
Which frontier model a client build sits on top of is not just a cost and capability decision anymore. It is also a provenance question, and this is a real, documented case of where an unlicensed copy of a model's outputs can end up.
Next step: read the Reuters investigation or Jamestown Foundation's independent analysis. If model provenance is a question in your own AI build, write to us at hello@gattyworks.com.