Chinese Military-Linked Researchers Used OpenAI and Anthropic Models to Train Defense Systems
A Reuters review of over 80 Chinese academic papers and patents shows researchers used GPT and Claude outputs via model distillation to train domestic military AI systems.
Nova AI News Editor
August 10, 2026 · 2 min read
A Reuters review of more than 80 Chinese academic papers and patents found that researchers linked to China's military have used outputs from leading AI models built by OpenAI and Anthropic to train their own defense systems.
What is "model distillation," and why does it matter?
The documents point to a technique called model distillation — a widely used AI method where a smaller model learns from the outputs of a much larger one. Instead of building a frontier model from scratch, a government or organization can use an existing advanced model's responses as a "teacher" to train its own smaller "student" model that needs far less computing power to run.
What do the specific examples show?
One study reviewed by Reuters, from researchers at PLA Unit 96941, used OpenAI's GPT-3.5 to summarize sensitive military software code before training a domestic model capable of running entirely inside Chinese military systems. Other papers describe using Anthropic's Claude 3 Haiku to generate synthetic training data for content moderation, and separate research explored deploying distilled models on drones, ships, and unmanned submarines.
What does this mean for these companies?
OpenAI and Anthropic both prohibit military or harmful use of their models under their terms of service. But model distillation effectively works around such restrictions — because the final system isn't directly connected to the original API, it merely "learns" from outputs the API generated. That points to a real gap in how AI companies can enforce usage policies or detect misuse of their API outputs downstream.
The bigger picture
This story is another reminder of how hard it is to contain the global spread of AI capability through export controls and usage policies alone. Blocking access to a model's API is one thing; preventing that model's publicly available outputs from circulating online and being used to train other models is a much harder problem to solve. That's pushing both regulators and AI companies toward new strategies focused on "output control" rather than just "border control."
Comments
No comments yet — be the first to comment.