Chinese Military Extracts US AI Model Data to Upgrade Defense Networks

Chinese military institutions are systematically utilizing outputs from leading American artificial intelligence models to train domestic defense networks.

Researchers linked to the People’s Liberation Army have extracted reasoning data from platforms developed by OpenAI and Anthropic to upgrade tactical systems.

This practice bypasses Washington’s technology export controls and reduces reliance on advanced semiconductor hardware.

A review of over 80 Chinese academic papers and patents reveals widespread reliance on a technique termed "model distillation."

The process involves capturing the outputs of massive frontier AI models to train smaller, specialized systems that operate locally.

Military Distillation Tactics

Documents compiled by the Jamestown Foundation and Reuters illustrate how the PLA leverages American AI platforms as technical blueprints.

By extracting data from U.S. models, Chinese military engineers eliminate the prohibitive computing power required to build foundational systems from scratch.

Sunny Cheung, a Jamestown fellow who reviewed the military literature, noted the strategic value of this approach.

"Teaching a model the right answer is one thing but teaching it the reasoning behind the answer is much harder," Cheung stated.

"These papers show Chinese military-linked researchers are trying to transfer that expensive, proprietary reasoning from Western models into smaller systems they can control and deploy locally," he added.

Cyber Warfare and Tactical Deployment

PLA Unit 96941, a Beijing-based cyber-warfare and intelligence division, applied OpenAI’s GPT-3.5 to process sensitive source code.

Military engineers used the American model to summarize classified software code. They subsequently trained a domestic model on those exact summaries to operate entirely within closed Chinese military networks.

The National University of Defense Technology detailed another application in a 2024 academic paper. Researchers shrank an image-processing model through distillation for deployment on unmanned aerial vehicles.

This lightweight model enables drones to analyze live video feeds for real-time navigation and targeting, even when communications links sever.

The Academy of Military Sciences similarly deployed distilled target-recognition models on tactical hardware during simulated maritime operations involving autonomous submarines and ships.

Intellectual Property Disputes

The unauthorized extraction of capabilities from American models has triggered significant bilateral friction ahead of U.S.-China talks on AI governance.

U.S. officials argue the practice infringes on intellectual property and nullifies existing technology export controls.

Beijing has firmly dismissed these accusations, framing Washington’s complaints as AI hegemonism while insisting American corporations execute identical extraction techniques.

Commercial AI developer Anthropic stated it blocks commercial access to its Claude platform in China and employs monitoring protocols to prevent misuse.

The company warned that distilled models often strip away the original system’s safety protocols.

Despite widespread utilization, industry executives highlight structural constraints within the extraction strategy.

Trevor Koverko, co-founder of AI data firm Sapien, observed that distilled networks inherit limited parameters.

"It is best understood as transferring selected capabilities into a cheaper, locally controlled system, not achieving independence from frontier AI," he stated.