Chinese Military Researchers Use OpenAI and Anthropic Outputs to Train Domestic Defense AI Systems

A 2024 paper from the People's Liberation Army's National University of Defense Technology detailed a method to shrink an image-processing model so unmanned aerial vehicles could analyze live video during communications blackouts. The underlying intelligence for that system did not originate in Beijing. It was extracted from American artificial intelligence models.[1][2]

Chinese military researchers are systematically bypassing U.S. export controls by using the outputs of frontier models developed by OpenAI and Anthropic to train their own domestic defense systems. A review of more than 80 Chinese academic papers and patents reveals that institutions linked to the PLA are leveraging a technique known as "model distillation" to transfer the reasoning capabilities of Western AI into smaller, locally deployed networks.[1][3]

The mechanism sidesteps the core logic of Washington's technology embargoes. U.S. export controls are designed to deny China the advanced semiconductors required to train massive frontier models from scratch. However, distillation does not require those chips. By feeding queries into existing American models and using the generated text and reasoning as synthetic training data, Chinese researchers can build highly capable "student" models on modest local hardware.[4][5]

The applications span multiple domains of modern warfare. Researchers at China's Academy of Military Sciences have used distillation to run target-recognition models on tactical hardware during simulated maritime operations involving ships, drones, and unmanned submarines. The goal is to enable autonomous swarms to make collective targeting decisions with minimal human intervention.[1][6]

In the cyber domain, the integration is equally direct. A paper published by PLA Unit 96941, a Beijing-based military intelligence and cyber-warfare unit, described using OpenAI's GPT-3.5 to summarize sensitive military source code. Because third-party models are considered unsuitable for handling classified information directly, the researchers used the American model to process the data before training a secure, domestic network that operates entirely within Chinese military infrastructure.[1][3]

Civilian-military fusion extends to social monitoring. At the North University of China, an institution with deep ties to the country's defense sector, researchers fed prompts to Anthropic's Claude 3 Haiku to generate synthetic data for text classification. That data was then used to train systems designed for social media monitoring and content moderation.[2][3]

The practice places Western AI developers in a complex position. Anthropic has stated that it does not provide commercial access to its systems in China or to Beijing-controlled firms, and it employs active monitoring to detect policy violations. The company also warned that distilled copies strip out the vital safety mechanisms built into the original models, allowing sensitive capabilities to proliferate beyond their control.[1][3]

U.S. officials argue that this unauthorized extraction undermines intellectual property rights and effectively neutralizes the hardware embargoes meant to preserve America's technological edge. The administration has grown increasingly critical of the approach, pointing to instances where Chinese open-weight models were found to contain traces of proprietary American systems.[4][5]

Beijing has dismissed the criticism, characterizing Washington's export controls and subsequent complaints as "AI hegemonism." Chinese strategists view the integration of artificial intelligence into military operations—a doctrine they term "intelligentized warfare"—as essential for algorithmic dominance and systems destruction in future conflicts.[5][6]

The distillation dispute highlights a structural vulnerability in the current AI arms race. While the U.S. maintains a lead in foundational model development and the hardware required to build them, the outputs of those models are inherently porous. As long as the reasoning generated by frontier systems can be queried and captured, the barrier to entry for developing specialized, military-grade AI will continue to fall.[4][6]