Chinese Military Researchers Used US AI Models from OpenAI and Anthropic to Train Defense Systems

Chinese military researchers are using model distillation to extract reasoning capabilities from advanced U.S. AI models, including OpenAI's GPT-3.5 and Anthropic's Claude, to train domestic defense systems. This practice allows Beijing to bypass U.S. chip export controls by running smaller, specialized models locally on tactical hardware like drones and secure military networks. The issue has intensified diplomatic tensions over intellectual property and export controls ahead of bilateral AI safety talks.

Chinese military AI model distillation

• ▪Chinese military-linked researchers utilize model distillation to transfer reasoning capabilities from Western AI models into smaller, locally controlled systems.
• ▪A 2023 paper by PLA Unit 96941 described using OpenAI's GPT-3.5 to summarize military source code to train a domestic model running on secure military networks.
• ▪Chinese military researchers used outputs from artificial intelligence models developed by OpenAI and Anthropic to train domestic defense systems.

US-China AI technology competition

• ▪United States officials accuse Chinese entities of using model distillation to extract capabilities from American AI models, potentially undermining export controls and intellectual property.
• ▪China rejected accusations of unauthorized AI extraction, accusing the United States of pursuing AI hegemonism and stating that United States firms engage in similar practices.

PLA defense surveillance applications

• ▪Researchers at the North University of China used Anthropic's Claude 3 Haiku to generate synthetic training data for social media monitoring and content moderation.
• ▪A 2024 paper from the PLA's National University of Defense Technology described using distillation to shrink an image-processing model for real-time drone navigation and targeting.

PLA drone targeting systems

• ▪Researchers at China's Academy of Military Sciences used distillation to run a target-recognition model on tactical hardware during simulated maritime operations involving unmanned vehicles.
• ▪A 2024 paper from the PLA's National University of Defense Technology described using distillation to shrink an image-processing model for real-time drone navigation and targeting.

US semiconductor export controls

• ▪China has embraced model distillation to compete with the United States in frontier AI while facing constraints on advanced computing resources due to United States export controls on high-end chips.
• ▪Chinese central and local governments have promoted model lightweighting and edge computing, directing subsidies and research funding toward technologies that enable AI to run on limited processing hardware.

AI governance diplomatic tensions

• ▪In January 2026, researchers at the Army Engineering University published a paper proposing defense mechanisms to counter the security risk of data-free distillation.
• ▪The unauthorized extraction of capabilities via model distillation has emerged as a major flashpoint ahead of scheduled United States-China talks on AI governance and safety.

1 source