Chinese Defense Researchers Tap U.S. AI Models to Enhance Military Capabilities

Chinese military and security-linked institutions are utilizing advanced artificial intelligence models developed by American companies like OpenAI and Anthropic to train their own domestic systems, according to a recent review of over 80 academic papers and patents by Reuters. This revelation offers an unprecedented look into how Chinese researchers are leveraging cutting-edge U.S. AI technology as a shortcut in developing specialized military applications.

The findings highlight the widespread use of a technique called "model distillation," which involves using outputs from powerful AI systems to train smaller, more localized models that can be deployed without requiring extensive computing resources. This approach allows Chinese institutions to bypass some of the technological hurdles they face due to Washington's efforts to restrict Beijing’s access to advanced chips and other strategic technologies.

Reuters' review included research compiled by the Jamestown Foundation in Washington, D.C., which revealed that military-linked researchers are extensively using model distillation. The documents indicate that these researchers see leading U.S. AI models as a source of technical insight and a means to quickly close the gap with their American counterparts.

The issue has become a significant point of contention ahead of upcoming talks on AI governance and safety between the United States and China. U.S. officials have accused Chinese entities of using distillation to extract capabilities from American AI models, potentially undermining export controls and intellectual property rights. In response, China has denied these accusations, arguing that Washington is pursuing an aggressive stance on AI while pointing out that similar practices are employed by U.S. firms.

Despite the controversy, Chinese developers maintain that their advancements in AI do not rely solely on foreign models. For instance, Moonshot, a prominent Chinese AI startup, recently refuted allegations from the Trump administration that its Kimi K3 model was built using distillation techniques, asserting instead that it is based on proprietary innovations.

Sunny Cheung, a fellow at the Jamestown Foundation who analyzed over 60 of these papers, noted that Chinese military scientists are systematically capturing and adapting the reasoning processes of Western models for various applications. "Teaching a model the right answer is one thing," said Cheung, "but teaching it the reasoning behind the answer is much harder."

One notable example is a paper published by researchers in PLA Unit 96941, a military intelligence and cyber-warfare unit based in Beijing. The study detailed how OpenAI’s GPT-3.5 was used to process sensitive military source code before training a domestic model on these summaries for deployment within Chinese networks.

The use of U.S. AI models extends beyond intelligence gathering to areas such as content monitoring and tactical decision-making. At the North University of China, researchers utilized Anthropic's Claude 3 Haiku to generate synthetic data for text classification models aimed at social media monitoring and content moderation. Meanwhile, PLA’s National University of Defense Technology used distillation to reduce an image-processing model for deployment on unmanned aerial vehicles, enabling real-time analysis even when communication channels are disrupted.

Despite these revelations, the White House, Pentagon, China’s foreign ministry, the People's Liberation Army (PLA), and OpenAI did not provide comments when contacted by Reuters. The implications of this practice underscore the complex interplay between technological innovation and national security in an increasingly competitive global landscape.