Chinese military researchers have utilized outputs from leading U.S. artificial intelligence models, including those developed by OpenAI and Anthropic, to train domestic systems for defense purposes. This finding is based on a review by Reuters of over 80 Chinese academic papers and patents, along with additional research compiled by the Jamestown Foundation, based in Washington.
The key method known as model distillation allows researchers to train smaller, specialized AI models using outputs from powerful systems, reducing the need for extensive computing power. Reuters found that this technique is widely used by researchers associated with the People’s Liberation Army (PLA) and other military institutions, as they leverage U.S. models for technical insights to enhance their capabilities.
Applications of this technique are varied. For instance, a PLA cyber-warfare unit used OpenAI’s GPT-3.5 to summarize sensitive military code, keeping classified work within Chinese networks. The National University of Defense Technology employed distillation to create a compact image model for drones, enabling real-time targeting even with disrupted communications. Other uses include maritime target recognition and social media monitoring.
Source: Reuters
So What
Distillation likely allows China to address specific gaps in capabilities without requiring the computing power needed to develop advanced AI technologies from the ground up. It serves as a shortcut rather than a means of achieving complete independence. Distilled models only inherit selected abilities and remain weaker than the original systems they emulate. This issue is expected to be a significant point of discussion during the upcoming AI governance talks in September. Tighter U.S. controls over access to models and their outputs are anticipated, although completely preventing the process of distillation is unlikely.
Follow us to join the intelligence community!