Despite US Restrictions: Chinese Military Continues to Use American AI

Despite US restrictions on China regarding advanced artificial intelligence (AI) technologies, Chinese military researchers have found a way to utilize models developed by OpenAI (ChatGPT) and Anthropic (Claude). A Reuters investigation revealed that Chinese military and security institutions have used the outputs of these models to train local AI systems and strengthen their military capabilities.

Reuters reviewed more than 80 Chinese academic papers and patents, revealing a rare glimpse into how Chinese military and security institutions rely on American AI models. Researchers use these tools as a shortcut to develop their own dedicated systems, despite Washington's attempts to impose restrictions.

"Distillation" of AI Models

The documents showed widespread use of a technique known as "Model Distillation." This involves using the outputs of a powerful AI model, such as ChatGPT or Claude, to train smaller and more specialized models that are available to the local military without restriction. This way, the Chinese do not need the massive computing power required to build advanced models from scratch.

However, "model distillation" is not a perfect solution. While it is an effective way to narrow the gap with the United States in AI, experts emphasize that this technology transfers only a portion of the original models' capabilities and does not provide full independence from advanced AI systems, which limits their use in tools like drones, satellites, and other devices.

To address this obstacle, Chinese researchers at the University of Military Engineering published a study dealing with the dangers of what is called "Data-free distillation" — a technique that allows for the reverse engineering of model capabilities without direct access to their underlying parameters.

But this solution also has limitations. In a recently conducted study, researchers argued that data-free distillation, to enable the reverse engineering of model capabilities, could be highly insecure. According to the researchers, reverse engineering in this way could expose all the logical information that the models generated — something that China does not intend to allow.