Chinese military-linked researchers have used outputs from advanced artificial intelligence models developed by United States companies to train domestic AI systems aimed at strengthening defence capabilities, according to a review of academic papers and patents.
The findings revealed that institutions connected to China’s military and security sectors have been using a technique known as “model distillation”, where a powerful AI system’s outputs are used to train smaller and more specialised models that can operate with fewer computing resources.
The review of more than 80 Chinese research papers and patents found that the method has been widely adopted by researchers linked to the People’s Liberation Army (PLA) and other defence institutions.
The practice has raised concerns in Washington, with US officials alleging that some Chinese entities may be using distillation methods to extract capabilities from American AI models, potentially affecting technology controls and intellectual property protections.
However, China has rejected such claims, accusing the US of attempting to maintain dominance in the AI sector while arguing that American companies have also benefited from similar approaches.
Researchers said Chinese defence institutions view leading US AI models as a source of technical knowledge and a way to narrow the technological gap with global competitors.
A researcher from the Jamestown Foundation, Sunny Cheung, said Chinese military scientists were attempting to capture the reasoning processes of Western AI models and adapt them for uses including surveillance, cyber operations and tactical decision-making.
One study by researchers from PLA Unit 96941, a military intelligence and cyber-warfare unit based in Beijing, reportedly used OpenAI’s GPT-3.5 to analyse military-related software code before training a domestic AI model using the generated summaries.
The researchers said external AI systems were unsuitable for processing classified information, leading them to develop a locally operated model that could function within Chinese military networks.
The use of AI distillation has also expanded beyond defence applications, with Chinese researchers applying the method for areas such as content monitoring, social media analysis and unmanned systems.
Researchers at China’s North University, which has links to the country’s defence industry, reportedly used Anthropic’s Claude 3 Haiku to create synthetic training data for a text classification system designed for monitoring online content.
Meanwhile, a 2024 study from the PLA’s National University of Defense Technology explored using distillation to reduce the size of image-processing AI models for unmanned aerial vehicles, allowing drones to analyse video and support navigation and targeting even when communication links are disrupted.
Experts noted that while distillation allows AI systems to become smaller and easier to deploy on devices with limited computing power, it does not fully replicate the capabilities of larger frontier models.
Trevor Koverko, co-founder of AI data company Sapien, said distilled models should be viewed as a transfer of selected capabilities into cheaper and locally controlled systems rather than complete independence from advanced AI technology.
The issue has become a major point of discussion ahead of US-China talks on AI governance and safety, as both countries continue competing over the development and regulation of emerging technologies.
