Chinese military researchers are using outputs from leading U.S. artificial intelligence models to train domestic defense systems [1, 2].
This development signals a strategic effort by Beijing to bypass lengthy development cycles by leveraging American innovation to enhance its own battlefield capabilities [1, 3].
Research institutions linked to the People’s Liberation Army have utilized a process known as model distillation [1, 5]. This technique involves using the responses from highly advanced "teacher" models to train smaller, specialized "student" models [5]. Researchers specifically targeted outputs from OpenAI’s GPT-4 and Anthropic’s Claude to refine their own AI systems, reports said [1, 4].
These domestic systems are intended for defense and battlefield applications [1, 2]. By distilling the logic and capabilities of U.S. models, China aims to keep its military technology on par with American advances [1, 3].
Evidence of this practice appears in academic and legal filings. Reuters reviewed more than 80 Chinese papers and patents that show military-linked institutions employing model distillation [6]. These documents suggest a systematic approach to integrating Western AI breakthroughs into the PLA's technical infrastructure [6].
The reports surfaced in late July and early August [2, 4]. The activity highlights the difficulty U.S. AI developers face in preventing their proprietary technology from being repurposed for foreign military use, even when direct access to the models is restricted.
“Chinese military researchers are using outputs from leading U.S. artificial intelligence models to train domestic defense systems.”
The use of model distillation allows China to extract the 'intelligence' of U.S. AI without needing the original source code or massive compute resources required to train a foundation model from scratch. This creates a significant challenge for U.S. export controls and safety guardrails, as the output of a model can be used as training data for a competitor's weaponized system, effectively subsidizing the adversary's military R&D with American intellectual property.
