Chinese military-linked researchers used outputs from leading U.S. artificial intelligence models to train specialized domestic defense and security systems [1, 2].
This activity suggests that despite efforts to build indigenous technology, China is leveraging American innovation to accelerate its own military AI capabilities. By utilizing models from companies such as OpenAI and Anthropic, researchers can bypass some of the immense compute requirements needed to build high-performing systems from scratch [1, 2].
According to a review of more than 80 Chinese papers and patents, the researchers employed a technique known as model distillation [1]. This process allows capabilities from larger, compute-intensive AI systems to be transferred into smaller models [2]. These smaller models can then operate with fewer computing resources, making them more practical for deployment in defense and security applications [1, 2].
"Chinese military researchers have tapped outputs from leading U.S. AI models to train defence systems, a review of more than 80 Chinese papers and patents shows," Reuters said [1].
The use of this technique indicates a strategic effort to optimize resource usage. The distillation process enables the military-affiliated institutions to maintain high performance while reducing the hardware overhead required for operation [2].
Reports on the matter highlight a tension between China's public push for technological self-reliance and its actual research practices. While China has released its own large-scale AI models to demonstrate indigenous capability, the evidence from these academic papers suggests a continued reliance on U.S.-developed logic and outputs to refine its military tools [1].
"The technique allows capabilities from larger AI systems to be transferred into smaller models that can operate with fewer computing resources," Reuters said [2].
“Chinese military researchers have tapped outputs from leading U.S. AI models to train defence systems”
This development underscores a paradox in the AI arms race: while the U.S. and China compete for dominance, the 'teacher' models from the U.S. are effectively training the 'student' models of the Chinese military. Model distillation allows an adversary to capture the reasoning and knowledge of a frontier model without needing the same massive datasets or compute power, potentially eroding the strategic advantage provided by U.S. hardware and software restrictions.
