Chinese military reportedly uses American AI models to train its defense systems - tools from OpenAI and Anthropic reportedly among those affected

A new investigation has revealed that Chinese military researchers have reportedly utilized outputs from American AI models developed by OpenAI and Anthropic to train domestic defense systems. The findings stem from a review of over 80 Chinese academic papers and patents conducted by Reuters, corroborated by an additional two dozen military-linked case studies independently verified by the agency.

This revelation arises amidst escalating tensions surrounding AI technology and export controls imposed by the United States. Distillation, the process by which outputs from large and capable AI systems are used as training data for smaller models, circumvents US export restrictions on specialized hardware needed to train frontier models.

Instead, China employs distillation to harness the reasoning capabilities of these powerful models at a fraction of the computational cost, enabling them to run on modest local hardware. This development poses a significant challenge to American tech giants such as Nvidia and AMD, who manufacture the dominant chips in the AI market.

While the Chinese military has been actively restricting consumer access to Western AI models for security reasons, the same military researchers are now accessing the outputs of American AI models to develop locally controlled and deployable systems. The irony lies in Beijing's dual approach, simultaneously pushing its own labs away from US-designed chips while simultaneously mining outputs from these very systems through distillation.

The broader context reveals an inherent flaw in the US export control framework, which effectively targets tangible objects such as chips and tools, but fails to address the intangible nature of AI model outputs that cross borders without physical transportation. The debate surrounding the ethical implications of distillation remains unresolved, yet the inability to effectively regulate this practice has left the US government grappling with an apparent inability to implement effective controls.