Revealed: China's Super-Advanced Weapons Are Actually Made in America
The People’s Liberation Army (PLA) and affiliated research bodies in China have been using advanced artificial intelligence models from American firms like OpenAI and Anthropic to sharpen the country’s weapon systems and domestic defence capabilities. The finding comes from a Reuters review of over 80 Chinese academic papers and patent documents, revealing a shortcut for China’s military to develop sophisticated warfare systems while bypassing Washington’s strict chip export restrictions.
Documents show Chinese military researchers are employing a technique called ‘model distillation’. This process extracts the output or reasoning logic from massive US AI models, such as OpenAI’s GPT-3.5 or Anthropic’s Claude 3 Haiku, to train smaller, specific local AI models. As a result, China does not need to build a frontier AI from scratch, a process requiring massive funding and thousands of advanced computing chips currently under US sanctions. ‘Teaching a model the right answer is easy, but teaching it the reasoning behind that answer is far more difficult,’ said Sunny Cheung, a researcher at the Jamestown Foundation. ‘Chinese military researchers are trying to transfer the expensive reasoning logic of Western models into cheaper, locally controlled systems.’
The use of distilled US AI by the Chinese military spans several critical sectors. PLA Intelligence Unit 96941 used GPT-3.5 to summarise sensitive military source code to train an internal, internet-isolated AI model. The National University of Defence Technology successfully miniaturised an image-processing model for installation on combat drones, enabling real-time target mapping even when communication is lost. The Chinese Academy of Military Sciences ran a target recognition model on unmanned ships, drones, and submarines during combat simulations. Additionally, the North University of China used Anthropic’s Claude 3 Haiku to train an AI for monitoring content and social media.
The findings have intensified geopolitical tensions between Washington and Beijing over global AI security governance. The US accuses China of unauthorised extraction that violates intellectual property rights and undermines export controls. Anthropic stated it has never provided commercial access to China and warned that distilled models can lose safety safeguards, making them dangerous in military hands. The Chinese government denied the accusations, countering that the US is pursuing AI ‘hegemony’. Top Chinese AI startup Moonshot also denied its Kimi K3 model used distillation from US models, claiming it is the result of independent innovation. Despite successfully circumventing US sanctions, experts caution that distilled models only inherit specific capabilities and cannot yet match the full intelligence level of the parent US AI systems.