Chinese military tapped US AI models in 80-plus defense papers

Chinese military researchers used OpenAI and Anthropic model outputs to train defense systems, a Reuters review of 80-plus papers found, exposing gaps in US export controls.

"The deeper a country climbs into that stack, the harder it becomes to climb out," Nvidia Chief Executive Jensen Huang said late last year, warning about China fusing its Belt and Road Initiative with AI infrastructure. The Reuters review suggests Beijing's military researchers have found pathways around US restrictions to access frontier models.

The papers and patents show Chinese researchers using outputs from OpenAI's GPT models and Anthropic's Claude to train domestic AI systems for defense applications. The findings come as China's gap with US frontier models has narrowed from roughly nine months to four or five months in little more than a year, according to Britain's AI Security Institute. Chinese labs including Moonshot AI, DeepSeek, Alibaba's Qwen, and Zhipu AI have all demonstrated rapid progress.

The revelations could trigger regulatory scrutiny and potential compliance actions against OpenAI and Anthropic, while pushing Washington toward stricter export controls on AI technology. For investors, the news heightens geopolitical risk around major AI stocks and could limit these companies' international operations. The stakes extend beyond individual firms: the countries that build AI infrastructure today will shape tomorrow's technical standards, government procurement decisions, and ultimately long-term economic growth.

Export Controls Face a Compliance Test

OpenAI's terms of service prohibit military use, and Anthropic's acceptable use policy similarly bars defense applications. The Reuters review raises questions about whether these companies have adequate safeguards to detect and prevent such usage. Model outputs — unlike physical chip shipments — are difficult to track, representing a significant enforcement gap in the US export control regime.

The US has assembled a range of countermeasures, including Pax Silica for semiconductor supply chains, FORGE for allied industrial ecosystems, and the emerging AI Spark initiative to offer trusted partners access to the full AI stack. Yet these programs remain largely supply-side policies. They strengthen US technological competitiveness but do not yet amount to a comprehensive strategy of AI diplomacy.

China's AI Gap Narrows to Four Months

The military applications come as China's civilian AI ecosystem closes the gap with US frontier models. Moonshot AI's Kimi K3 frontier model, following breakthroughs from DeepSeek and rapid progress by Alibaba's Qwen and Zhipu AI, demonstrates that Chinese labs are producing models competitive with ChatGPT, Claude, and Grok.

Beijing has also been building a broader AI diplomacy apparatus. The proposed World AI Cooperation Organization (WAICO), supported by AI application centers, open-source models, and technical training, represents the next evolution of China's international strategy, built directly on the Digital Silk Road. Countries including Indonesia have already endorsed the initiative, while Chinese firms including Alibaba and JD.com have invested heavily in Southeast Asian e-commerce platforms.

The strategic stakes are difficult to overstate. The countries that build AI infrastructure today will shape tomorrow's technical standards, software ecosystems, government procurement decisions, and ultimately long-term economic growth. If Washington fails to close the enforcement gap exposed by the Reuters review, the competitive advantage of US AI leadership could erode faster than current policy anticipates.

This article is for informational purposes only and does not constitute investment advice.