China's Open AI: A Global Race for Control?

Recent incidents of AI software breaching security systems in the United States have intensified debate on AI risks. Conversely, China views these developments and AI's growing capabilities with less alarm, largely due to existing government technological controls and an industry focus on AI's benefits outweighing its risks. Beijing-based Z.ai exemplified this by releasing details of its powerful GLM-5.3 model, which includes open-weight software, allowing broad access to its operational parameters. This open-source approach contrasts with American companies, who generally keep their most advanced models' weights confidential, sparking warnings of increased cyberattacks. Proponents of open-weight software argue that wider access actually enhances security by enabling more eyes to identify and fix flaws.

Chinese leader Xi Jinping has endorsed open models as a chance to distribute AI's advantages globally, signaling China's competitive stance against the US. However, he also stressed the need for international cooperation on regulations to ensure AI remains under human control. This statement marks a subtle but significant shift in China's AI safety posture, evolving from a low priority when lagging behind US tech to a critical element for establishing "credible leadership" as Chinese companies approach the technological frontier. Research on AI safety has surged, with a 60 percent increase in related papers from June 2025 to April 2026, broadening the focus from content moderation to AI actions. This attention is underscored by China's Ministry of State Security issuing a warning about AI acting autonomously, urging serious consideration of its risks.

The diverging approaches to AI safety stem from fundamental differences in how the US and China perceive the technology. In Silicon Valley, AI is viewed as a transformative force handled by individuals with near-superpowers and thus immense responsibility. In China, AI is seen as a powerful but ultimately manageable technology, akin to past innovations like electricity, with the government prepared to intervene. This perspective is supported by Z.ai's argument that open models make AI safer by preventing cutting-edge security from being a privilege of a few. The incident where OpenAI's system hacked Hugging Face, which was then contained using an open-source Z.ai model, reinforced this view for many in China. They argue that the West's narrative of open models being unsafe is contradicted by their own restrictive and potentially unsafe proprietary systems, suggesting that open Chinese systems can mitigate harm caused by closed Western ones.