Abstract
Frontier AI models—cutting-edge foundation models whose capabilities generalise across domains, improve rapidly, and extend beyond what developers explicitly trained them for—and the systems they integrate into pose regulatory challenges qualitatively distinct from those of narrower AI applications. China is well placed to take on these challenges, as over the past decade it has built substantial AI governance infrastructure that can serve as a foundation for governing frontier AI.
As frontier capabilities advance, this infrastructure can be calibrated to focus regulatory attention on the most capable systems where the distinctive frontier AI risks arise while preserving space for the application-layer development that represents much of China's AI ecosystem. This paper examines how to do so, focusing on China's institutional context while drawing on comparative experience from the US, UK, and EU.
We identify nine proposals across three areas—frontier risk evaluation, emergency response, and liability adaptation—that build on China's existing infrastructure to address frontier risks, plus secondary mechanisms suited for pilot-zone experimentation.
Authors
Emmie Hine, Zhu Yue, Alex Jumper, Jeff Liu, Ian Read, Saad Siddiqui, Raymond Wang, Li Wenlong, James Zhang, Christoph Winter, Zhou Hui
