China Hasn't Decided Whether to Curb Open-Weight AI Models

Xi Jinping used a July 17 speech at Shanghai's World AI Conference to endorse both "open source and openness, cooperation and sharing" and to "strengthen risk awareness, and ensure security and controllability." Graham Webster argues in TechPolicy.Press that the pairing is unresolved tension, not a blessing for the country's open-weight labs. Three AI experts in our Who's Who directory have shared the piece.

The tension is old. Xi's 2014 formulation, "development is the foundation of security, and security is a precondition of development," is echoed in China's current generative AI regulation, which instructs the state to "maintain equal emphasis on development and security." Webster's read is that Beijing has not decided where the security threshold sits for models like DeepSeek V4, GLM 5.2, Kimi K3, and Alibaba's Qwen line, which he places roughly six months behind US frontier systems.

He walks through four hypothetical triggers. Political dissent is unlikely, he argues, because "getting online in China and communicating over the platforms that reach large numbers of Chinese citizens entails multiple layers of identity verification and surveillance." Cyber risk is live but bounded: a US-UK analysis found that Kimi K3's "safeguards did not prevent it from attempting cyber exploit development or offensive cyber operations," though the Chinese models trail their US counterparts. A biological or chemical proof of concept, "hopefully in a lab and not in the wild," could shift policy fast. And Beijing may prefer to be the fast follower, expecting Chinese officials can "see what kind of trouble awaits and swerve before they get there."

The closing move is a caution against Western analysts who keep predicting a Chinese pivot. Betting on a crackdown, Webster argues, may say more about the observer than about Beijing.