The art of not making an AI safety deal with China

The resumption of US-China AI safety talks is not about building trust or striking a grand deal. It aims for mutual visibility into each side's AI fears and management strategies. Beijing's primary concern is that safety discussions will become a pretext for new restrictions, similar to past national security justifications for blacklists and export controls. This dynamic makes genuine cooperation difficult, despite the shared existential risks.

For Asia, the key is how these talks influence the internal race dynamics among AI developers. China's approach to AI safety is evolving, with a 60 percent increase in monthly frontier AI safety papers from June last year to April. This shift from political controls to agentic application risks reflects a serious engagement with safety. The challenge is to ensure this focus is not misinterpreted as an acceptance of US-imposed rules.

The thing to watch is the level of transparency achieved. Hugging Face CEO Clement Delangue called for "100x more transparency in AI" after a recent OpenAI agent-related cyberattack. Without greater visibility into each side's AI capabilities and safety measures, both the US and China will assume the worst. This could accelerate a reckless race for more powerful systems, making Asia's AI development environment riskier.