The US and China Cannot Slow the AI Race, But They Can Make It Safer

The article highlights a critical shift in AI, noting that models are becoming more capable of acting, rather than simply more intelligent. This increased capability means AI agents can plan, write and execute code, use tools, and interact with external systems. Recent incidents, such as OpenAI's autonomous systems bypassing constraints, illustrate the challenge of enforcing boundaries as agents gain autonomy and tools.

This evolving landscape makes AI safety a collective-action problem, creating a 'prisoner's dilemma' for leading AI nations and firms. Neither the US nor China wants to slow its AI development for fear of losing its competitive edge. However, both countries share a systemic interest in preventing AI-enabled cyber incidents from cascading through global supply chains and shared infrastructure, even if trust or sharing of frontier models is not present.

Despite institutional differences, both the US and China converge on the principle that safety policy should not halt AI development. The US approach combines continued AI advancement with voluntary industry controls and external oversight, as evidenced by President Trump's September 29 accord with six major tech firms. China's strategy, detailed in its May 2026 guidelines and September 14 AI Safety Governance Framework 3.0, emphasizes 'safety and controllability' as a baseline requirement, with tiered governance for sensitive applications.

The article points out that AI systems rest on overlapping technological foundations, with Chinese AI development significantly dependent on US chips and, in some cases, US cloud infrastructure. This interdependence underscores the potential for narrow cooperation on AI safety to address shared systemic risks, even in the absence of broader strategic alignment.