China Is Setting America’s Tolerance for AI Risk

Leading US artificial intelligence developers are framing some of their safety commitments around their competitive position relative to international counterparts. Anthropic's policy, for instance, allows the company to delay a highly capable model to ensure catastrophic risks are contained, but this obligation ends if it no longer believes it holds a significant competitive lead. This suggests that the perceived pace of rival AI development could directly influence when American companies deem it acceptable to release a new system.

OpenAI's April 2025 Preparedness Framework also reflects this dynamic. It permits adjustments to existing safeguards if another developer, regardless of location, achieves specified dangerous capabilities without comparable protections. While the framework stipulates that overall risk of severe harm must not meaningfully increase, and OpenAI's protections must remain stronger, these provisions give foreign capabilities an explicit role in American companies' safety decisions.

This approach means that if a breakthrough by a competitor, such as a Chinese firm, is perceived, American developers might release models they previously considered too dangerous. This dynamic effectively allows competitor progress to influence the risks Americans face from AI systems. However, reliably assessing these capabilities is crucial, as shown by the US government's Center for AI Standards and Innovation, which found DeepSeek V4 Pro's performance closer to an eight-month-older American model, despite the Chinese developer's published claims of parity with more recent US systems.