China shares parts of America’s AI-risk diagnosis but rejects an American-designed slowdown

Anthropic chief Dario Amodei’s proposal for coordinated AI pacing arrives in China bundled with measures to restrict advanced chips, semiconductor equipment, smuggling and remote access to foreign data centres. Beijing’s rejection is therefore predictable: Chinese officials and engineers read the package as an attempt to preserve American technological dominance, not as a neutral safety bargain.

The disagreement is less about whether advanced systems can be dangerous than about who bears the cost of restraint. China’s latest national Safety Governance Framework warns that systems could learn autonomously and improve recursively beyond human expectations or control. Security officials also emphasize cyberattacks, information manipulation and military applications. The vocabulary overlaps substantially with concerns inside US laboratories.

But Chinese developers occupy a different competitive position. A MiniMax engineer asked why American firms got to pursue AGI and then change the rules once ahead. A DeepSeek engineer argued that the greater dystopia would be a few western companies controlling scarce frontier intelligence, and called for powerful systems to be cheap and widely available. Huawei chair Eric Xu said Chinese developers may need to catch up before they can even observe the frontier risks described by US labs.

A workable agreement must separate reciprocal safety measures from unilateral industrial containment. Incident notification, shared evaluation protocols and verified restrictions on clearly dangerous deployments could benefit both sides. A plan that asks China to accept a permanent compute disadvantage while US companies retain the lead will be interpreted as strategy wearing the language of safety—even where the underlying risk assessment is genuinely shared.