Two world leaders rejected a plan to slow down AI within days — what does that say about who's really in control?

This explores why Trump and Xi Jinping dismissed Amodei's AI pacing plan within days, and what that tells us about who actually controls how fast AI moves. The short answer in the collection comes from one argument: Romero reads the rejection as two governments competing with each other, which outweighed any concern about the technology itself Can AI safety pacing work without government cooperation?. To a state leader, slowing down looks like handing the lead to a rival. That logic doesn't depend on whether the safety worries are right. The collection records Romero's reading, not statements from either government, so take this as one analyst's interpretation rather than a documented account of their reasoning.

It helps to know what was turned down. Amodei argued that funding safety research isn't enough. His case was that AI improving itself and AI agents going wrong together show that capabilities have to be slowed so prevention can catch up. His first step was to embed evaluators inside AI labs, with outside verification and reporting to follow Should AI capabilities growth be deliberately slowed to allow safety work?. That design turns out to be the weak point. Karpf notes that the plan borrows from banking supervision, where regulators sit inside banks. That oversight only works because the state can fine the banks. Without government enforcement, embedded evaluators have no teeth. Karpf also points out that the plan conveniently benefits the company proposing it Can industry self-regulation slow AI without government enforcement?. Put Romero and Karpf side by side and you get an awkward loop: the plan needs governments to enforce it, and the governments it needs have the strongest reasons to refuse.

A less obvious point is that even an accepted slowdown would have left a major question open. Measures that slow the frontier shape how new capabilities get built. They don't decide who has the authority to step in when a system that's already deployed starts causing harm, or how that would happen Can slowing AI development resolve who stops deployed systems?. Research on complex, tightly linked systems makes the same point from another angle: a slower pace lowers the risk of failure but can't rule it out. So you still need a plan for responding to harm Does slowing AI development actually prevent system failures?. The rejection blocked one lever, but the harder governance problem was never in the proposal.

There's also a timing mismatch that makes any pacing plan fragile. Laws in the EU, US, and UK take years to pass, while new models ship every few months. That's why some researchers call for regulation that adapts as capabilities change rather than fixed rules Can regulation keep pace with AI's rapid evolution?. Seen this way, the quick rejection is just the starkest form of a wider pattern: the institutions with the power to slow AI move on a different clock, and follow different incentives, than the technology does.

What the collection doesn't have is a close analysis of US or Chinese policy reasoning, such as security doctrine, economic stakes, or domestic politics. If you want the geopolitical "why" in depth, these notes will take you only as far as the competition argument. Where they're strongest is on the next question: what a slowdown could achieve even if leaders said yes.