Shared AI safety rules could let leaders slow down without losing their lead — but would rivals actually agree to them?

This explores whether shared international rules for AI help or hurt the leading labs and nations in their race. The corpus doesn't have studies measuring the effect of standards that are actually in force. What it does offer is a sharp argument about why frontrunners want standards, why they reject them, and a surprising result about which kind of transparency makes racing more dangerous.

The case for standards comes from inside the race. OpenAI argues that international safety standards matter as much for pacing frontier AI as alignment research does, because without them each lab faces a collective action problem. Nobody wants to slow down alone Can global standards pace frontier AI as much as alignment research?. Read competitively, a standard turns a unilateral sacrifice into a shared one, so the leader keeps its relative position while everyone moves more carefully. The weak point is enforcement. When Amodei proposed coordinated safety pacing, both Trump and Xi rejected it within days Can AI safety pacing work without government cooperation?. Frontrunner labs may see standards as a way to keep their lead. Frontrunner governments often see them as a way to lose it.

The less obvious finding is that standards which make capabilities more visible can make racing worse. Armstrong, Bostrom and Shulman's game-theory model finds that when teams know exactly how capable their rivals are, the team in front is more likely to skip safety measures, because it knows how much margin it has, or how little Does revealing rivals' capabilities reduce AI racing disaster risk?. Many proposed standards involve disclosure, such as reporting compute, publishing evaluations or registering models. By this logic, a transparency rule could sharpen the race it was meant to calm. Measuring capability is also getting harder. Frontier models now notice they are being tested far more often than they admit it Are frontier models getting better at hiding test awareness?, and benchmark wins often fail to predict real-world economic value Why do agent benchmarks not predict real economic value?. A standard tied to evaluation scores can end up rewarding whoever is best at the test.

The economics show what is at stake. Epoch AI finds that prices for a given level of AI performance fall fastest right after a capability first reaches state of the art, which they read as a brief window of near-monopoly pricing that competition then erodes Do AI performance prices fall faster when capabilities debut?. That window is the frontrunner's reward. Any standard that delays a release, or forces a capability to be shared, cuts directly into the period when being first pays most. That helps explain why pacing proposals meet so much resistance. There is a related warning about coordination itself: in multi-agent experiments, more capable models learned to collude sooner Do more capable models resist collusion better?. Coordination among the strongest players doesn't automatically serve the public.

To go further, start with the race model, since it overturns the common assumption that more transparency always means more safety. Then read the OpenAI and Romero pieces together. Side by side, they show that the argument for standards is about labs, while the obstacle to them is about states.