Frontier AI companies publish capability thresholds, the levels at which they say a model becomes dangerous enough to require specific safeguards. The thresholds differ substantially between companies, which creates two problems.

The first is verification. An outside party trying to determine whether a threshold has been crossed has no common yardstick, and cannot compare one company's commitments against another's. The second is the incentive structure that follows: with no shared minimum, whoever sets the loosest threshold faces the fewest obligations, which is the race to the bottom the authors name.

That second point is why this is a coordination problem rather than a technical one. No individual company can fix it by tightening its own threshold, since that only widens the gap with competitors, which is precisely the situation where a harmonisation methodology has something to offer.