The comparison is unfashionable among technologists and inevitable among diplomats: the last time humanity built a technology whose worst case was civilisational, it answered with a treaty architecture — imperfect, gamed, endlessly renegotiated, and yet the reason the worst case stayed hypothetical. Frontier AI now poses the same institutional question, on a faster clock.
The Geneva draft's critics are right about its weaknesses: evaluation without enforcement, a registry without teeth, reservations wide enough to drive a training run through. They are wrong about the implication. The non-proliferation regime began as exactly this kind of document; its power accumulated through decades of verification practice that no first draft could have specified.
What cannot be recovered later is the norm itself — the shared expectation that frontier systems are evaluated before they are deployed. Norms are cheapest at the moment of technological adolescence and become unpurchasable after the first catastrophe. The drafting time is now; the ratifying time is short.