A model built for the world cannot be licensed by a city
A frontier model is trained once and served everywhere. City-by-city validation will not make it safer. It will make it unavailable, or make the rule a formality.
Nobody in the council chamber on 5 October argued that the risks are imaginary. The witnesses from the four companies said, in different words, that no one could promise a risk-free system. The question is not whether frontier AI needs rules. It is whether a city council is the body that should write them.
A model is trained once and served everywhere at the same moment. Speaker Julie Menin's bill would make it unlawful to sell or deploy one in New York City without third-party validation, with the criteria set by the city's Cyber Command and a $25,000 penalty per instance. Multiply that by every city and state with a view, and a developer would face a crowd of validators applying different standards to the same system.
The bill does not say what an instance is. For a product used by millions of New Yorkers, that is the difference between a nuisance and a ruinous fine, and the bill leaves the choice to whoever enforces it. Good law tells a company in advance what compliance looks like.
There is a competence problem too. Judging whether a frontier model can be turned to a cyberattack or a weapon is specialist work done by a small number of teams, some of them in national governments. Asking a municipal cyber agency to accredit those validators, and to keep pace with every release, is asking it to do something it was not built for.
Meta's witness asked the council for a uniform federal standard. That is the right answer, and the companies should be held to it: a national validation regime, written by Congress, that binds every developer and every new entrant alike. New York can make the loudest case for one. It should not mistake a local ordinance for a solution to a global technology.