The Law of Stop

The essay is called 'The Law of Stop', it appears in Verfassungsblog, and its central demand is that AI governance acquire a legal framework for stopping systems that pose serious risks or cause harm. Full marks for the title; it has the virtue of saying what it means. The research rests on an analysis of forty AI governance instruments — forty documents read in search of a functioning off switch, with the author emerging to propose that somebody build one. Forty separate paragraphs explaining, with footnotes, that stopping is somebody else's department.

The one piece of binding text doing any work here is the EU AI Act, which requires providers of general-purpose AI models with systemic risk to report serious incidents to the AI Office. To the AI Office. Not to the public, not to a website, not to anyone who might be affected — to an office, which receives the tidings in the manner of a confession and files them. The essay notes, with admirable restraint, that publication is not required. So the arrangement is a duty of candour toward the AI Office and a duty of silence toward everybody else, which is roughly how you'd design a smoke alarm that only rings in the landlord's kitchen.

This is the natural sequel to the question Columbia Magazine asked about regulating an industry you don't understand: the follow-up is how you stop one, and the answer so far is a well-argued essay. Forty instruments, one law named after the brake pedal, and the stopping itself still delegated to whoever happens to be holding the plug. A stop mechanism is the sort of thing you would assume came free with the product, like brakes on a hatchback, and it instead requires peer-reviewed advocacy. Someone should draft the law of 'no' and see how far it gets before the consultation closes.