What an emergency stop switch for AI could achieve

Regulating Artificial Intelligence: What a Kill Switch for AI Could Achieve

AI development is moving at a rapid pace. Calls for regulation are growing louder — including calls for a so-called kill switch. But how might that work, and more importantly: who would be allowed to activate it?

A kill switch for artificial intelligence sounds simple and sensible at first. The idea: simply shut down a model or an AI agent that becomes dangerous. In practice, that could be considerably more difficult, experts say.

AI policy researcher Miles Brundage emphasizes the challenges involved. If, for example, all kinds of hospitals were using AI in a few years and it turned out that something was wrong with the latest model, shutting it down could already become a problem, Brundage told US broadcaster Bloomberg. In such a case, people might die, making shutdown hardly a viable option.

Technical Implementation

On the technical side, the question is also whether shutting down individual data centers would be sufficient — that might work at best for limited problems. Or whether a hardware solution would be needed, for example directly on the chips being used.

AI safety researcher Mark Nitzberg of the University of Berkeley told broadcaster CNBC: "Every chip could be delivered in such a way that it needs a specific code every minute, and if that code doesn't arrive, the chip stops working."

Who Decides on the Kill Switch?

Another important question: who decides whether the kill switch is actually activated? Many experts and politicians do not want to leave that decision to the companies.

AI pioneer Geoffrey Hinton goes one step further. He is not optimistic that the concept would still work once AI becomes smarter than humans, he stressed on US broadcaster CNN. AI could be very persuasive — and would convince those responsible that there is no reason to shut it down.

Safety Standards Demanded

In any case, the much-cited kill switch is just one of many control mechanisms being called for.

It is very difficult to demonstrate AI safety in advance because it is still a black box, says Stuart Russell of the University of Berkeley, president of the International Association for Safe and Ethical AI. Nevertheless, companies must be held responsible for scientifically demonstrating the safety of their models before they come to market, just as is done with pharmaceuticals.

Many experts also point out that international agreements are necessary when it comes to AI safety. Nevertheless, California Governor Gavin Newsom insists that regulations in California are also important. Many of the most advanced AI developers — including OpenAI, Anthropic, and Google — are headquartered there. He called on the Trump administration to take action.

Trump Dismisses Concerns

President Donald Trump has so far dismissed concerns related to AI, announced an "AI Force," and is searching on social media for a new name for artificial intelligence. Newsom himself, however, vetoed a comprehensive AI regulation bill just two years ago. Among other things, it had envisaged an independent government oversight body and a requirement for companies to incorporate a kill switch.

Newsom is nearing the end of his term as California's governor. However, he is being mentioned as a possible Democratic candidate for the next presidential election in 2028.