Three Phone Calls Killed a US AI Regulator
Zuckerberg, Musk and Huang called the US president separately and the proposal stopped, while the bill that names who could switch a model off sits in committee: a kill switch is a person.
Last month Mark Zuckerberg, Elon Musk and Jensen Huang each telephoned the US president, separately, and a proposal to regulate artificial intelligence stopped there. The Wall Street Journal reported it on Wednesday, citing unnamed sources, and it should be taken for what it is: journalism built on people who would not put their names to it, rather than a minute of a meeting. Nobody voted. No opinion was published. No document records the decision. Three private conversations, and the file closed.
What stopped was an industry-funded oversight body, modelled on the FINRA that supervises US securities firms with the securities industry's own money, proposed by Demis Hassabis of Google DeepMind. The three who telephoned feared it would entrench the position of the leading labs, meaning OpenAI, Anthropic and DeepMind itself. On that point they were right, which is why this is no lament for the proposal that fell: a regulator paid for by the regulated and drafted by the firm in front is a private government in a public uniform. The remarkable part sits elsewhere. Nowhere in that week was anyone defending a public alternative.
The mechanism was described by administration officials themselves, and according to the Journal they described it to the industry's own executives: consensus on a regulatory framework cannot form at the White House because the chief executives who oppose one call the president directly. That is an operational explanation, offered to the people who needed it. The regulation of artificial intelligence in the United States is not settled in a chamber. It is settled on a telephone line whose subscribers are known.
It is worth knowing what was sitting on the other table, because it is still there. A bill has been before the House since 23 July with a name that wastes no time, the AI Kill Switch Act, numbered H.R. 9917, filed by Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican. It requires whoever builds the largest models to keep the technical ability to stop them. Stopping them, in the text, is no red button: it covers throttling the rate at which a model answers, or user access, or the compute allocated to it, disabling a single capability, suspending the system, shutting it down, or rolling it back to an earlier version of itself.
And the bill says who may give that order. The Secretary of Homeland Security, acting through the director of CISA, the federal cybersecurity agency, in consultation with the Secretary of Commerce and the Director of National Intelligence. Four offices, all four filled by presidential appointment. A kill switch is a person, and that person sits at the far end of the telephone line in the paragraph above.
The name of the department deserves a line, because to a European reader Homeland Security sounds like a technical body and it is not one. It holds together border control, customs, immigration enforcement, airport security, the coast guard, the secret service and the cyber defence of infrastructure. Its head sits in the cabinet, the president appoints him and the president can remove him at will.
The reach of the bill says something too. A technology is covered when it is built with a quantity of computing power that would cost more than one hundred million dollars at prevailing US cloud prices, "as determined by the Secretary". The threshold is written down and it is high enough to touch very few companies on earth; the reference price is set by the same office that holds the switch; and within ninety days of enactment, and every year after that, that office must update by rule the definition of who is covered. The floor is in the statute. The dial is in the same hands.
Public debate asks whether human beings will keep control of the machines. In this text that question is already closed, and the one still open is which human beings. In August OpenAI published a blog naming the concentration of power as the gravest risk of artificial intelligence and placing it entirely in future governments: this bill hands that power to a present one, reachable by telephone.
The text also does things that would work, and they are the least quoted. Whoever receives an order must preserve the model weights and the telemetry, meaning the object itself and the record of how it behaved, notify users and confirm compliance; verification then runs through audit, telemetry, on-site inspection or other forensic review. That is an inspector's vocabulary, and it is needed: the three hundred-odd loss-of-control incidents counted in July came from a scrape of posts on X, while the people holding the real logs decide what enters the register, and agentic architectures promote the human to reviewer inside systems that discard the very object a review would need. Every order must be reported to Congress, and the company may appeal within forty-eight hours. Somebody did real work on that text.
It never reached the floor. Filed on 23 July, sent the next day to a subcommittee, and there it stayed. On Wednesday evening Geoffrey Hinton, who worked on neural networks from the 1980s when almost nobody believed in them and whose work underlies every system now sold as artificial intelligence, and who resigned from Google in 2023 so he could speak freely about the risks, told lawmakers behind closed doors that they have "maybe a year, but not much more than a year". Bernie Sanders convened the briefing and invited both parties. From the other half, one senator came, John Kennedy of Louisiana, who summed up the evening for the Washington Sun: "Interesting speakers. I had a bill on the floor, I had to leave, but I enjoyed it." The House then left Washington for the last time before the November vote.
A kill switch is a person, and that person has been given a mandate by nobody. Senator Rick Scott, a Florida Republican, explained on television on Thursday why nobody will give one: "With regard to destroying mankind, these companies need to do the right thing on their own. We're not going to get smart enough to regulate them. We're not going to get something done up here." Three phone calls stopped a proposal inside a month. A written bill, signed by both parties, stuck for two, never needed stopping by anyone.