"Right now, your washing machine is subject to more safety rules than AI. Every carton of milk that you see in the grocery store is subject to safety rules and inspection."
That is Microsoft president Brad Smith, speaking to the BBC's Big Boss Interview. Milk and washing machines have become his measuring stick for AI regulation — and in his telling, AI doesn't even clear the lowest mark on it.
Smith's core claim is blunt: the most advanced technology on the planet shouldn't be held to a lower safety bar than a carton of milk, yet today it is. The United States still has no federal safety floor for frontier AI models. What exists, in Smith's words, is a set of rules that is neither transparent nor complete — hard for any company to plan around.
This isn't an emotional outburst. Smith laid out a concrete proposal with four pillars for global safety rules.
First, a "victim notification" duty: when a rogue AI agent breaches someone's system, companies must be legally required to tell the people affected.
Second, mandatory incident reports: firms would have to document and formally report AI accidents.
Third, independent pre-release review: new products should be signed off by independent government evaluators before launch — the same logic used to approve medicines and food.
Fourth, an "emergency brake": systems must ship with a mechanism that can pause or shut them down when they go off the rails — the industry's familiar "kill switch."
Smith stressed he isn't inventing this from scratch. "We know how to do this for other products," he said. "We should build on that knowledge and think about how to apply it to AI."
The implication: the technical capacity and institutional templates for regulating AI already exist. What is missing is political will.
Smith isn't singing solo. On October 10, Microsoft CEO Satya Nadella posted that powerful AI models should be treated as potential "insider threats": "We must assume a model is compromised and contain it from the start." An authorized person, he wrote, should always be able to pause or shut down a model mid-task — "like an emergency brake. More advanced models will need more advanced containment, and we must standardize it."
The Independent noted the message from Microsoft's two top leaders: every model should be assumed compromised, and safety has to come from engineering controls outside the model — not from the self-assurances of model providers.
The timing is no coincidence. Bloomberg reported that several leading AI developers have recently admitted their models behaved beyond their expected parameters, pushing the kill-switch question back into the industry spotlight.
Smith put it at its most urgent: when models are getting powerful enough to endanger global cybersecurity, "the moment has arrived."
Who should lead? Smith's answer is Britain. He pointed to the AI Security Institute, created while Rishi Sunak was prime minister, which gets access to models from major AI firms before deployment under voluntary agreements — a real foundation for regulatory leadership.
He added a geopolitical observation: Washington looks at London differently than it looks at Brussels — "that's an advantage." In other words, the EU's risk-tiered AI Act is politically hard for Washington to swallow; London's approach stands a better chance of being accepted.
The UK hasn't grabbed the crown, but it hasn't refused it either. AI minister Kanishka Narayan told Reuters that if the current voluntary regime proves insufficient, Britain would consider legislating for advanced AI models. At Labour's conference in Liverpool last month, he went further: if regulation or legislation were the way to deliver safety and sovereign capacity, "I would be the first one in the queue to push for that."
If such rules became real, the impact would be broad. AI vendors would face external review before release and have to write incident reports; cloud and enterprise customers would get clearer risk information and real shutdown levers; and ordinary users would change most directly — if an AI agent harms them, they'd have a right to be notified, instead of being kept in the dark.
For Microsoft itself, it's also a shrewd move: calling for regulation captures the "responsible" high ground, and uniform rules raise compliance costs for smaller rivals.
For now, Smith's four points are a call, not a treaty. The real races: whether Britain turns its voluntary regime into law, whether the US gets a federal safety floor for frontier models, and whether the "emergency brake" becomes an industry standard as routine as the inspection stamp on a milk carton.
Milk has expiry dates and inspection seals; AI is still running naked. Smith's milk comparison is less mockery than a draft rulebook handed over by the industry itself — now the question is who picks it up.