Microsoft's Brad Smith says AI has fewer safety rules than milk -- days after the White House briefly cut Britain off from frontier models

Milk is regulated more strictly than artificial intelligence. That's the comparison Microsoft president Brad Smith used on a BBC podcast published October 12 to argue that the world's most powerful technology is also one of its least supervised -- that a carton of milk, or a washing machine, clears more mandatory safety checks before it reaches a shelf than the systems now running inside hospitals, trading desks and government agencies. His conclusion: "the moment has arrived" for binding global AI safety rules, not the voluntary pledges the industry has leaned on so far.

Smith's wish list is specific. He wants a legal requirement that companies notify people when a rogue AI agent breaches their systems. He wants independent, outside evaluators approving a model's safety before release rather than after. And he spoke favorably of "emergency brake" mechanisms -- the ability to pause or slow a system if humans lose the ability to control it. None of that is currently required by law anywhere in the world. The closest thing on paper is the agreement Smith himself calls a "good step": a voluntary, "morally binding" document that OpenAI, Anthropic, Nvidia, Google and Meta signed at a White House event last month, promising to work with independent auditors on safety assessment.

Why "good step" and not "enough": Smith's argument is that voluntary commitments only bind the companies that were already inclined to behave. "It's the companies that are not as responsible that you need to worry about," he told the BBC -- the same logic that has driven every call for mandatory disclosure rules in every regulated industry, applied here to a technology that currently has none.

That sequence is the context Smith's interview landed in, and it sharpens his argument considerably. Nine days before he called the UK "best placed" to lead, the country's own AI safety testers had been locked out of the models they'd need to test -- by the US government, not by the labs. The access was restored only after Narayan applied direct diplomatic pressure, and nothing in the reporting suggests the White House has reversed its underlying position that oversight bodies outside America shouldn't get a first look at American frontier models. Smith's pitch for UK leadership is, in effect, an argument that the institution best equipped to referee frontier AI is the one Washington just tried to sideline.

Smith's case for the UK specifically, rather than the US or the EU, rests on relationships as much as capability: "In Washington, D.C., people don't look at the U.K. and they don't look at London the way they look at Brussels," he told the BBC. "That's a strength." The UK's AISI, launched in 2023, already gets pre-deployment access to frontier models from the major US labs -- access that just survived its first real test of whether that arrangement outlasts a change in American political mood.

That last row is the one Smith keeps returning to. A pledge that only binds its signatories is, by construction, a pledge that can't touch the actor it's designed to catch. OpenAI, Anthropic, Nvidia, Google and Meta signing a safety document changes very little about what a sixth company, one with no reputation to protect and no seat at that table, is free to ship.

Smith closed the interview without naming an adversary. "Let's speak calmly, maybe even softly, but firmly, and start to bring people together," he said -- a notably conciliatory register for an executive publicly contradicting the administration whose AI policy his own company has to operate under. Dario Amodei and Sam Altman made similar slowdown arguments to Trump and adviser David Sacks last month and were dismissed; Smith's version arrives wrapped in a milk carton instead of an extinction warning, but it's the same disagreement, from inside the industry the administration says doesn't need the rule.