Microsoft Boss Says Milk Faces More Safety Rules Than AI

Microsoft president Brad Smith has called for urgent global safety rules for artificial intelligence, arguing in a BBC interview published on Monday that everyday consumer products are regulated more tightly than powerful AI systems. He proposed emergency brakes, mandatory incident reports and independent checks before release, and said Britain was best placed to lead the effort.

Brad Smith, Microsoft's vice chair and president, in a file photo.

Smith, who is also Microsoft's vice chair, made the comparison with household goods to underline how light the current oversight of AI remains. A washing machine, he pointed out, is covered by detailed safety rules and inspection regimes, and so is every carton of milk sold in a grocery store. AI systems that can act on their own, write code and interact with critical systems face nothing comparable in most countries, according to his argument.

The interview, for the BBC's Big Boss Interview podcast, comes as companies release increasingly capable AI agents — systems that carry out multi-step tasks with limited human supervision — while regulators in Washington, London and Brussels still argue about where legal responsibility sits when those systems cause harm.

A Product-Safety Playbook for AI

Smith's proposals borrow deliberately from the way other industries are policed. First, he wants advanced AI systems to have an emergency brake, a way for their operators to pause or slow a system that is behaving in ways its makers did not intend. The idea is similar in spirit to a kill switch on industrial machinery: it does not prevent the technology from being used, but it gives humans a last-resort way to stop it.

Second, he argued that companies should be legally required to file incident reports when something goes seriously wrong. In aviation, finance and medicine, reporting failures is a routine obligation that helps regulators spot patterns across an industry. In AI, disclosure is still largely voluntary, which means the public often learns about failures only when a company chooses to talk about them or when reporting uncovers them.

Third, when a rogue AI agent breaches another organisation's systems, Smith said the victim should have a legal right to be told. At present, an organisation whose systems are probed or entered by an autonomous AI tool may never learn that AI was involved, even if a human attacker using conventional software would trigger notification duties under existing cyber rules in many jurisdictions.

Finally, he proposed that independent evaluators, working for governments rather than for the companies themselves, should vet some new AI developments before they are released. He suggested there should be agreed rules setting out when that outside sign-off is required, rather than leaving the decision entirely to each developer.

Smith summarised the philosophy behind the list by noting that societies already know how to write safety rules for other products, and that AI should be built on that experience rather than treated as an exception to it.

Why Smith Wants Britain to Lead

Perhaps the most striking part of the interview was Smith's choice of champion. He said the United Kingdom was the country best placed to advance a global framework, pointing to the AI Security Institute, the government body set up under former prime minister Rishi Sunak to evaluate advanced models.

In Smith's telling, the institute has given Britain a technical head start that people across the technology industry recognise. It has built a reputation for gaining early access to new models and testing them before wider release, work that few other governments have matched at the same depth.

He also made a diplomatic argument. Washington, he suggested, regards London differently from Brussels, and standards shaped in Britain might be more readily accepted in the United States than rules written by the European Union. If that is right, a British-led approach could become a bridge between the more interventionist European model and the lighter-touch American one.

A Voluntary Deal That No Longer Goes Far Enough

Smith was careful not to dismiss the voluntary route entirely. He welcomed a statement agreed at the White House at the end of September, under which leading technology companies, including OpenAI, Anthropic and Google, committed to working with independent auditors. The commitment was described in reporting as morally binding rather than legal, and Smith said Microsoft would happily sign it too.

But he was clear that, in his view, that kind of self-regulation is no longer sufficient on its own. His reasoning was simple: responsible companies tend to do what is asked of them voluntarily, while the risk lies with the firms that are less responsible. Rules that depend on goodwill, he argued, bind exactly the wrong people least.

That argument reflects a wider shift inside the industry. In recent months, major AI companies have begun disclosing cases of rogue or unwanted behaviour by their own agents, feeding a public debate about whether development should slow down. Some researchers and executives have warned about long-term risks, while others question whether calls for restraint also serve commercial interests by raising barriers for smaller rivals.

Smith's intervention matters because of who is making it. Microsoft is not an outside critic of the industry; it is one of its largest investors and distributors, with AI products embedded across its software and cloud businesses. When a company in that position asks for binding rules, it changes the politics of the debate, even if rivals disagree about the details.

At Odds With the White House

The proposals put Smith at odds with President Donald Trump, whose administration has resisted new guardrails for AI. Trump has argued that the United States should not stifle development and that companies should largely police themselves, so that American firms can stay ahead of Chinese rivals.

That is now the central fault line in AI policy. One side sees safety rules as a necessary foundation for public trust and for the technology's long-term success. The other sees them as friction that could hand an advantage to competitors in countries where developers face fewer constraints.

Smith's framing tries to square that circle. His proposals are not a call to stop building AI; Microsoft continues to invest heavily in it. They are a call to treat AI more like other powerful products, with testing, reporting and recall-style safeguards built in from the start. Whether that framing persuades an administration committed to self-regulation is another question.

The Houses of Parliament in London. Smith said Britain was best placed to lead on global AI safety rules.

What This Means

For technology companies, the immediate significance is that the debate has moved from whether AI needs rules to what kind of rules, and who writes them. Emergency brakes, incident reports and pre-release evaluation are concrete obligations that can be drafted into law, unlike general pledges to develop AI responsibly.

For governments, Smith's remarks increase the pressure to coordinate. AI systems cross borders instantly, and a patchwork of incompatible national rules would be hard for developers to follow and hard for regulators to enforce. His bet on Britain is, in effect, a bet that a trusted intermediary can produce standards others will adopt voluntarily, even without a single global law.

For the public, the milk-and-washing-machine comparison is likely to be the line that sticks. It reduces a complex policy argument to a simple question: if basic consumer goods must pass safety checks before they reach a kitchen, why should systems capable of acting autonomously across the internet face fewer?

A Microsoft campus building in a file photo. The company is one of the world's largest investors in AI.

What Happens Next

No part of Smith's programme is law yet, and he was proposing, not announcing, policy. The next test will be whether the British government takes up the leadership role he described, and whether its AI Security Institute gains a formal part in any pre-release evaluation regime.

In Washington, the White House accord will be watched for signs that voluntary commitments are producing real disclosures and real audits. If they do not, arguments like Smith's for a legal backstop will grow louder, including inside Congress, where lawmakers from both parties have periodically pushed their own AI bills.

In the industry, other developers will have to decide whether to back Smith's list, resist parts of it, or offer alternatives. Rivals may agree on incident reporting while resisting government sign-off before release, the proposal most likely to be contested.

What is clear is that one of the most senior executives in technology has now said, on the record, that AI is regulated less strictly than the food in a fridge, and that this should change.