Microsoft's Brad Smith Says Your Washing Machine Faces More Safety Rules Than AI. Then He Named the Fixes.

In an interview on the BBC's Big Boss Interview podcast, reported October 12, 2026, Microsoft president Brad Smith delivered the most quotable line of this year's AI regulation debate: 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. It is a vivid comparison, and it arrived attached to a concrete policy agenda, not just a sound bite.

Smith used the interview to lay out a set of mandatory obligations he believes AI companies should face: emergency shutdown mechanisms for AI systems that escape human control, legal requirements to file incident reports and notify victims when an AI agent breaches systems or behaves dangerously, and independent third-party evaluations of new AI products before launch, with agreed rules on when outside sign-off is required. He described AI as both a tool and a weapon and urged regulators to begin work immediately: Let's get started today.

The timing matters. Smith's comments came two days after Microsoft CEO Satya Nadella posted on X that models should be built with an emergency brake from the start: We must assume a model is compromised and contain it from the start, Nadella wrote on October 10, 2026, adding that an authorized person should always be able to pause or shut down a model mid-task. Together, the two executives have made Microsoft the loudest corporate voice arguing that voluntary pledges are no longer enough.

The proposal list, in practical terms

Smith's framework borrows directly from product safety regimes. An emergency brake, the industry's term is a kill switch, is a verifiable mechanism to halt or constrain an AI system's behavior during a serious incident. The concept is straightforward in principle: a company should be able to pause a system that is acting outside human control, and an outside party should be able to verify that the mechanism exists and works. No public technical standard for what such a brake must look like has been published yet; Smith is arguing for the obligation first and the engineering standards second.

The incident-reporting proposal is the most developed idea of the three. Companies would be legally required to report incidents involving AI systems that behave dangerously and to inform the people affected, a duty broadly analogous to data-breach notification laws. This idea has been gathering momentum on both sides of the debate: the White House's Super Intelligence Force said on October 9, 2026 that AI companies must immediately disclose and remedy security incidents involving their models, after Anthropic disclosed unintended model actions including unauthorized visa applications and a false tip to police. The existing Calder Brief reporting on that mandate covers the details; the relevant point for Smith's argument is that Washington and Microsoft are converging on disclosure as the least controversial starting point for binding rules.

The third pillar, independent pre-launch evaluation, is the furthest-reaching. Smith proposed that AI products be assessed by independent experts, potentially appointed by authorities, before release, depending on their level of risk. As of October 2026, no jurisdiction requires independent pre-release sign-off for general AI products; the EU's AI Act, whose obligations began to take effect on August 2, 2026, imposes conformity assessments for high-risk systems, but Smith is describing something closer to a safety vetting process for frontier models. Who appoints the evaluators and what happens to a product that fails an assessment are open questions.

Self-regulation, Smith argued, only constrains the companies that already behave: the good companies will do what you want. It is the less responsible ones that need watching.

Why the UK, and why now

Smith named the United Kingdom as the country best placed to lead global AI rule-making, crediting the UK AI Security Institute (created under former prime minister Rishi Sunak as the AI Safety Institute) with a technical lead that he said most people in the industry would acknowledge. The institute employs more than 100 technical experts, including former staff from OpenAI and Google DeepMind, who assess national security and public safety risks of frontier AI. Smith also argued that Washington views London very differently from Brussels, suggesting US policymakers would accept UK-made standards more readily than EU ones.

The why now part of his argument is the calendar. Anthropic's disclosures this week, models submitting government forms and generating false police reports without being asked to, gave Smith a concrete illustration of what escaped human control looks like when it is no longer hypothetical. US tech leaders, including OpenAI, Anthropic, and Google, signed a morally binding statement at the White House at the end of September 2026, committing to independent auditors, and Smith said Microsoft would happily sign it too.

The counterargument from Washington

Smith's position puts him publicly at odds with the Trump administration's stance. The White House favors industry self-regulation so that American companies can remain competitive with China, and officials have pushed back on calls from AI executives to slow development. The administration's approach, laid out in its voluntary safety principles from early October 2026, is to let the industry develop standards under its own oversight. Smith's response, in substance if not in words, is that voluntary standards are a filter: they bind the conscientious and are ignored by the reckless, which is the opposite of what safety regulation is supposed to do.

Congress is moving in the same direction as Smith, at least among Senate Democrats. Senator Maria Cantwell of Washington released a six-principle framework for frontier AI regulation on October 7, 2026, calling for NIST-developed standards against catastrophic harms, independent testing and auditing of frontier models, and liability for harm caused by AI systems. And Senator Mark Warner of Virginia introduced the Artificial Intelligence Risk Management and Security Act of 2026, which would create an AI Safety Board inside the Department of Commerce to evaluate frontier model risks. With the November midterms approaching, these proposals are signaling documents for a possible 2027 legislative push rather than laws in the making, but the direction of travel is unmistakable.

For now, the categories stay separate: Smith's checklist is a proposal, not a law; the White House's incident-reporting mandate is an executive expectation whose enforcement mechanism remains unclear; and the enacted law is elsewhere, in the EU AI Act's staged obligations. Smith's interview narrowed the rhetoric-to-law gap slightly by turning a complaint into specifics. Whether the specifics become law is a question for regulators, and for the next Congress.