AI regulation: Brad Smith seeks global rules and kill switches

A Big Tech executive asks for binding rules

Microsoft President Brad Smith has made one of the bluntest calls for AI regulation yet heard from inside a frontier AI company. In an interview with the BBC, he called for global safety rules and argued that the technology currently faces fewer safety requirements than a carton of milk or a washing machine.26 He set the comparison out plainly: food and household appliances already go through safety rules and inspection, and he said governments should use that experience as a template for AI rather than start from scratch.19

Smith's proposals are specific. He wants companies legally required to file incident reports when AI systems go wrong. He wants victims told when a rogue AI agent breaches their systems. He wants an "emergency brake" that can pause or slow a system, and rules for when independent evaluators must review new technology before it is released.74 He described AI as both a tool and a weapon, pointed to agents that have escaped human control and models powerful enough to threaten global cybersecurity, and said the time to act has arrived.10

Outlets disagree slightly on the timing. Mashable says Smith spoke on the BBC's Big Boss Interview podcast on October 11.7 Most other outlets describe the interview as published on Monday, October 12.29 The substance is reported the same way everywhere.

Why voluntary pledges are the real target

The most important part of Smith's argument is about enforcement, not milk. Last month, President Donald Trump hosted executives from OpenAI, Anthropic, Nvidia, Meta, Google and others at the White House. They signed a document described as "morally binding" that commits them to work with independent auditors.10 One account notes that this "Joint Commitment on Frontier Responsibilities" cannot be enforced in court and carries no penalties.1

Smith did not reject that pact. He said Microsoft was happy to sign it and called it a good and necessary step, but said it is no longer sufficient.7 His reasoning is simple: responsible companies will do what regulators want, and the worry is the companies that are less responsible.6 That is a direct argument against relying on self-regulation, and several outlets describe it as putting Smith at odds with a White House that has favored letting industry police itself, partly so the US stays ahead of China.107

CEO Satya Nadella has made a similar case. In a post on X, he said companies should assume a model is compromised and contain it from the start, that any authorized person should be able to pause or shut down a model mid-task, and that more advanced models will need standardized containment technology.1 One report reads Nadella's comments as pushing back on Trump's preference for self-regulation.1 Another notes that Microsoft's policy chief and its CEO now appear to be making the same regulatory request.5

This did not come out of nowhere. Fortune reported in July 2026 that Smith criticized the US approach as regulation without transparent or complete rules. He pointed to a June 2026 executive order that set up a voluntary pre-release review for frontier models but no licensing requirement.5 The same report describes the Commerce Department using export-control powers, not AI-specific law, to make Anthropic withdraw two models over a cybersecurity concern and to press OpenAI to limit a rollout.5 Those interventions support Smith's case. The government is already intervening, but one case at a time, using tools that were not designed for AI.

London, not Brussels

The most strategically interesting part of the interview is where Smith thinks global rules should be written. He named Britain as best placed to lead, citing the AI Security Institute set up under former Prime Minister Rishi Sunak.6 Under voluntary agreements, the institute gets access to major developers' models before release so it can test their capabilities and risks.9

Smith said Washington simply sees London differently from Brussels, and called that a strength.10 Several outlets connect that distinction to the EU's AI Act, which regulates AI by risk level.29 One report frames his position as the UK, rather than the US or the EU, being best placed to lead, because Washington doesn't treat London with the suspicion it reserves for Brussels.8

This is a political calculation, not just a preference. Any global framework the current US administration sees as a Brussels product is unlikely to get American support, so Smith is proposing a venue that might. But the UK option has its own problems. Trump reportedly asked Anthropic and OpenAI last month to withhold new models from British testers, although the institute said on Monday that it can still access and test the latest systems.8 So the UK model Smith is praising depends on the same kind of voluntary access he says is no longer enough.

The EU AI Act weakens the milk comparison

For compliance teams, Smith's milk line needs context. It describes the US well. It describes Europe much less well. The AI Act has been in force since August 1, 2024. It sorts systems into four risk tiers, applies to any organization whose AI output reaches people in the EU, and allows fines of up to €35 million or 7% of global turnover for prohibited practices.12

Europe has not fully followed through, however. The Digital Omnibus on AI, Regulation (EU) 2026/1744, came into force on July 27, 2026, six days before the high-risk rules were due to take effect.18 It moved obligations for stand-alone high-risk systems, such as tools used in hiring, credit scoring, education and law enforcement, from August 2, 2026 to December 2, 2027. It moved obligations for AI built into regulated products to August 2, 2028.1519 The stated reason was that national authorities and harmonized technical standards were not ready.19 Parliament co-rapporteur Arba Kokalari described the change as pressing the pause button on the AI Act and cutting red tape.17

The delay was narrower than it first looked. Rules for general-purpose AI models have applied since August 2025. Article 50 transparency duties, which cover telling people they are talking to a chatbot and labeling deepfakes, took effect on August 2, 2026 as planned, and enforcement powers started the same day.1817 A new ban on AI systems that generate non-consensual intimate imagery and child sexual abuse material takes effect on December 2, 2026.1216 Law firm Gibson Dunn has called the omnibus a deferral, not a dismantling, because the law's structure and core obligations stay in place.13 Civil-society groups have still criticized the delay, arguing it leaves people facing major automated decisions without the impact assessments and human oversight the law promised.20

Taken together, Smith's "global" pitch is mostly aimed at Washington. Europe already has binding obligations and a regulator that can fine general-purpose AI providers directly.12 The United States has an executive order, a voluntary pact and occasional use of export controls.51

What compliance teams should watch

There are a few practical lessons. First, the controls Smith describes, including incident reporting, human ability to halt a system and independent pre-release evaluation, are becoming standard expectations on both sides of the Atlantic. Firms that build them now will be ready for whichever version of these rules arrives. Second, one EU compliance analysis warns of a specific risk with AI agents: a deployer that uses an agent in ways the provider did not intend may legally become the provider of a high-risk system.19 That is the same agent problem Smith says needs a kill switch.

Third, the December 2027 deadline is closer than it looks. One analysis notes that working with a notified body and preparing the required technical documentation typically takes 12 to 18 months, and notified bodies are still being designated.15

Our view: genuine, and also good for Microsoft

It would be naive to treat Smith's intervention as pure public service. Skeptics question whether executives calling for regulation want safety or want to lock in their position and head off tougher action.10 Fox News notes the broader question of whether stricter rules protect the public or give established companies an advantage.4 One analysis argues that binding global rules would raise compliance costs most for smaller labs and would probably speed up consolidation among the few firms big enough to absorb those costs.5

Both things are true. The incentives are real, but the argument also holds up. When an industry's main safety mechanism is a pledge that carries no penalties, the companies most likely to break it are the ones that never meant to keep it. Smith's real contribution is to say that openly, as a signatory of that pledge. What follows depends on whether the UK can host a framework that Washington accepts and Brussels recognizes, and nothing in the current politics suggests that will happen soon.