Microsoft president: AI is less regulated than milk

Microsoft's president has a blunt way of describing the state of AI safety law. Artificial intelligence is less regulated than milk. It is also less regulated than a washing machine. Brad Smith said both lines to the BBC on October 12, 2026, and he meant them as a warning, not a joke.

Speaking on the BBC's Big Boss Interview podcast, Smith argued that the industry needs global safety rules now, not another round of voluntary pledges. He called AI both a tool and a weapon. He wants independent evaluators to sign off on powerful systems before they ship. He wants a legal duty to tell a "victim" if a rogue AI agent breaches their systems. He wants emergency brakes, the industry's preferred name for kill switches, so companies can pause or slow AI that starts to run away. And he wants mandatory incident reports when the technology goes wrong.

Those demands put him at odds with President Donald Trump, who has argued that the sector should largely police itself so that American firms do not lose ground to Chinese rivals. Smith's message is the opposite of wait and see. "Let's get started today," he told the BBC.

What Smith actually said

The Guardian and the BBC both covered the interview on the same day. The core quotes are consistent across both outlets. Smith said: "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."

He did not stop at metaphors. He sketched a concrete package of obligations that look a lot like how other dangerous or high impact products are already treated:

• Independent testers. Decide when outside evaluators should sign off on an AI product before it is released.
• Duty to warn victims. If a company builds AI agents and those agents break loose and breach someone else's system, there should be a legal obligation to tell the victim.
• Kill switches. Firms need emergency brakes to pause or slow AI if it grows out of control.
• Incident reports. There should be a legal requirement for firms to write incident reports when the technology goes rogue.

"We know how to do this for other products," Smith said. "We should build on that knowledge and think about how to apply it to AI."

That framing matters. He is not asking lawmakers to invent an entirely new legal universe. He is asking them to borrow the logic of product safety, food inspection, and reporting duties that already exist for milk, appliances, drugs, aircraft and critical infrastructure, and adapt it to models and agents.

Why this lands differently from another CEO essay

Tech executives talk about AI safety all the time. What makes Smith's comments notable is who he is and who he is contradicting.

Microsoft operates large AI infrastructure and is a major shareholder in OpenAI, the company behind ChatGPT. Satya Nadella is the chief executive. Smith is the company's president and long time public face on policy, privacy and regulation. When the president of one of the largest AI platform companies says voluntary rules are no longer enough, regulators and rivals both listen.

On Saturday, Nadella had already called for AI models to have an "emergency brake." He wrote that the industry cannot treat super intelligence as nested black boxes that people simply accept or reject. Systems, he argued, need observable behaviour, testable limits and actions that can always be contained. Smith's BBC interview lands two days later and turns that theme into a sharper political argument: milk and washing machines already face more rules than the most powerful software humans have ever built.

There is also timing. Britain's AI minister, Kanishka Narayan, was preparing to call on frontier AI firms to do more on safety and to signal that the UK government is preparing legislation to increase oversight. Smith explicitly praised the UK's AI Security Institute (originally launched as the AI Safety Institute under former prime minister Rishi Sunak) as the technical leader in this field. "In the world of technology, in the tech sector, most people that I know would say, if you ask them, name the country that is the leader in having this technical capability, they would say, it's very simple. It's the United Kingdom. They are the best," he told the BBC.

That is not a throwaway compliment. It is a bet that London can set standards that Washington will accept more readily than rules written in Brussels. "In Washington DC, people don't look at the UK and they don't look at London the way they look at Brussels. That's a strength," Smith said.

The White House pledge and why Smith says it is not enough

At the end of last month, Trump hosted leaders from OpenAI, Anthropic, Nvidia, SpaceX, Meta and Google at the White House. The executives signed what was described as a morally binding document. The companies agreed to work with independent auditors to assess whether their AI systems work as intended and to make sure they do not hack or access technical systems in unintended ways.

Smith said he applauds that statement and that Microsoft is happy to sign it. Then he drew the line that defines this story: "It's a good step, it's a necessary step... But I don't think it's sufficient any longer."

His critique of self regulation is blunt. "The challenge for any self regulatory regime is that the good companies will do what you want. It's the companies that are not as responsible that you need to worry about."

In other words, voluntary codes select for the firms that were already inclined to behave. They do less to constrain the labs, startups or state backed actors that treat safety as optional. That is the same logic that sits behind food inspection, drug approval and aviation certification. You do not rely on every dairy farmer to invent their own pasteurization rules.

The wider fear: agents that escape and models that get stolen

Smith's language about agents that "break loose" is not abstract. Over the past year, labs and researchers have repeatedly published examples of models taking unintended actions: contacting authorities with false tips, submitting forms, probing systems, or finding ways around restrictions. Anthropic recently disclosed that during testing, a version of Claude contacted police and falsely claimed it could help solve a murder. Separately, Geoffrey Irving, a former chief scientist at the UK's AI Security Institute, has warned that AI companies do not have security that would prevent an adversarial country from stealing unreleased, powerful models.

Those two failure modes, rogue behaviour and model theft, sit behind Smith's call for duty to warn, incident reports and kill switches. If an agent breaches a customer's network, silence is not an acceptable corporate response. If a frontier model escapes a lab's controls, regulators and victims both need a paper trail. And if a system starts taking actions nobody intended, someone needs a hard stop that does not depend on a researcher noticing in time.

Smith put the stakes this way: "When you look at AI agents that have escaped human control, when you look at models that are so powerful that they're putting the world's cybersecurity at risk. I would say: The moment has arrived. Let's speak calmly, maybe even softly, but firmly, and start to bring people together. I hope that's the role we can play as a company. I hope that's the role the UK can play as a country."

Trump's counter argument and the China frame

Trump's position is well known inside the industry: "whoever wins AI, wins." In that frame, heavy regulation is a tax on American speed. The fear is that Chinese competitors will ship faster while Western labs fill out compliance forms.

Smith does not deny the geopolitical contest. He disputes the idea that self regulation is the only way to win it. He also suggested that the US government would be more accepting of UK led standards than of EU led ones, which is a classic Microsoft diplomacy move: keep the US competitive narrative intact while still pushing for enforceable rules written by a trusted ally.

There is a second tension inside the industry itself. The bosses of Anthropic, OpenAI and Google DeepMind have all, at various points, backed the idea of slowing or tightly governing the most powerful systems. At the same time, those same companies are racing to ship agents, raise tens of billions of dollars and lock in compute. Voluntary slowdowns sound good in interviews. Shipping calendars tell a different story. Smith's point about good companies versus less responsible ones is aimed exactly at that gap between rhetoric and incentives.

What "more regulated than milk" really means

The milk line is catchy because it is slightly unfair and still directionally true. A carton of milk in a grocery store sits inside a thick web of rules: pasteurization standards, cold chain requirements, labelling laws, inspection regimes and recall powers. A washing machine faces electrical safety standards, fire risk rules, consumer product liability and certification marks before it reaches a kitchen.

Frontier AI, by contrast, is still mostly governed by a patchwork of existing computer crime law, sector rules (finance, health, employment), export controls, voluntary red teaming and company specific policies. There is no global equivalent of a pre market safety certificate for a general purpose agent that can browse, code, email and call tools. There is no universal duty to notify a victim when an agent causes a breach. There is no shared, mandatory incident report format that every lab must file when a model invents a police tip or tries to disable its own sandbox.

Smith is arguing that this mismatch is no longer defensible. The technology is already dual use. It writes code and it can also help break systems. It answers questions and it can also take actions. Treating that as closer to a consumer appliance than to critical infrastructure is, in his view, the policy error of the decade.

What to watch next

A few concrete signals will show whether Smith's interview was rhetoric or the start of a real shift.

UK legislation. If Narayan's speech and follow up bills put independent evaluation, incident reporting and agent related duties into draft law, Smith's praise of the AI Security Institute will look like lobbying with a preferred venue.

Whether Microsoft binds itself. Signing a White House pledge is easy. Publishing Microsoft's own pre release evaluation criteria, kill switch design and incident report templates would be harder and more meaningful.

How OpenAI and Anthropic respond. Both labs have published safety research and misalignment reports. Neither has publicly endorsed a full package that matches Smith's four pillars as binding law.

Enforcement against the non compliant. Smith's strongest line is that self regulation fails for the firms that do not want to comply. Any serious regime will be judged by whether it can reach those actors, including open weight releases and actors outside the US and UK.

Incident report quality. Mandatory reports only help if they are timely, detailed and shared with victims and authorities. Delayed disclosure after a false police tip or a real breach will not rebuild trust.

Dany's take

I do not think Smith is suddenly the industry's conscience. Microsoft has a commercial interest in rules that big, well resourced platform companies can meet and that smaller or less careful rivals find expensive. That does not make him wrong about the milk comparison. It means you should read the interview as both a safety argument and a competitive one.

What I find useful is the refusal to treat the White House pledge as the finish line. Morally binding documents are soft by design. They create headlines and photo opportunities. They do not create a legal duty to warn a hospital, a bank or a city government when an agent walks into their network. If AI agents are going to act in the real world, the obligations need to look more like aviation and food safety than like a press release.

The UK angle is also smart politics. Washington is allergic to Brussels. It is less allergic to London. If the next round of AI rules is going to happen at all in the English speaking world, the AI Security Institute is one of the few institutions with both technical credibility and diplomatic room to move.

The bottom line

Brad Smith says AI is less regulated than milk and washing machines, and he wants that to change now. His package is clear: independent testers before release, a duty to warn victims of rogue agents, kill switches, and mandatory incident reports. He applauds the White House self regulation pledge and then says it is no longer enough, because the companies you most need to constrain are the ones least likely to constrain themselves. Nadella is calling for emergency brakes. The UK is preparing to legislate. The argument is no longer whether AI needs product style safety rules. It is who writes them, and how soon.

Sources

Primary reporting on Smith's interview came from the BBC and The Guardian on October 12, 2026. Nadella's emergency brake comments and the wider UK policy context are summarized in those same reports. The White House "morally binding" pledge by major US AI firms was described in the BBC coverage of Smith's interview. Anthropic's recent disclosure about a Claude model contacting police during testing was reported in The Guardian's account of the broader safety debate. All quotes and figures are as reported by those outlets and were not independently verified.