I was standing in a supermarket aisle when a worker slapped a safety sticker on a gallon of milk. You trusted that sticker the way you trust a red light at an intersection. It hit me: that carton of milk is under stricter rules than the software shaping our world.
I’ll be blunt with you: this isn’t academic. I follow the briefings, the memos, the interviews—and I want you to know what’s at stake before the next headline hits. I’ve worked sources who watch this industry closely, and I’ll walk you through what Brad Smith’s warning really means for you.
I watched a grocery worker scan milk and apply a safety label. Brad Smith used that image to make a point about mismatched rules.
In a recent BBC interview Brad Smith, president of Microsoft, argued that AI lacks the kind of baked-in oversight we accept for everyday goods: mandatory incident reporting, victim notification when systems are breached, independent pre-release reviews, and emergency brakes for runaway systems. He compared the situation bluntly: washing machines and milk cartons face stricter checks than many AI systems now do.
Smith’s list of fixes reads like product safety playbook: require reporting when AI goes wrong, mandate third-party evaluations for new models, and build kill switches so humans can pause or stop a model mid-task. Satya Nadella echoed the emergency-brake idea in a post on X, arguing humans should be able to pause a running model.
Are washing machines more regulated than AI?
Smith’s claim is literal and political. Consumer appliances and food products are governed by well-established federal and state rules, inspection regimes, and liability norms. AI mostly lives in a patchwork: voluntary pledges from industry players like Google, Meta, OpenAI, Anthropic, xAI, and Nvidia, plus sparse regulatory guidance. That gap is what he wants lawmakers to close.
At a White House meeting, CEOs signed a voluntary safety agreement — I read the Reuters report the same day. But voluntary pacts have limits.
On September 29, executives at a White House session signed a nonbinding safety compact described by President Trump as “morally binding.” It’s a public-relations shelf for now: no enforcement, no penalties. Axios sources say some insiders privately expect a major AI incident within six to 12 months, and that companies are already rehearsing damage control scenarios.
Meanwhile, OpenAI CEO Sam Altman told Politico the world should accept “some bad things happening” as trade-offs for AI benefits—an argument that sets expectations about who bears risk and how much tolerance society will accept.
Should AI have “kill switches”?
Both Smith and Nadella say yes. Practically, that means design requirements for pause-and-stop controls, logging, and oversight that are legally enforceable. Companies can build such safeguards—the aerospace and automotive industries regularly use verification and emergency-stop systems—and Smith argues we’ve already done similar work for other technologies.
I called a colleague in London who follows the UK’s AI Security Institute. Smith thinks the U.K. could become a model for global rules.
Smith praised the U.K. AI Security Institute and suggested Washington might prefer working with London rather than Brussels. That’s tactical: harmonized Anglo-American standards could pressure other nations and private firms to adopt common practices, while avoiding the slower, more prescriptive EU path.
Companies are trying to square two things: the need for speed to innovate, and the legal and reputational fallout if a system behaves badly. If regulators make reporting mandatory, firms no longer get to decide whether an incident is “public-facing.”
Who is responsible when AI causes harm?
Responsibility will be split: developers, deployers, and sometimes customers. Courts and regulators will parse contracts and design choices. Smith is arguing that law should make some responsibilities non-negotiable—incident reporting, victim notice, and independent audits—rather than leaving everything to internal compliance memos.
I watched a washing machine sit unused on a showroom floor with a child-lock sticker visible. Product safety can be mundane and strict — why not software?
Smith’s point is practical: regulators already write requirements for consumer safety, and the mechanisms exist to extend similar rules to software. That doesn’t mean copying appliance rules verbatim; it means adapting principles like testable safety standards, transparent reporting, and a public record of incidents.
AI today is a sleeping volcano—mostly quiet but capable of sudden, destructive pressure. Regulation should act like a seatbelt for software on public roads, limiting harm without banning travel.
Companies will resist mandatory reporting and pre-release audits; governments will fear stifling innovation; voters will worry about jobs, national security, and global competitiveness. But the debate is no longer theoretical. Brad Smith, Satya Nadella, Sam Altman, and the major platform players have publicly staked out their positions. The question now is whether you want those answers decided by voluntary agreements or by law—who do you trust to pull the brake when the code crosses the line?