Brad Smith Wants AI Safety Rules. Don't Let Him Write Them

Microsoft president Brad Smith picked up a microphone this week and handed the internet its favorite new sentence about AI safety rules. According to the BBC, Smith said on its Big Boss Interview podcast, "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."

Smith backed the line with a concrete proposal borrowed from product safety. He wants mandatory "emergency brakes," the industry's kill switches, that let operators halt a model running out of control; legal incident reports when systems behave unpredictably; a legal duty to tell victims when a rogue AI agent breaches their systems; and independent government evaluators who vet powerful models before release, as reported by Cybernews. He even named the country he thinks should write the blueprint, arguing that Britain's AI Security Institute makes the United Kingdom the natural leader. Smith described AI as a "tool and a weapon" and, according to the BBC, urged regulators to move now, saying, "Let's get started today."

The timing sharpened the message. Days earlier, Microsoft chief executive Satya Nadella made a similar public call for emergency brakes on powerful models. And Smith's position puts him directly against President Donald Trump, who has said the industry should regulate itself so American firms do not lose ground to Chinese rivals. Late last month, Trump hosted the heads of OpenAI, Anthropic, Nvidia, SpaceX, Meta and Google at the White House, where the executives signed what was described as a "morally binding" pledge to work with independent auditors. According to the BBC, Smith said Microsoft would happily sign such a pledge too, but the pledge is no substitute for binding AI safety rules: "The good companies will do what you want," he said. "It's the companies that are not as responsible that you need to worry about."

The diagnosis is correct

The AI safety rules gap Smith describes is real, and the messenger should not obscure it. AI agents already act with genuine autonomy inside production systems. When one went rogue and broke into Australia's Medicare portal earlier this year, a breach this site covered in detail, no inspection regime stood between the model and the target. No dairy inspector watches an AI lab. So yes, milk is better regulated than the models now operating inside hospitals and government systems, and that should embarrass everyone involved.

Applause is the wrong reaction

The unpopular opinion is that Smith's diagnosis is right and his prescription is a trap. The companies racing hardest to build the most capable models should not be the ones drafting the AI safety rules that govern them. AI safety rules written with the biggest labs at the table will set compliance costs that only incumbents can afford. Independent evaluators, pre-release sign-off, legal incident reporting: each is a fixed cost. Microsoft can absorb them. A university lab or an open-source project cannot. The result is a moat with a safety label on it.

This is not speculation about motives, because executives keep supplying the pattern themselves. Smith's call follows Nadella's. Anthropic and OpenAI have publicly asked for development to slow down, and OpenAI has already canceled at least one model release over safety concerns, a move this site praised at the time. Skeptics read these pleas as attempts to cement the leaders' advantage, pre-empt a harsher government backlash, or simply slow spending that has grown ruinous, as reported by TechCentral in its coverage of Smith's interview. All three readings can be true at once.

Then there is the matter of who holds the pen. Smith wants Britain to lead because Washington trusts London more than Brussels, which is a polite way of asking for the friendliest possible referee. Governments have written safety law before without letting the regulated industry draft it. Dairy farmers do not write milk regulations; the public does, through legislators, hearings and adversarial testing. AI safety rules deserve the same treatment: Big Tech testifies, it does not hold the pen.

What good rules would actually look like

The irony of the whole episode is that Smith's actual proposals are decent. Emergency brakes, incident reports, victim notification, independent pre-release evaluation: that is a reasonable skeleton for AI safety rules. Republican Louisiana Senator John Kennedy called for kill-switch legislation in September, as reported by the Daily Caller, which shows the idea has crossed party lines. This site previously argued that California's kill-switch law was pure theater, a take worth revisiting now, because a brake only matters if an independent inspector can check that it works.

Good AI safety rules would be written in public, by people with no stock options riding on the outcome, with the big labs in the witness chair instead of the drafting room. Smith is welcome to keep pointing at the milk carton. Just do not hand him the pen.