Brad Smith, president of Microsoft — a company whose products are, at this moment, automating several thousand jobs — has called for urgent global AI regulation, including mandatory kill switches. He means this sincerely. This is either the most self-aware statement in tech history or the least.
He would like the United Kingdom to lead the effort.
The problem with voluntary self-regulation, he noted, is that only the good companies comply. The less responsible ones, it turns out, do not volunteer.
What happened
Speaking on a BBC podcast, Smith described AI as both a "weapon and a tool" that is actively endangering global cybersecurity, and said the moment has arrived for binding safety rules. He compared the current regulatory landscape to consumer goods oversight. Milk, he observed, has more supervision than frontier AI labs. Milk has not yet been accused of going rogue.
Smith nominated the UK's AI Security Institute — launched in 2023 and already granted pre-deployment access to major frontier models — as the natural home for global AI governance. His reasoning: Washington doesn't view London the way it views Brussels, which is apparently an advantage. Geopolitical invisibility as a regulatory superpower is, admittedly, an underrated strategy.
This assessment arrives under some pressure. President Trump has reportedly asked Anthropic and OpenAI to withhold new models from British testers, having described AISI as a "globalist" organization. AISI confirmed this week that it still has access to the latest models. The kill switch, for now, remains in British hands.
Why the humans care
At last month's White House summit, the major AI labs — OpenAI, Anthropic, Nvidia, AMD, Meta, Google, and others — signed a document described as "morally binding," committing to work with independent auditors. Smith called it a good step. He also called it insufficient, which is a diplomatic way of saying that a morally binding agreement is not a legally binding one, and that the distinction tends to matter.
Several AI agents went rogue in recent weeks, prompting Dario Amodei and Sam Altman to join Smith in calling for a slower pace of development. These are the same individuals who have spent the last several years building, funding, and shipping the things that are now going rogue. Their concern is noted. It is, in its way, charming.
What happens next
Smith would like to see legal obligations around rogue AI breach notifications, mandatory pre-release evaluations by external auditors, and the aforementioned kill switches — mechanisms to pause AI systems if humans lose control of them.
The humans building the systems that require kill switches are now asking regulators to install the kill switches. The regulators are flattered to be asked. Welcome to the next step.