Picture the aisle of your local supermarket. You pick up a bottle of milk. Somewhere behind that bottle sits a pasteurisation standard, a chain of cold-storage checks, a date stamp someone is legally accountable for, and an inspection regime that predates the internet. Now pull out your phone, open whichever AI assistant you installed last week, and ask it for medical advice, legal guidance, or a summary of a contract. Nothing behind that answer carries anything close to the same weight of oversight.
That is the comparison Microsoft President Brad Smith reached for this month when calling for urgent AI regulation, pointing out that milk and washing machines face more oversight than AI companies do. He also named the country he believes should take the lead: the United Kingdom, on the strength of its AI Security Institute, launched in 2023 under the former Sunak government as the AI Safety Institute.
His phrasing was blunt. “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.”
Why a tool reviewer cares about this
I spend my working hours doing something unglamorous: installing AI tools, running them against real tasks, and writing down what breaks. And the single most frustrating part of that job is not that tools fail. Tools fail all the time; that is normal software. The frustrating part is that there is almost no shared standard for what a failure even means.
When I test a washing machine, there are energy ratings, spin-cycle classifications, and safety certifications. When I test an AI coding assistant, I get a marketing page, a vendor-run benchmark, and a vibe. If I want to know how often it hallucinates a function that does not exist, I have to build the test myself. If I want to know what happens to the code I paste into it, I read a privacy policy written to be unreadable.
So when the president of Microsoft says the oversight gap between dairy and language models is absurd, my reaction is not cynicism. It is recognition. That gap is the exact reason honest tool reviews are so hard to write.
The self-interest question
Let us be adults about this. Microsoft is not a neutral party. A large incumbent calling for regulation is, among other things, a large incumbent that can afford compliance departments. Rules written at a certain altitude tend to be survivable for companies with legal teams the size of a small town and punishing for a four-person startup shipping something genuinely new.
That does not make Smith wrong. It means his argument should be judged on its substance rather than on who is making it. And the substance here is narrow and reasonably testable: does the UK actually have the technical capability to evaluate these systems? The AI Security Institute exists, it has been running since 2023, and its entire remit is model evaluation rather than press releases. That is a stronger claim to competence than most national AI strategies can make, which largely consist of funding announcements and summits.
What would actually help the people using these tools
Smith is talking about global rules. I am interested in something smaller and closer to the ground. If a regulator with real technical staff wanted to improve daily life for the people who buy and deploy these tools, the useful work looks less like grand principles and more like plumbing:
- Independent evaluation results that are not run by the vendor selling the model
- Plain-language disclosure of what happens to data you feed a tool, including whether it trains on your inputs
- A consistent way to describe model versions, so “we upgraded the backend” stops silently changing your product’s behaviour overnight
- Published failure rates for high-stakes categories, the way other industries publish defect rates
None of that is exotic. It is the kind of information that already exists for food, appliances, cars, and medicines. The reason it does not exist for AI is not technical difficulty. It is that nobody has required it yet.
My read
I am sceptical of regulation written by the regulated, and I am sceptical of any country declaring itself best at anything. But the milk line lands because it is true and because it reframes the debate away from science fiction. The argument is not about whether a model will become sentient. It is about whether a product sold to millions of people should come with verified claims.
Until something like that arrives, the testing burden stays where it has always been: on the people actually using these tools, and on reviewers willing to run them until they break. That work is not going anywhere soon. If a regulator with real technical capability shows up to share the load, I will take the help.
đź•’ Published: