Microsoft’s president says AI companies face less oversight than a carton of milk or a washing machine. The same president says one country already has the technical chops to write the rules for all of it. Those two statements sit about as comfortably together as a production deploy on a Friday afternoon, and yet both come from Brad Smith in the same breath.
Smith’s argument is that the UK is best placed to drive AI regulation, pointing to its AI Security Institute, launched in 2023 under Rishi Sunak’s 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.”
I build bots for a living. Nothing about that statement changes my Tuesday. But the gap it describes — between what gets measured and what ships — is something I run into every single week.
What the milk comparison actually means for builders
The washing machine line lands because appliance safety is boring, specific, and enforced. Someone decided what counts as a pass. Someone tests against it. Someone’s name is on the certificate. When I ship an agent that touches a customer database, there is no equivalent. There is me, my test suite, and whatever I decided “good enough” means last Thursday.
That is not a complaint about missing paperwork. It is an observation that the burden of defining safe behaviour currently sits entirely with the person writing the code. If you have ever written a prompt-injection filter by hand and wondered whether you caught everything, you already know what unregulated means in practice. It means you are the regulator, and you are also the one shipping on a deadline.
The things I end up testing myself
- Whether a tool-calling agent can be talked into calling the wrong tool with attacker-supplied arguments
- Whether retrieved documents can carry instructions the model treats as commands
- Whether a bot’s refusals hold up when the same request is rephrased five times
- Whether logs capture enough to reconstruct a bad decision after the fact
- Whether any of the above still passes after a model version bump
None of these have a shared pass mark. Every team invents its own. That duplication is expensive and it produces wildly uneven results across the industry.
Why a technical institute matters more than a policy document
Smith’s case for the UK rests on technical capability rather than legislative ambition, and that distinction is the useful part for anyone building systems. A body that can actually evaluate models produces something developers can use: methods, benchmarks, failure taxonomies. A body that only publishes principles produces a PDF you skim once and forget.
The difference shows up in architecture decisions. If an evaluation method exists and is public, I can wire it into CI. I can fail a build. I can show a client a number instead of a vibe. If all that exists is a statement of intent, I am back to hand-rolling my own red-team scripts and hoping my imagination covers the attack surface.
Smith is also expected to announce new steps the UK is taking on cyber defences, and to argue the country needs to build more of its own technology to have a real say in how AI develops. That second point cuts in an interesting direction. Regulatory influence tends to follow technical participation. A country that only consumes models has limited standing to specify how they should behave. The same is true of development teams — if you only call someone else’s API, your influence over its behaviour is a support ticket.
What I’d actually do with this
Treat any public evaluation work coming out of a safety institute as free engineering. Those organisations are doing the expensive, unglamorous job of figuring out how to measure model behaviour, and the output is usually open. Pull their methods into your own test use. Borrow their failure categories for your incident taxonomy. You will get better coverage than you would designing it alone, and you will be aligned with whatever rules eventually arrive rather than scrambling to retrofit.
Regulation will land unevenly and slowly. The measurement techniques are useful now, today, in the repo you have open. That is the part worth paying attention to while the politics sort themselves out.
🕒 Published: