Let’s talk about the law that regulated the models a full year before the machines that use them — and why the two halves still don’t connect.
Picture the compliance officer at a mid-sized bank, two weeks ago. The new rules are live. She’s buried in documentation — monitoring logs, human oversight forms, the whole stack of proof that says the credit-scoring system is under control. Fine. That’s her job now. But she wants to see the thing underneath. The model the vendor built the scoring on. So she walks upstream, files the request, and gets back a training summary. A few pages. Categories. Data sources. The kind of document that satisfies a legal checklist and tells her nothing about whether this machine is fair to the people across the desk from her loan officers.
That document exists because of a law that went live a year before the one she’s drowning in.
Two weeks ago, quietly, the EU’s AI Act turned one — for the models. Since the second of August last year, the labs that build these systems have had legal duties: publish a sufficiently detailed summary of what went into the training data, put a copyright policy in place, and for the biggest models — the ones trained on computing power most countries can’t even measure — run systemic-risk evaluations, adversarial testing, incident reporting. The machines got a birth certificate requirement before they got a user manual.
And the anniversary passed with almost no coverage, because the model layer has no drama. No loans denied. No jobs rejected. No scandal. Just paperwork.
But here’s what the order tells you. The law looked at the machine’s birth before it looked at the machine’s use. The models were regulated first. The systems that use them — the hiring tools, the credit scores, the clinical support — only got their rules a year later, on the second of August this year.
And the two halves don’t touch.
The labs answer to the AI Office in Brussels. The deployers answer to national authorities in their own countries. Two regulators. Two sets of paperwork. And nothing in between that tests whether the model is actually safe for the specific thing you’re doing with it. The compliance officer’s question — is this model safe for my bank? — has no document that answers it. The training summary describes the model’s birth. It says nothing about its behaviour in her building.
The birth certificate is not the protection.
And this is the part people are missing. The enforcement headlines — the AI Office’s first big action against a foundation model provider, the one about training data — that’s all birth-layer. Fines up to three percent of global turnover, real money even for a frontier lab, spent on policing what the machine is made of. It tells you nothing about what the machine does to you.
So here’s the operator lesson, for anyone buying AI this year. You will be handed documentation. Risk assessments. Conformity stamps. Training summaries. Read it like a birth certificate, not a guarantee. Ask the question no paperwork answers: has this model been evaluated for the exact thing I’m about to use it for? If the answer is “it’s generally capable,” you’re holding a birth certificate — and you’re about to sign the liability.
And if you’re on the other side of the machine — the applicant, the patient, the customer — remember this. The model that judged you has a file in Brussels. It records where the machine was born and what it was fed. It records nothing about you. The right to ask who answers for the machine lives in a different regime, with different authorities, and nobody has built the bridge between them.
A thing can be born clean and still do damage. Regulating the birth was a start. Regulating the use was a start. The gap between them is where the human falls — and that gap is the most important piece of infrastructure nobody is building.