Let’s talk about the countdown nobody’s watching.
There’s a date running out right now that almost no one in the AI conversation has mentioned. December the 9th, 2026. Under a hundred days away. And on that day, in every EU member state, a law goes live that does something the AI Act — for all its noise — never really did.
It makes the machine a product you can hold accountable.
Picture the scene the law is built for. A man in his sixties, recovering from surgery at home, wearing a wearable that’s supposed to monitor his recovery and flag danger. The software was updated in the background one night. A model retrained. A threshold shifted. It stops flagging the thing it used to catch. Nobody notices until it’s too late.
Now ask the question that has haunted every AI story for five years: who do you sue?
The hospital says they just bought the thing. The maker says the software is licensed, not sold, it’s constantly learning, the terms are in the fine print, and anyway the logic inside is a trade secret. And right now, in most of the world, that’s where it ends. You get a shrug wrapped in a terms-of-service and a mountain of costs.
That’s what changes in December.
Because the revised EU Product Liability Directive does two quiet, brutal things. First, it rewrites what counts as a product to explicitly include software — including AI systems that keep learning after they’re sold. No more hiding behind “it’s not a toaster, it’s a service.” The machine is now a thing with a maker.
Second — and this is the part nobody’s talking about — it gives the courts teeth to order the maker to hand over the evidence. The training data. The model information. The logs. The thing the industry has always defended as a trade secret. When a harmed human asks a court to look under the hood, the court can compel disclosure, weighing secrecy against the right to a real case.
The AI Act can tell a company to write an impact assessment about itself. This law can tell a court to open the machine. Those are not the same thing.
And because the machine learns, the law had to get clever about blame. If an AI genuinely modifies a product after it ships — retrains, updates, drifts — and that change causes the harm, the original maker isn’t automatically off the hook. The burden shifts. Somebody has to own the state the machine was in when it hurt you.
Which means the era of the defect with no owner is ending. The quiet, learned, unlogged failure that just kind of happened — that now has a price and a name.
Now here’s the human consequence underneath the legal mechanics. For four years we’ve been told the machine’s reasoning is unknowable, and unknowable means unaccountable, and unaccountable means you eat the cost. That was never a law of physics. It was a business decision dressed up as a technical fact.
December takes the “unknowable” excuse and hands it to a judge.
For the builders listening — and I know some of you are — this isn’t an EU compliance footnote. If you ship software that touches a body, a bank account, a home, a life, your risk just went from “the user signed the waiver” to “the defect can be opened in court.” The right answer isn’t to lawyer up and obfuscate harder. The right answer is to build the way you’d have to if the logs could be read aloud in front of a stranger. Provenance. Audit trails. The receipts of every update, every retrain, every drift.
Not because a regulator asked. Because a person might.
So here’s what the conversation keeps missing. Everyone argues about whether AI should be regulated, whether we need more rules, whether the machines are safe. And the quieter, uglier question — the one nobody wants because it costs money — is who pays when the machine is wrong.
That question is about to have an answer with a deadline attached.
The AI Act decided the machine must behave. The Product Liability Directive decides the machine must answer. And answerability — real, evidenced, open-in-court answerability — is the thing that makes every other right real. A right to an explanation is a paragraph. A right to pull the trade secret open in front of a judge is leverage.
So watch December 9th. Not because a single judgment will change everything. But because it’s the first day the machine that hurt you has a maker who has to show up.
The mystery was never sacred. It was a shield. And shields, eventually, get opened.
We’re still human. Let’s keep the evidence on our side.