The previous article closed on who already holds what will remain scarce. Right after comes the twin question: who writes the rules of the game.
AI Act.
And it arrives with a piece of fresh news that at the beach will become a single sentence, said with the air of someone who always knew: “they’ve postponed everything, Europe realised it was going too far.”
That’s not what happened. On 27 July the AI Omnibus came into force, pushing forward the obligations for high-risk systems: over a year’s extension, through to December 2027. On 2 August, however, the transparency obligations actually kick in: anyone interacting with an AI system must know it, and generated or manipulated content must be declared. And the reason for the delay is written in black and white: the technical standards and tools to demonstrate compliance are not yet in place.
They moved the date of the inspection. They didn’t say the lift can’t fall.
Nobody accuses fire safety regulations of hating architecture. They don’t ban skyscrapers — they ask where the exits are. And you don’t regulate a lift because it’s worse than the stairs; you regulate it because when it fails, it fails catastrophically.
The AI Act does the same thing with systems that touch people’s lives. It doesn’t ask whether the machine is better than a human. Often it is, on average. It asks what happens the day it fails with confidence — and who pays.
A recommendation engine that gets five per cent of films wrong is an annoyance. A CV filter that gets five per cent wrong silently closes careers, and nobody ever notices. A medical triage system that gets five per cent wrong kills someone. Same percentage, incomparable stakes. That’s why the law doesn’t regulate AI — it regulates the radius of action: the bigger and more irreversible the consequence, the more a human who answers for it must remain in the loop.
The AI Act isn’t improvisation — it’s written by someone who understood everything very well: that the machine produces the plausible, not the true; that bias is the centre of gravity of its data; that where it doesn’t know, it doesn’t stay silent but fills the gap. It isn’t bureaucratic whim. It’s the sensible legal consequence of the machine’s properties.
Then there’s the other half. Regulating first risks cementing the giants: compliance is a rounding error for whoever has a legal department in Mountain View, and a hammer blow for a six-person startup in Emilia. And the familiar dissonance remains: we have the scaffolding of the regulation — valid, absolutely valid — but before having a European model capable of competing. The delay, read carefully, says this too. The rules arrived before the tools to verify them.
December 2027 seems far away. It’s the time you have to ask yourself the question the law will eventually ask you.
Over twelve instalments I’ve taken apart one word at a time — to leave you with the only thing that really matters: the right question instead of the right word.
In September the sun umbrella closes. The words come to the office with you.