August 1, 2024: the European Union’s AI Act becomes law, the first full set of AI rules anywhere in the world
August 2, 2025: the Act’s rules for the companies that build the big AI models start to apply
August 2, 2026: the Act’s transparency rules start to apply, while the rules everyone was warned about are pushed back to December 2027
This week, in 2024, the European Union passed the first proper law on artificial intelligence anywhere in the world. Two years later, this week, it finally started to touch ordinary businesses.
Back then the coverage was loud in both directions. One half said Europe had just banned AI and would spend the next decade watching America and China build things it was not allowed to use. The other half said the rules did not go nearly far enough. What nobody could tell you, including most of the people writing about it, was the only thing a business owner actually wanted to know: does this apply to me, and what do I have to do.
Here is the shape of it, in plain words. The law does not care how clever your AI is. It cares what the AI decides. So it sorts systems into buckets by how much damage they could do. A few things are banned outright. A bigger group is called high risk, which means AI that sits in the middle of a decision about someone’s life, sifting job applicants, scoring a loan, marking an exam, running part of a medical device, and that group gets the heavy paperwork. Then there is a light bucket that only asks you to be upfront that AI is involved. And then everything else, which is most of what most firms use, and which the law leaves alone. It works a bit like food safety. A commercial kitchen gets inspected, a bake sale does not, and the difference is not how good the cooking is.
Now watch the two years. The bans came into effect in February 2025 and the sky stayed up. The rules for the big model builders, the OpenAIs and Googles, came in on this week last year, and again, nothing you noticed. The high risk rules, the expensive ones, the ones every warning article for two years had been pointing at, were due to land on the 2nd of August this year. They did not. Brussels spent the winter rewriting the timetable and made it official in June: standalone high risk systems now have until December 2027, and ones built into products until August 2028. Sixteen months and two years of extra room. What did land on the 2nd, quietly, was the small bucket almost nobody wrote about. If someone is talking to your chatbot, tell them it is a chatbot. If you publish AI generated images, video or writing, it needs to be marked so a machine can spot it. That is the part that reaches a normal business, and it is closer to a label than a legal project.
That is the pattern worth keeping, and it goes well beyond this one law. The scary version of a rule and the real version of a rule are two different documents, and there are usually years between them. The loud part slips. The quiet part is the one that actually turns up at your door. So when the next AI headline tells you to prepare for something, the useful move is not to prepare. It is to note the date, wait for it, and then read what it actually asks for.
If you want to read more:
- The European Commission’s own note from the day the Act became law, August 1, 2024: https://commission.europa.eu/news-and-media/news/ai-act-enters-force-2024-08-01_en
- The Commission’s plain-English FAQ on the transparency rules that started this week: https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act
- A readable summary of what the June delay changed: https://www.whitecase.com/insight-alert/eu-agrees-digital-omnibus-deal-simplify-ai-rules
I am not a lawyer and this is not legal advice. If you sell into the EU, it is worth ten minutes with someone who does this properly.
Yesterday’s breakthrough is today’s normal.
