The EU's AI rules just became real: what you'll actually notice

The European Union passed its AI Act two years ago, but for most of that time it was a law with a waiting period — written down, not yet enforced. That changed on 2 August 2026, when the Commission's AI Office and national regulators started actually applying it. Here is what is different, in terms of what you will see.

First, honesty from chatbots. Any interactive AI system has to tell you that you are dealing with a machine rather than a person. No more wondering whether the support agent is real — if it is AI, it has to say so.

Second, labels on synthetic media. Images, videos and audio that were generated or edited with AI — deepfakes in the broad sense — must be marked. The Commission has even produced a standard set of icons for it.

Third, and more technically interesting: AI-generated content must also carry a hidden, machine-readable marker. The idea is that even if a visible label gets cropped out or screenshotted away, software can still recognise the content as synthetic. That is much harder to do reliably than it sounds, and it is the requirement companies are working hardest on.

Fines, and substantial ones: up to €15 million or 3% of a company's worldwide annual revenue, whichever is larger in practice, with smaller businesses treated proportionately. Enforcement is handled by regulators in each EU country, plus a central European AI Office for the largest general-purpose systems like the ones behind ChatGPT and Gemini.

This is a staged rollout, not a single switch. The next date is 2 December 2026, when a new list of banned uses takes effect. Rules for 'high-risk' AI — systems used in areas like employment, education, credit and essential services — follow in December 2027, and rules covering AI built into regulated products arrive in August 2028.

One of the December bans is worth knowing about because it is drawn unusually widely: so-called 'nudifier' apps that create non-consensual intimate images are prohibited, and so are tools marketed without reasonable safeguards against being used that way. That shifts responsibility onto how a general tool is sold, not just what it was designed for.

The bigger picture: Europe has taken a very different path from the United States, where federal policy has moved towards voluntary arrangements and away from binding rules. For now, if you want to know what enforceable AI regulation looks like, Europe is the only place with a full set of it — and the rest of the world is watching how the first cases go.

It is worth understanding why the invisible label matters more than the visible one. A visible badge on an AI-made image does its job until someone crops it, screenshots it or re-posts it somewhere else — which, on the internet, is roughly immediately. A machine-readable marker is meant to travel inside the file itself, so that a social network, a news organisation or a fact-checking tool can ask 'was this made by AI?' and get a reliable answer without relying on anyone's honesty.

The catch is that files get compressed, resized, re-encoded and screenshotted constantly, and each of those steps can strip or scramble hidden data. Making a marker that survives all of it, while being impossible to forge, is genuinely unsolved. The requirement is written as an obligation on providers to do it, which in practice means doing the best current technology allows and documenting the effort — not guaranteeing a perfect result.

For you as a reader, the practical takeaway is modest but real: over the next year, expect more tools and platforms to start showing 'this was AI-generated' notices automatically, rather than relying on whoever uploaded it to mention it. That is the point of the rule, and it is a reasonable thing to want.