Since 2 August 2026, a chunk of the European Union's AI Act has moved from "written down" to "actively enforced." The part that took effect is the transparency layer — and it changes what any company putting AI in front of European users is allowed to leave unsaid.

Here is what is now required.

You have to be told when you're talking to a machine

Under Article 50 of the Act, an AI system that interacts directly with people — a chatbot, a voice agent, a support bot — must make clear that it is AI, unless that is already obvious. No more support chats where you can't tell whether "Alex" is a person or a script.

AI-generated content has to be labelled

Three related rules cover synthetic media:

• Synthetic content — images, audio, video or text made by AI — must carry machine-readable markings so it can be detected downstream, and providers have to offer a way to detect it.
• Deepfakes, and AI-generated content on matters of public interest, must be openly disclosed as artificial, unless the material has been through substantive human editorial review.
• Emotion-recognition and biometric-categorisation systems must inform the people they are used on.

There is one bit of breathing room: generative systems already on the market get until 2 December 2026 to add the machine-readable marking. The rest applies now.

The teeth

This is not guidance. The EU's AI Office enforces the rules on general-purpose AI models — it can demand technical documentation, evaluate a model, require changes, and fine non-compliance. Penalties for breaching the transparency obligations run to €15 million or 3% of worldwide annual turnover, whichever is higher.

There is a softer path running alongside the stick: companies that sign the voluntary Code of Practice on AI-generated content get a presumption of conformity and a gentler enforcement posture. Those that don't can expect closer scrutiny.

Why it matters beyond Europe

Two reasons.

First, the "Brussels effect": global companies tend to build to the strictest major market and ship that version everywhere, because maintaining a separate, more secretive product just for the EU is rarely worth the trouble. A disclosure rule written in Brussels has a way of becoming a disclosure rule in a lot of other places too.

Second, it is one of the most consequential regulatory answers yet — China's synthetic-media labelling rules got there earlier — to a question the whole industry has been dancing around: does a person have the right to know when the thing they are reading, watching or talking to was made by a machine? The EU has now answered yes, in a binding law, and attached a number to getting it wrong.

The Act phases in over several years, and its heaviest obligations — the rules for "high-risk" systems like hiring, credit scoring and biometric identification — run on their own, later timeline: a July 2026 "Digital Omnibus" revision pushed the main high-risk deadline out to December 2027, while leaving these transparency rules exactly where they were. But the transparency principle is now live and enforceable: if it's AI, and it isn't obvious, you have to say so.