Chatbots in Europe Now Have to Tell You They’re Chatbots

You know that customer service chat where you spend four minutes being pleasant to something before realizing it was never a person? In the European Union, that's now against the law. As of August 2nd, the transparency obligations in the EU's AI Act are live and being enforced, and the headline requirement is refreshingly blunt: if you're talking to a machine, the machine has to say so.

The rules cover three things. Chatbots and other interactive AI systems must disclose that a user is dealing with AI rather than a human. Deepfakes, meaning AI-generated or AI-edited images, video, and audio, must be labeled. And AI-generated or altered content has to carry machine-readable marks so platforms and researchers can detect it downstream.

The disclosure has to be somewhere you'd actually see it

This is the part that gives the rule teeth. Under the Commission's guidelines, burying the admission in your terms of service doesn't count. A metadata watermark alone doesn't count. Vaguely calling the thing an "assistant" doesn't count. The disclosure has to be perceivable inside the interaction itself, and it has to arrive at or before the first exchange.

There's a narrow carve-out for cases where the AI nature is already obvious, but it's tested against a reasonably informed person drawn from the actual audience, with a lower threshold when children, elderly users, or people with disabilities are part of that audience. In other words, you don't get to argue that everyone knows by now.

Deepfakes that are clearly artistic, satirical, or fictional get a lighter disclosure duty rather than a pass. AI-assisted text escapes the labeling requirement only where a human has actually reviewed it and someone holds editorial responsibility.

Who this reaches

Further than you'd think. The AI Act applies to anyone placing AI on the EU market or whose AI outputs are used inside the EU, so a US or UK company serving European users is in scope. Penalties run up to €15 million or 3 percent of worldwide annual turnover, whichever is larger.

The Commission paired the enforcement date with its first published list of signatories to the Code of Practice on transparency of AI-generated content. Roughly 190 organizations have signed, which suggests this is being operationalized rather than treated as a suggestion. Signing gives companies a defined path to demonstrate compliance. Everyone else has to prove their own approach is adequate.

Notably, the higher-risk categories (recruitment, credit scoring, law enforcement) got pushed back to December 2027 under a simplification package agreed in May. So the systems most likely to decide whether you get a job or a loan have more time, while the everyday consumer-facing tools are being regulated first. Reasonable people disagree about whether that's the right order.

What changes for you

If you're in Europe, the practical difference is visibility. The chatbot handling your airline complaint should identify itself. An AI-edited photo in your feed should carry a label. The EU has even published a set of icons companies can use.

Whether that actually reduces deception is the open question. Labels only work if people read them, and the internet has spent two decades training everyone to click past disclosure banners without looking. But there's a version of this that works quietly and well: the machine-readable marking, which lets platforms and journalists detect synthetic content at scale even when a human never notices a label. That's the part with real leverage, and it's the part nobody will ever see.

For everyone outside Europe, watch what companies do next. Building two products, one that discloses and one that doesn't, is expensive. Most will pick one behavior and ship it everywhere, which means Brussels may have just written the labeling rules for your feed too.