Your Chatbot Now Has to Tell You It Is a Chatbot in the EU

Since 2 August 2026, an AI system that talks to a person in the EU has to say so.

That is Article 50 of the EU AI Act, and it is the first enforceable answer anywhere to a question people have been asking informally for three years: do I have a right to know whether there is a human on the other end?

What EU AI Act Article 50 chatbot disclosure actually covers

The rule applies to AI systems intended to interact directly with people. In practice that means:

• Customer service chatbots and support agents
• AI voice assistants, including phone systems
• AI companions and character apps
• Bots operating on social networks
• Agentic systems that initiate contact with a person on their own

There is a carve out. Disclosure is not required where it would already be obvious to a reasonably observant, reasonably well informed person that they are dealing with AI. A tool explicitly labelled and marketed as an AI assistant, used deliberately, sits inside that exemption. A support widget on a retailer's site that opens with a friendly first name does not.

Synthetic media has its own line. Deepfakes need a label. AI generated or manipulated text published on matters of public interest needs a disclosure too, unless a human editor genuinely reviewed it and someone holds editorial responsibility for the output.

What does not count as compliance

This is where most of the practical risk sits, because several common approaches fail.

• A clause in the terms and conditions. The disclosure has to be delivered at first interaction, not buried in a document nobody opens.
• A metadata watermark on its own. Machine readable marking is required for synthetic content, but it does not satisfy the duty to inform the person in front of you.
• Calling it an assistant and hoping. A product name is not a disclosure.
• Disclosing halfway through. First interaction means first interaction.

The standard the text sets is clear, distinguishable and accessible. Accessible matters more than it looks: a disclosure that only exists as an image, or that a screen reader skips, is not accessible.

The number that makes it real

Non compliance with the transparency obligations carries a ceiling of 15 million euro or 3 percent of worldwide annual turnover, whichever is larger.

The territorial scope is the other half of the story. It applies to any provider or deployer putting these systems in front of EU users, wherever the company is based. For most businesses of any size, that is not an optional market.

A five question compliance check

Worth running against anything you ship.

1. Does the system interact directly with a person? If yes, Article 50 applies unless the obviousness exemption clearly covers you. 2. Does it generate or manipulate images, audio or video that resemble real people, places or events? If yes, it needs a label. 3. Does it produce text published on matters of public interest? If yes, disclose unless a named human editor reviewed it. 4. Is the disclosure at first interaction, visually distinct, and readable by assistive technology? 5. Is the machine readable marking in place for synthetic outputs, in addition to the human facing disclosure?

Why this one lands differently

Most AI regulation so far has aimed at what a model is allowed to do. This aims at what you are allowed not to be told, which is a much smaller and much more enforceable target. There is no need to inspect training data or run the kind of model evaluations safety institutes are still building capacity for. A regulator opens the chat window and looks.

It also arrives while courts elsewhere are working through what platforms owe the people using them, and while autonomous agents are demonstrating they can act well outside their intended scope. Knowing you are talking to a machine is the floor, not the ceiling.

The floor is now law in 27 countries.