Your chatbot must say it is an AI: what Article 50 requires

What the provision says

Article 50(1) requires that AI systems intended to interact directly with natural persons be designed and developed so that the persons concerned are informed that they are interacting with an AI system — unless this is obvious from the point of view of a reasonably well-informed, observant and circumspect person, taking into account the circumstances and the context.

Two things stand out. First: the obligation is one of design, not of notice. Adding a line to your terms of use is not enough; the information must reach the person who is interacting, while they are interacting. Second: the "obvious" exception exists, but it is a narrow door.

The "it's obvious" exception — and why you cannot rely on it

Many cling to the exception: "everyone can tell it's a bot". The trouble is that the exception is assessed from the standpoint of a reasonably well-informed, observant person in the circumstances — not from the standpoint of whoever built the system.

And the technology pulls the other way: the better the assistant, the less obvious it is that it is a machine. A chatbot that answers well, with a name and a natural tone, is precisely the case where the exception fails. Investing to make your assistant more human narrows your own escape route.

Rule of thumb: if your assistant has a name, an avatar and a conversational tone, the "obvious" exception is not available to you. You need the disclosure.

The most common mistake: "you are talking to an AI"

It is the wording you see everywhere, and it is wrong in most real installations. Almost every serious system is hybrid: the AI answers, and a human takes over when needed. At that moment "you are talking to an AI" becomes false — and a false disclosure is worse than none, because it is a statement the company wrote and that is not true.

The correct wording for a hybrid system says the conversation may be handled by a virtual assistant or by an operator, and then distinguishes, message by message, who answered. It is the kind of detail you only learn by implementing it.

Doing it properly: four technical requirements

• Disclosure at the start of the interaction, visible without hunting for it — not in the footer, not in the privacy policy, not behind an icon.
• Per-message distinction between an AI reply and a human operator's reply. In a hybrid system it is the only way the initial disclosure stays true for the whole conversation.
• The system must not pass itself off as human. Asked "are you a person?", it must answer honestly. That takes a constraint in the system instructions, not a hope: commercial personas, left unconstrained, tend to say yes rather than break the illusion.
• The disclosure must not be switchable off by whoever installs the system. A "hide AI notice" toggle in the admin panel is a button that makes your own customer non-compliant.

What about AI-generated content?

That is paragraph 2 of the same article, and it covers systems generating synthetic text, images, audio or video: the content must be marked as artificially generated, in a machine-readable format. If you publish AI-generated product descriptions or images, it applies to you.

For systems already on the market on 2 August 2026, the deadline on this specific point moves to 2 December 2026. It is the only breathing room provided, and it does not extend the interaction-transparency obligation.

Who answers for it: provider or deployer

The provider designs the system so transparency is possible and on by default. The deployer — the business putting the chatbot in front of its customers — keeps it on and does not work around it. The right question to ask your vendor is not "are you compliant?" but three precise ones: is the disclosure on by default and impossible to disable? are AI replies distinguishable from human ones? what does the bot answer if I ask whether it is a person?

If the vendor cannot answer those three, the answer is no.