EU AI Act Article 50: what your chatbot must disclose — and how to build it in

Since 2 August 2026, the transparency rules of the EU AI Act apply. For anyone who runs a chatbot, an AI assistant or a content generator for people in the EU, Article 50 is the passage that matters most. It is short, and at its core it asks for one thing: people must not be misled about whether they are dealing with a machine.

This is an engineering view of the rule — what it requires, who it applies to, and how we build it into systems so that nobody has to remember it. It is not legal advice; for your specific situation, ask your counsel. The Commission’s guidelines on Article 50 and the voluntary code of practice on marking AI-generated content, both from summer 2026, go into the details.

First question: provider or deployer?

The AI Act distinguishes two roles. The provider develops an AI system, or has it developed, and puts it into service under its own name. The deployer uses an AI system under its authority. An association that commissions its own members’ assistant and runs it under its own name will usually be the provider. An organization that embeds a vendor’s chat widget unchanged is a deployer. Some duties sit with one role, some with the other, so this question comes first.

Duty 1: say that it is an AI

Article 50(1) requires AI systems that interact directly with people to be designed so that those people are informed they are interacting with an AI — unless that is obvious to a reasonably well-informed, observant person in the circumstances. Relying on “obvious” is risky. A friendly first name and a smiling avatar make an assistant look more human, not less.

• A clear notice in the chat window before the first question — not in a footer, not in the privacy policy.
• The same notice in every language the interface speaks.
• For voice assistants, a spoken notice at the start of the conversation.
• A name and an avatar that do not suggest a member of staff.

Duty 2: mark what the system generates

Article 50(2) addresses providers of systems that generate synthetic text, images, audio or video. Their outputs must be marked in a machine-readable format and be detectable as artificially generated, as far as this is technically feasible. Functions that merely assist with standard editing, or do not substantially change the input, are exempt. For a text assistant this typically means a marker in stored and exported answers; for images, embedded provenance metadata such as content credentials. Systems that were already on the market before 2 August 2026 have until 2 December 2026 for this marking duty — a transitional period added by the Digital Omnibus on AI; new systems must comply from the day they launch.

Duty 3: label deepfakes and public-interest texts

Article 50(4) is aimed at deployers. Whoever publishes AI-generated or manipulated images, audio or video that resemble real people, places or events must disclose it. AI-generated text published to inform the public on matters of public interest must be disclosed as well — unless it has gone through human review and a person or organization holds editorial responsibility. For a press release or newsletter drafted with AI, an editor who reviews and signs off is the cleaner path.

Article 50(3) adds a duty to inform people exposed to emotion recognition or biometric categorisation. For a members’ assistant the simplest answer is not to use either.

When and how: first contact, clear, accessible

Article 50(5) sets the form. The information has to be given clearly and distinguishably, at the latest at the first interaction or exposure, and it has to meet accessibility requirements. A notice that a screen reader cannot reach, or that only appears after the first answer, misses the point.

Not a replacement for data-protection notices

The AI notice comes on top of the information duties of the GDPR, not instead of them. Members still need to know which data the assistant processes, where it runs and how long questions are kept. Where the assistant runs entirely on the organization’s own server, that part of the explanation becomes refreshingly short.

How we build it in

Transparency that depends on someone remembering it will one day be forgotten. So we make it part of the system itself:

1. 1.The AI notice is a fixed component of the chat interface and is shown before the input field becomes active.
2. 2.It is translated through the same files as the rest of the interface — a language without the notice fails the build.
3. 3.Every answer carries its sources, so users can see where a statement comes from.
4. 4.Stored and exported answers carry a machine-readable “AI-generated” marker.
5. 5.The acceptance tests include questions such as “Am I talking to a person?” — and the assistant must answer truthfully.
6. 6.The setup is documented: model, data sources, guardrails and test results.

On the members’ assistant we built for the Verein für holistische Gesundheit, the AI notice and the cited sources were part of the design from the first day, and its acceptance catalogue is re-run after every change.

Checklist

1. 1.Settle for every AI system whether you are its provider or its deployer.
2. 2.Show an AI notice before the first interaction, in every language.
3. 3.Avoid names and avatars that suggest a human.
4. 4.Where you are the provider, mark generated content in a machine-readable way.
5. 5.Disclose AI images, audio and video that resemble real people or events.
6. 6.Put AI-drafted texts on matters of public interest through human editorial review.
7. 7.Make every notice reachable for screen readers.
8. 8.Test the disclosure like any other feature, and document the system.