EU AI Act chatbot disclosure examples for Article 50

What Article 50 asks teams to achieve

Article 50(1) addresses AI systems designed to interact directly with natural persons. The system should be designed and developed so that the person is informed that they are interacting with AI, unless that fact is obvious to a reasonably well-informed, observant and circumspect person in the circumstances.

The implementation goal is therefore not a magic sentence. It is an understandable disclosure delivered at the relevant moment through an interface that people can perceive and revisit.

Short disclosure examples

These examples are implementation starting points, not legal opinions. Adapt them to the actual system, audience, capability and channel.

Customer support chatbot

You are chatting with Acme Assist, an AI system. It can answer questions about our products. A human support agent can review the conversation when you request help.

AI product assistant

This assistant uses AI to generate responses. Check important information before relying on it.

Voice assistant opening

Hello, I am Acme's AI voice assistant. I can help with account questions or connect you to a person.

Compact persistent label

AI assistant

The compact label works best as a persistent reinforcement. Do not assume that a small label alone is sufficient when the nature of the interaction could still be unclear.

Where to place the disclosure

Use a two-layer pattern:

1. Put a plain-language notice in the launcher, opening message or voice introduction so it is encountered before or at the start of the exchange.
2. Keep a persistent AI label or an easily available information panel during the interaction.

Avoid hiding the only disclosure in terms, a privacy policy, a tooltip that is never opened, or a footer outside the interaction. If the assistant starts the conversation automatically, the AI identity should be part of that opening experience.

Accessibility and localisation

The disclosure should be clear, distinguishable and accessible. Match the language of the surrounding product experience and test the component with keyboard navigation, screen readers, zoom and high-contrast settings. For voice experiences, provide the information audibly and consider a text equivalent where a visual interface exists.

Do not translate only the noun "AI" while leaving the rest of the notice in another language. Preserve the meaning, the system identity and any route to human assistance across every supported locale.

Evidence to retain

For each production surface, retain:

• the exact approved wording and locale;
• the component or prompt version;
• screenshots or recordings showing the first-interaction experience;
• the deployment URL, environment and timestamp;
• accessibility test results;
• the control owner and approver;
• a change record when wording, placement or assistant behaviour changes.

Article 50 Hub can turn these records into a versioned control and evidence workflow. Re-test after changing the model, interface, launcher behaviour, supported languages or human-handoff design.

Common failure modes

• Calling the feature an "assistant" without stating that it is AI.
• Showing the disclosure only after several messages.
• Relying on a privacy policy instead of the interaction itself.
• Updating the chatbot UI without updating the evidence record.
• Using wording that suggests a human is responding.
• Treating a disclosure as proof that every other AI Act duty is satisfied.

Run the free Article 50 readiness checker to determine whether interaction disclosure is the relevant control for the facts you provide.