EU AI Act Chatbot Requirements: What Applies Now
Article 50 has applied since August 2, 2026. Chatbots and AI phone assistants must disclose the AI interaction from the start, clearly and accessibly. Final Commission guidance now makes the practical standard more concrete.
The current rule: if an AI system directly interacts with a person, the provider must design it so the person is informed that it is AI. The disclosure must appear from the start of the first interaction, be clear and distinguishable, and meet accessibility requirements. Do not bury it in terms and conditions or rely on a vague “assistant” label.
What changed since this article was first published
The August 2 application date has passed and the European Commission has published final Article 50 guidance. This update replaces the earlier draft-guidance wording, explains the narrow “obvious” exception and separates the limited transition for machine-readable AI-content marking from the disclosure duty for directly interactive systems.
What Article 50 requires now
Article 50 covers AI systems intended to interact directly with natural persons, including conversational chatbots, virtual assistants and AI agents. The Commission's final guidance translates the obligation into four practical criteria:
- Direct: the notice belongs in the interaction itself, where the user can perceive it.
- Immediate: it must be given at the start of the first interaction, not after several messages.
- Clear and distinguishable: users should not have to infer AI involvement from a brand name or generic “assistant” label.
- Accessible: the presentation must account for the relevant accessibility requirements and channel.
For web chat, use a persistent visual notice in or immediately beside the conversation. For an AI phone assistant, make the disclosure audible in the opening. If the caller interrupts the greeting, the system should repeat it in the next response rather than assuming it was heard.
The “obvious” exception is narrow
The Act does not require a notice when it is obvious to a reasonably well-informed, observant and circumspect person that they are interacting with AI. The Commission interprets this exception narrowly and in context. A robot avatar, synthetic-sounding voice or product name is not a dependable compliance strategy because users, devices and accessibility needs differ.
Explicit language is easier to understand, test and document: “You are chatting with an AI assistant” or “You are speaking with our AI phone assistant.” Repeat the disclosure when the interaction changes channel or when the user could reasonably believe a human has taken over.
No general grace period for chatbot disclosure
The Commission's FAQ describes a limited transition for certain providers' machine-readable marking of AI-generated or manipulated content when systems were placed on the market before August 2, 2026. That is a different Article 50 obligation. It is not a general extension for telling people that they are interacting with a chatbot or AI agent.
If an interactive AI system is live today, the disclosure workflow should already be active. Treat a missing notice as a production issue, not as backlog for a later redesign.
Provider and deployer responsibilities
The design obligation in Article 50(1) falls on the provider of the interactive AI system. The business deploying the chatbot still controls the customer-facing context, content, channels and configuration. In practice, both sides need evidence: the provider needs a disclosure mechanism that works by design, while the deployer must not hide, disable or contradict it in the published experience.
Document who owns the greeting, widget label, accessibility review, language versions and human-handover state. Buying software does not answer these operational questions by itself. Transparency violations can fall within the AI Act's fine category of up to €15 million or 3% of worldwide annual turnover, subject to the Act's rules and the circumstances of the case.
Six checks to run today
- Inventory direct AI interactions. Include website chat, embedded assistants, messaging channels and AI phone numbers.
- Test the first contact. Open every channel as a new visitor and verify that the AI notice appears before or with the first substantive response.
- Use explicit words. State “AI” rather than relying on “digital”, “virtual” or “assistant”.
- Check every language and accessible view. Test keyboard navigation, screen-reader wording, contrast and audible phone disclosure.
- Keep human states truthful. When a human takes over, label that state clearly and retain the conversation context.
- Save evidence. Keep dated screenshots, call-test notes, approved wording and the deployed configuration version.
How WebChatAgent supports the disclosure workflow
WebChatAgent identifies AI replies inside the chat experience and supports a distinct human takeover state. The AI phone assistant can introduce itself audibly and route callers to a person. Your team should still verify the published wording, languages and channel configuration against the actual use case and keep the evidence with its compliance documentation.
This article provides general information, not legal advice. High-risk use cases, sector rules, employee contexts and personal-data processing may create additional obligations beyond Article 50.
Make AI disclosure part of the product
Keep the AI notice visible from the first interaction, maintain a truthful human-takeover state and test the exact customer experience in every published language.