EU AI Act Transparency Obligations for Chatbots and AI Agents

By the UXreviewer.io team4 min read

Under Article 50 of the EU AI Act, chatbots and AI agents that interact directly with people must make clear that the user is dealing with an AI system, unless this is obvious from the context. Providers of generative AI must mark synthetic output in a machine-readable way, and deployers must disclose deepfakes and certain AI-generated public-interest text. These duties apply from 2 August 2026.

This article breaks down the obligations by role and content type and then turns them into interface patterns. Guidance and a code of practice on marking AI-generated content were still being developed as of late 2026, and the digital omnibus proposals may adjust details, so check the current status. This is not legal advice.

Who has to disclose what under Article 50?

Article 50 splits obligations between providers (who develop or place the system on the market) and deployers (who use it under their authority). Knowing your role determines your to-do list.

• Providers of interactive AI systems: design the system so people are informed they are interacting with AI, unless obvious to a reasonably well-informed person.
• Providers of generative AI: mark synthetic audio, image, video and text output in a machine-readable format and make it detectable as artificially generated, as far as technically feasible.
• Deployers of emotion recognition or biometric categorisation: inform the people exposed to it.
• Deployers creating deepfakes: disclose that the content is artificially generated or manipulated, with lighter rules for evidently artistic or satirical works.
• Deployers publishing AI-generated text on matters of public interest: disclose it, unless it has undergone human review and someone holds editorial responsibility.

When is it obvious that a user is talking to AI?

The exception for obvious cases is narrow in practice. A product called an AI assistant with a clearly robotic persona may qualify; a support chat with a human first name, a profile photo and casual typing indicators almost certainly does not. Agents that send emails or make phone calls on a user's behalf are especially sensitive, because the recipient never saw your interface.

As a rule of thumb, if any reasonable user could believe they are talking to a person, disclose. The cost is one sentence; the cost of a misled user is trust, and potentially a fine.

Remember that disclosure and handoff belong together. Users who learn they are talking to AI often want to know how to reach a person, so name the route right away.

How do you design a disclosure that users actually notice?

The Act requires the information to be clear, distinguishable and provided at the latest at the first interaction, and it must meet applicable accessibility requirements. That rules out the footer link and the terms of service as the only place.

Patterns that work well in reviews:

• A first message or header line such as 'You are chatting with an AI assistant. A human can take over at any time.'
• A persistent, compact AI badge next to the assistant name, not only on the first screen.
• Text that screen readers announce, not just an icon.
• A short 'What can this assistant do?' link that explains limits and data use.
• For voice agents, a spoken disclosure at the start of the call, before collecting any information.

What about agents that act outside your interface?

Agents increasingly send messages, fill forms or negotiate with third parties. Those recipients are also people interacting with an AI system. Practical measures include signatures in outgoing emails that state the message was drafted or sent by an AI agent on behalf of a named user, spoken disclosures in outbound calls, and identifiers in automated web interactions where relevant.

This connects to UX design for the owner too: users should know in advance how their agent introduces itself. Our article on agent transparency and explainability covers that side.

How should generated content be labelled?

Machine-readable marking is mostly a provider and engineering topic: watermarks, metadata and detection interfaces. If you build on a foundation model, check what the model provider already does and what you must add, for example preserving metadata when you resize or re-encode images.

Visible labels are a UX topic. Use consistent wording such as 'AI-generated' close to the content, make labels survive sharing and export where feasible, and avoid labels so small they function as decoration.

Disclosure checklist for chatbots and agents

Use this sequence as a release gate for any new AI touchpoint, not only as a one-time project before August 2026.

1. Identify every place where a person meets your AI: chat, voice, email, generated media.
2. Decide for each whether disclosure is needed and whether you act as provider or deployer.
3. Write disclosure copy in plain language and test it with users.
4. Check visibility on mobile, in dark mode and with screen readers; see our guide to WCAG 2.2 for AI products.
5. Verify machine-readable marking survives your processing pipeline.
6. Record screenshots and decisions as evidence, and re-check after every redesign.

Frequently asked questions