Your Chatbot Needs an AI Disclosure. Here's Exactly What to Write.

What Article 50 actually says

Article 50(1) of the EU AI Act is one sentence, and it's worth reading in full: providers must ensure that AI systems intended to directly interact with people are designed so that the person is "informed that they are interacting with an AI system, unless this is obvious from the circumstances and the context of use."

There's a lot packed into that sentence. "Informed" means told — not tricked, not obscured, not buried. "At the latest at the time of the first interaction" means before or during the first exchange, not after. And "unless this is obvious" is a narrow exception — if your chatbot has a human name, a human avatar, or talks like a person, it's not obvious.

This applies to deployers too, not just providers. If you're using someone else's AI (OpenAI, Anthropic, whatever) in your product, the disclosure obligation is yours.

Where to put the disclosure

The regulation doesn't specify exact placement, but the intent is clear: it has to be visible where the interaction happens. That means:

• In the chat interface itself. A banner or message at the start of the conversation. Not a link to a separate page. Not a footer nobody reads.
• Persistent, not dismissible. If a user can close the banner and forget, new users arriving later won't see it. Keep it visible or re-show it at the start of each session.
• On the widget or embed too. If your chatbot lives on a third-party site via a widget, the disclosure needs to be inside the widget. You can't rely on the host site's terms of service.

The wrong approach: putting "we use AI" in your privacy policy and calling it done. That's not "informing" the user in any meaningful sense. The Article specifically says at the time of interaction.

Copy you can steal

Here are actual disclosure strings you can drop into your product. Adjust the tone to match your brand, but keep the core message: this is AI, not a human.

Customer support chatbot

For a support bot that handles first-line queries:

Sales or onboarding assistant

For an AI that guides users through a product:

Internal AI tool (employee-facing)

For AI tools used by your own staff:

Content generation feature

For a feature that drafts emails, reports, or marketing copy:

In your settings or about page

A longer-form disclosure for your product's settings or transparency page:

What not to do

A few patterns I've seen that would probably not satisfy a regulator:

• "Powered by AI" in 8px font in the footer. Not visible, not at the time of interaction. The disclosure has to be where the user is actually engaging with the AI.
• A one-time popup that can be dismissed forever. New sessions should re-establish the disclosure. A user who last visited six months ago shouldn't have to remember that your chatbot is AI.
• Giving your AI a human name and photo with no AI label. If your bot is called "Sarah" and has a headshot, a reasonable person would think they're talking to Sarah. The "obvious from the circumstances" exception does not apply here.
• Only disclosing in your terms of service. The Article says "at the time of the first interaction." Nobody reads the terms of service at the time of a chat interaction.
• Disclosing only in English on a product used across the EU. If you serve French or German users, the disclosure should be in their language. The regulation doesn't say this explicitly, but existing EU consumer protection rules strongly imply it.

Content labeling is separate

Article 50 has two different requirements, and people mix them up:

• Interaction disclosure (Article 50.1): Tell people they're talking to AI. This is the chatbot rule.
• Content labeling (Article 50.2): AI-generated content must be labelled as such and made machine-detectable. This is the deepfake/synthetic media rule.

If your product does both — a chatbot that also generates images or documents — you need both. The chatbot needs an interaction disclosure, and the generated content needs metadata labeling (C2PA for images, structured metadata for text).

The interaction disclosure is the easier one. It's just copy in your UI. The content labeling is more technical — it requires machine-readable metadata, which means integrating with standards like C2PA or adding schema.org annotations. We wrote a separate guide on the multi-layered marking strategy if you need to implement that side of things.