AI transparency in ChatLab and the EU AI Act

Since 2 August 2026, the transparency obligations in Article 50 of the EU AI Act apply. If you run a chatbot on your website for visitors in the European Union, this affects you. This article explains what the rules ask for, what ChatLab does about it out of the box, and what stays on your side.

What Article 50 actually requires

Article 50 is about honesty, not about restricting what your chatbot may do. It has two parts that matter for a website chat widget.

The first part says that an AI system meant to interact directly with people must be designed so that those people know they are talking to an AI, unless that is already obvious. The information has to reach the person no later than their first interaction, and it has to be clear and easy to notice.

The second part says that outputs generated by AI should be marked in a machine-readable format, so that automated systems can also tell that content was produced by AI. This obligation comes with an explicit qualifier: it applies as far as it is technically feasible, taking into account the state of the art and available technical standards.

There is also a rule about AI-generated text published to inform the public on matters of public interest. A customer service chatbot answering questions in a chat window is not that, so it does not apply to a typical ChatLab deployment.

Who is responsible for what

ChatLab is the provider of the AI system. The duty to design the product so that disclosure happens sits with us.

You, as the business running the chatbot on your website, are the deployer. You decide the bot name, the tagline, the welcome message and your privacy policy, so the disclosure that visitors actually see is shaped by your configuration.

In practice this works well, because our defaults already carry the disclosure and you only need to avoid removing it.

What ChatLab does by default

A persistent AI label in the chat header. Every bot is created with a tagline that sits directly under the bot name and by default reads "AI support assistant". It stays visible for the whole conversation, not only in the first message. It is set in the language of the interface used when the bot was created, and every language version we ship contains the letters AI.

The label follows the visitor's language on multilingual bots. If your bot has our multilingual feature enabled, the tagline is translated along with the rest of the interface, so a visitor writing in Spanish sees the disclosure in Spanish.

The widget footer names the product as an AI chatbot. Where our branding is shown, the footer link describes the widget as an AI assistant or AI chatbot.

Machine-readable marking in the page. Every message rendered by the widget carries a data-ai-generated attribute set to true or false, plus a data-message-origin attribute that says whether the text came from the AI, from a human agent or from a system notice. Messages generated by AI additionally carry data-ai-provider. The widget container itself is marked with data-ai-system. Any automated reader of the page can therefore tell, per message, what was machine-generated and what was not.

Human replies are never marked as AI. When a conversation is escalated and one of your agents takes over in live chat, those messages are explicitly marked as not AI-generated. The same applies to configured static text such as welcome messages, thank-you messages and error notices. We consider a false AI marking to be worse than no marking, so the distinction is made at the point where each message is created, not guessed from its appearance.

An honest note on watermarking

For images, audio and video there are emerging technical standards for embedding provenance signals. For plain conversational text there is no established, interoperable watermarking method today, and any signal embedded in the words themselves disappears the moment someone copies and pastes the reply. This is exactly the situation the regulation anticipates when it limits the marking duty to what is technically feasible and reflected in recognised standards.

Our approach is therefore to mark at the level where marking is reliable: the page markup that delivers the message. As standards for text provenance mature, we will follow them.

What we recommend you do

Keep the tagline in place. It is the single most effective thing, because it is visible from the first moment and throughout the conversation.

Consider putting AI in the bot name as well, for example "Virtual Assistant AI" instead of "Virtual Assistant". The name sits directly above the tagline, so the two reinforce each other in the same field of view.

Say it once in the welcome message. Something like "Hi, I am the AI assistant of [your company]" costs nothing and removes any doubt.

Add a sentence to your privacy policy describing that the site uses an AI chatbot and what happens with conversation data. If you are already reviewing your privacy documentation, this is the natural place for it.

If you integrate ChatLab through our API rather than the widget, you are building your own interface, and the disclosure in that interface is yours to implement.

Where to change these settings

The tagline and the welcome message live in your bot settings under Appearance and Conversation. The bot name is in the general settings. All of them are editable at any time and take effect immediately.

If you have questions about your specific setup, or you need a written summary of the technical measures for your compliance documentation, contact us and we will prepare it.