EU AI Act 2026: Chatbot Disclosure Duties Explained

From 2 August 2026, Article 50 of the AI Act applies: users must be able to tell they are chatting with an AI, not a human. What it means for your website chat.

In plain language

The EU AI Regulation requires transparency. People should know when they are talking to an AI. For a website chat, this means the notice belongs visibly in the chat window. The privacy policy is not enough. The rule asks for the notice, but a handover to a person is not a must. In practice, the two belong together. Write down how you put the rules into practice. Breaches can lead to high fines.

July 8, 2026 11 min read By XICBOT

What Applies From 2 August 2026

The AI Act does not take effect all at once but in stages. After it entered into force on 1 August 2024, the bans on certain AI practices applied first, followed by obligations for general-purpose AI models. The transparency obligations under Article 50, which cover chatbots and AI-generated content, become applicable on 2 August 2026 (European Commission). By that date, a website chat should be set up to meet the requirements. For operators this means the clock is ticking, and the effort is manageable if you address it in good time instead of retrofitting just before the deadline.

Article 50 distinguishes between providers and deployers. Providers are those who develop and supply the AI system; they must design the chatbot technically so that disclosure is possible. Deployers are the companies that use the assistant on their website; they are responsible for ensuring the disclosure is actually visible in the specific use. In the practice of a website chat, both roles often coincide or interlock closely. The consequence matters: neither the provider nor the deployer can point to the other. Both sides should clarify from the outset who fulfils which obligation.

Article 50: the Four Transparency Duties at a Glance

Article 50 bundles several transparency rules that look different at first glance but share a common goal: people should know when they are dealing with AI. For a website chat, the first duty is decisive; the others concern AI-generated content and are also relevant depending on the use.

Chatbot disclosure (Art. 50(1))

Users must be informed that they are interacting with an AI, and in good time before or at the very beginning of the conversation.
Marking AI content (Art. 50(2))

Text, images, audio or video generated by generative AI must be marked in a machine-readable format as artificially generated.
Emotion and biometric systems (Art. 50(3))

Anyone using emotion recognition or biometric categorisation must inform the individuals exposed to it.
Disclosing deepfakes (Art. 50(4))

Artificially generated or manipulated image, audio and video content, so-called deepfakes, must be identified as such.

For chatbot disclosure under paragraph 1, the deadline of 2 August 2026 applies without delay. For the machine-readable marking of AI-generated content under paragraph 2, the AI Act, as amended by the so-called Digital Omnibus on AI, provides a short transition period until 2 December 2026 (Art. 111 AI Act) for systems that were already placed on the market before that date. A website assistant that mainly answers questions and guides visitors to the right offer usually falls under chatbot disclosure; if it also produces visible AI content, the marking duty is added.

Why Users Must Know They Are Writing With an AI

The idea behind disclosure is simple: those who know an AI is answering at the other end can classify the answers correctly and decide consciously which questions to ask and which details to share. This is no theoretical concern, because AI assistants are becoming everyday tools. According to Bitkom, 44 percent (Bitkom) of AI users turn to the technology for private questions and 34 percent (Bitkom) even for health-related matters. At the same time, Gartner forecast in 2022 that by 2026 one in ten contact-centre agent interactions would be automated with AI, up from an estimated 1.6 percent (Gartner) at the time. The more natural AI chat becomes, the more important it is that no one is left unclear about the nature of their counterpart.

People should be informed that they are interacting with an AI so that they can decide with knowledge of the circumstances.

In line with the recitals on Article 50 of the AI Act

For companies, disclosure is more than a duty. A visible, honest notice builds trust and removes the uncertainty from the interaction. Visitors who know they are talking to an assistant often phrase things more clearly and are less disappointed when a question is passed to a human. Transparency and a good chat impression are therefore not a contradiction. How an individually trained assistant differs from a generic standard chatbot is shown in the article on the difference between a custom assistant and a standard chatbot.

What This Means Concretely for Your Website Chat

Article 50 and the accompanying European Commission guidelines yield some concrete requirements for a website chat. They sound technical, but they are manageable to implement if you consider them from the outset rather than afterwards.

The notice appears in good time, that is, before or at the beginning of the conversation, not only after several messages (European Commission).
The disclosure is clear and distinguishable, not hidden in a tiny footer text or buried in the terms and conditions (European Commission).
The notice is designed to be accessible and thus perceivable by people with impairments too.
The assistant remains recognisable as an AI throughout and does not pretend to be human.
For sensitive or legally delicate matters, a handover to a human is provided.
The deployer can demonstrate that the disclosure is actually visible in use.

The European Commission guidelines also state explicitly what does not suffice: a tiny snippet of text in the footer, a faint, barely legible label or a notice buried in the terms and conditions. Such token labels do not meet the requirement for clear and distinguishable information. For a website chat this means the notice belongs visibly at the start of the chat window, in plain language and with sufficient contrast. A transparently labelled assistant should also answer reliably instead of speculating freely; how to avoid false answers through a well-maintained knowledge base is shown in the article avoiding hallucinations with a knowledge base.

In practice, disclosure can be implemented with a few but deliberate decisions. What matters is that the notice appears where the visitor is already looking, namely in the chat window itself, and before the first question is answered.

Notice before first contact

As soon as the chat opens, a clear sentence appears stating that an AI assistant is answering here, not a human.
Visible, not hidden

The disclosure sits prominently in the widget and not in the footer or terms and conditions, as the guidelines require.
Accessible and legible

Sufficient contrast, plain language and a screen-reader-friendly implementation make the notice perceivable for everyone.

Technically the disclosure is no major intervention, but it wants to be designed deliberately. The notice should appear in the visitor's language, which a multilingual assistant handles automatically, and it should meet the same accessibility standards as the rest of the website. How the integration embeds the assistant cleanly into an existing page without disturbing the layout is described in the relevant article. And how to make an assistant fully accessible in line with accessibility legislation is shown in the article on the accessible AI assistant under the BFSG.

Disclosure belongs at the start, not the end

Check your chat from the perspective of a first-time visitor: is it immediately clear on opening that an AI assistant is answering? A notice you only find after scrolling or in a linked document does not serve the purpose. The best place is the first visible line in the chat window, supplemented by a permanently recognisable AI label.

Enabling Handover to a Human

Article 50 requires disclosure, not necessarily a human handover. In practice, however, the two belong together. An assistant that openly appears as an AI should also offer a way to switch to a human as soon as a matter becomes too sensitive, too complex or legally delicate. This is no contradiction to automation but its sensible frame: the assistant handles the frequent, clearly answerable questions, and a human is available for everything else. A Gartner forecast suggests that this division of labour can pay off: in 2022 the analyst firm expected conversational AI in contact centres to reduce agent labour costs by 80 billion dollars (Gartner) in 2026. Human contact does not disappear in the process; it concentrates on the matters that need it.

A good handover passes along the previous conversation so the human does not start from scratch. That keeps the matter flowing even though the channel changes. Which questions the assistant can reliably handle itself and where a handover makes sense can be read over time from real conversations; how that works is shown in the article on analysing chat transcripts. How an assistant relieves support overall without replacing the human is described in the article on the support assistant.

Documentation for Market Surveillance

Transparency obligations are only worth as much as they can be demonstrated when in doubt. The AI Act requires member states to designate at least one market surveillance authority by 2 August 2025 (AI Act). In Germany, the Bundesnetzagentur (Bundesnetzagentur) takes on this role, with a coordination and competence centre for the AI Act being set up that supports other competent authorities and is intended to ensure a uniform interpretation. The German Bundestag adopted the necessary national supervisory structure in the AI Act implementation law (German Bundestag).

For operators of a website chat this means it is worth recording your own implementation in a traceable way. This includes how and where the AI disclosure is shown, when it appears, how the handover to a human is arranged and which system sits behind the assistant. This documentation need not be a thick manual, but in the event of a market surveillance enquiry it should be able to show that the assistant meets the requirements of Article 50. Those who have the assistant operated by a service partner should clarify that the partner provides the necessary evidence and technical details.

Fines: What a Breach Costs

The AI Act provides for tiered fines in Article 99. A breach of the transparency obligations under Article 50 can be penalised with up to 15 million euros (Art. 99 AI Act) or three percent of global annual turnover, whichever is higher. For breaches of the prohibited AI practices under Article 5, the ceiling is even 35 million euros (Art. 99 AI Act) or seven percent of turnover, and incorrect or misleading information to authorities can be penalised with up to 7.5 million euros (Art. 99 AI Act) or one percent.

Breach	Ceiling	Classification
Prohibited AI practices (Art. 5)	35m euros or 7 % of turnover	Highest tier
Breach of transparency (Art. 50)	15m euros or 3 % of turnover	Covers website chats
Incorrect information to authorities	7.5m euros or 1 % of turnover	Duties to provide information

The regulation emphasises that fines should be effective, proportionate and dissuasive and take into account the situation of small and medium-sized enterprises; for them the lower of the two values applies. Whether and at what level a fine is imposed depends on the individual case, and the ceilings are theoretical maxima, not standard penalties. The real point is different: the requirement itself is achievable with manageable effort. It is much easier to set up a chat transparently from the start than to have to react later to a complaint or an inspection.

Standard Widget or Transparent by Default

Many off-the-shelf chat widgets can be fitted with a disclosure notice after the fact. The difference lies in whether transparency is a detail screwed on afterwards or considered from the outset. An assistant that already brings the AI disclosure, the handover to a human and the necessary documentation saves retrofitting and makes the implementation traceable for market surveillance.

Aspect	Retrofitted standard widget	XICBOT assistant
AI disclosure	Optional, often in the footer	Visible before the first contact
Handover to humans	Not always provided	A fixed component with context
Accessibility	Depends on the provider	Legible, high-contrast, accessible
Documentation	To be compiled yourself	Traceable details on the use
Hosting	Often outside the EU	Hosting in Germany

XICBOT assistants are labelled as AI by default, with a clean handover to a human and hosting in Germany. That way the disclosure duty is not a makeshift solution but part of a considered setup. How data protection and the server location interplay is shown in the article on data protection and hosting; the technical foundation of an AI assistant is explained in the article what is an AI chat assistant. Which package fits is shown in the pricing overview. We do not promise absolute legal certainty, but a concept that takes the transparency obligations seriously from the start.

Your Checklist for the Disclosure Duty

From the points so far, a manageable list of questions emerges that should be clarified before 2 August 2026. It does not replace individual legal advice but offers reliable guidance on whether a website chat is set up transparently.

Do visitors recognise before or at the beginning of the conversation that they are writing with an AI?
Is the notice visible in the widget and not in the footer or terms and conditions?
Is the disclosure accessible, that is, legible, high-contrast and screen-reader-friendly?
Does the assistant offer a handover to a human at any time?
Is it documented how and since when the disclosure is shown?
Does the notice appear in the visitor's language?
Is it clear who is the provider and who is the deployer and who fulfils which obligation?
Does the assistant produce AI content that additionally has to be marked in a machine-readable way?

2 August 2026 is no reason for nervousness, but a good occasion to look at your own website chat from the perspective of the transparency duty. Those who answer the few points on this list cleanly operate their assistant on a solid basis. Which setup suits your sector is best clarified in a conversation; an overview of typical fields of use is given on the page on industry solutions, and specific questions are answered by direct contact.

Frequent questions

When does the disclosure duty for AI chats apply?

The transparency obligations under Article 50 of the AI Act become applicable from 2 August 2026 (European Commission). The regulation itself already entered into force on 1 August 2024, but its rules apply in stages. For a website chat this means it should be recognisable by the deadline that an AI assistant is answering. The precise understanding of the deadlines may still sharpen with further guidelines, which is why this article offers orientation and does not replace legal advice.

Where exactly must the AI notice appear in the chat?

Under Article 50 and the European Commission guidelines, the notice must appear in good time, that is, before or at the beginning of the conversation, and be clear and distinguishable. A tiny footer text or a notice buried in the terms and conditions is generally not considered sufficient. In practice, the disclosure belongs visibly at the start of the chat window, in plain language and with sufficient contrast. How strictly individual wordings are interpreted may depend on the individual case.

Must an AI chatbot be able to hand over to a human?

Article 50 primarily requires disclosure, not necessarily a human handover. In practice, however, the two belong closely together, because an assistant that openly appears as an AI should offer a route to a human for sensitive or legally delicate matters. A visible option to request a human at any time strengthens trust and makes the boundary of automation transparent. Whether and how a handover is needed depends on the sector and the use case.

What does a breach of Article 50 cost?

Under Article 99 of the AI Act, a breach of the transparency obligations under Article 50 can be penalised with up to 15 million euros (Art. 99 AI Act) or three percent of global annual turnover, whichever is higher. For small and medium-sized enterprises the lower value applies. These amounts are ceilings and not standard penalties; whether and at what level a fine is imposed depends on the individual case. A transparent implementation from the start is the simpler path.

Who enforces compliance in Germany?

In Germany, the Bundesnetzagentur (Bundesnetzagentur) is designated as the central market surveillance authority for the AI Act, supplemented by a coordination and competence centre that supports other competent authorities. The AI Act requires member states to designate at least one such authority by 2 August 2025; the German Bundestag adopted the national structure in the implementation law. Operators should document their implementation so that it can be demonstrated in a traceable way in the event of an enquiry.

Is a standard chatbot with a small AI notice enough?

That depends on the specific implementation. A barely visible notice in the footer tends not to meet the requirement for clear and distinguishable information. A more sensible option is an assistant that brings the AI disclosure, the handover to a human and the necessary documentation from the outset. XICBOT assistants are labelled as AI by default and set up with hosting in Germany. This does not promise absolute legal certainty, but it does provide a setup that takes the transparency obligations seriously.

Ask it yourself, Assistant of this website

Automated assistant

Settings