From 2 August, your chatbot has to tell people it's AI

On 2 August 2026, article 50 of the European AI Act comes into force. From that moment on, anyone talking to your chatbot or AI search function has to know there is AI on the other side.

Whether you call the thing a chatbot or an AI search function makes no difference to the law. Article 50 covers AI systems intended to interact directly with people. Our own widget falls under it just as much as a classic chatbot does. A different name is no escape.

The notice itself is a matter of one line of text. The difficulty lies elsewhere. In June, Brussels pushed a large part of the AI Act back to 2027 and 2028, and since then there is a widespread impression that there is still time. For this part, that doesn't hold. On top of that, the text of the law contains two requirements that barely appear in any checklist.

What the law asks

Article 50(1) covers AI systems intended to interact directly with people. The provider has to design such a system so that the person concerned is informed that they are dealing with an AI system. There is one exception: when this is already obvious to a "reasonably well-informed, observant and circumspect" person, given the circumstances and the context of use. We'll come back to that exception, because it is easy to lean on it too heavily.

Paragraph 5 says when and how. The information has to be given "at the latest at the time of the first interaction", in a clear and distinguishable manner. And the last sentence of that paragraph adds something that often falls away: the information has to meet the applicable accessibility requirements. See the full text of article 50.

Let it slide and you fall under the second tier of fines: up to 15 million euros or 3% of worldwide annual turnover, whichever is higher. In the Netherlands, supervision of the transparency obligations sits with the Dutch Data Protection Authority, within a system of ten regulators in which the RDI is the central point of contact.

Less than two weeks ago, on 20 July 2026, the European Commission published the final guidelines on this article: 51 pages explaining what providers and deployers have to do. That document is the best source when you are unsure about a specific case.

Three things that often get skipped

1. June's postponement wasn't about this article

On 29 June 2026 the Council of the EU approved the Digital Omnibus, the first substantive amendment to the AI Act since it was adopted in 2024. With it, the obligations for stand-alone high-risk systems under Annex III moved to 2 December 2027, and AI in regulated products under Annex I to 2 August 2028.

Article 50 wasn't in that package. The AI literacy obligation in article 4 stayed as well. Anyone who read in the news that the AI Act has been postponed and concluded that nothing is needed until 2027 is looking at the wrong article.

1. 2 August 2026

   Transparency obligations (article 50) and AI literacy (article 4). Visitors have to know they are talking to AI. National regulators start enforcing.

2. 2 December 2026

   Machine-readable marking of AI-generated content (article 50(2)) for generative systems already on the market before 2 August.

3. 2 December 2027

   Obligations for stand-alone high-risk systems under Annex III. Moved by the Digital Omnibus; it used to be 2 August 2026.

4. 2 August 2028

   AI built into regulated products under Annex I. Also moved.

2. The notice itself has to be accessible

Paragraph 5 requires the information to meet the applicable accessibility requirements. That is one sentence with considerable consequences. A grey label in small type, a text that only appears on hover, or a line baked into an image: none of that reaches the visitor working with a screen reader or navigating by keyboard. The notice has to land with everyone, and for this group especially, an AI answer that comes across as human is the most confusing.

There is a concrete signal that this will be the weak point. Digimonitor assessed 39 municipal chatbots against 29 accessibility benchmarks. Not one met the requirements. That same research also showed how those chatbots perform on substance: of 312 answers, 10% were correct and 64% were wrong.

If your AI's interface isn't accessible today, the legally required notice inside it won't be either. So we tested and improved the standard vragen.ai front end on exactly this point: keyboard navigation, visible focus, text labels on buttons and status updates through ARIA live regions.

3. "With us it's obvious it's AI"

The exception in paragraph 1 genuinely exists, and it is narrower than it looks. The measure isn't what you know about your own widget, but what a reasonably well-informed, observant and circumspect person would assume in that context of use. A neatly designed input field in your own brand styling, without a name or a label, is exactly the case where it is not obvious. The better your widget is integrated into the site, the weaker your appeal to this exception becomes.

The final guidelines of 20 July also make two things explicit that run against intuition. Human involvement in the process doesn't remove the transparency obligation for AI answers. And an AI agent also has to say on whose behalf it acts.

What such a notice can look like

The law prescribes no particular form and a popup isn't required. The difference is in the placement and how easy it is to find.

• Not enough

  "Hi! How can I help you?"

  Nowhere does it say this is AI. A mention in the terms and conditions, or halfway through the conversation, is too late: paragraph 5 asks for the first interaction.

• Enough

  "You are asking the AI assistant of Municipality X. I look the answer up in our own information and show you where it comes from. If I can't work it out, I'll point you to a colleague."

  Right when it opens, in plain text a screen reader reads out. It says this is AI and on whose behalf it speaks, and offers a route to a person.

Where you put the notice is freer than what it looks like. An opening message works. A fixed line under the input field works just as well and draws less attention: it is already there before anyone types their first question and stays for as long as the conversation runs. An extensive information panel next to it is allowed too, but that is heavy artillery for this one statement.

What doesn't work is a block that only appears below the first answer. By then the first interaction has already happened, and paragraph 5 asks for that moment at the latest. The same mistake sits in a notice that only lives in the terms and conditions or on a separate explainer page.

What you can check now

• Does it say at the first interaction that this is AI, rather than only after the first question?
• Is that notice real text in the page, visible without hover and readable by a screen reader?
• Does it also say which organization the AI speaks for?
• Is it clear how a visitor reaches a person when the answer isn't there?
• Do you use generative AI for images, audio or video? Then machine-readable marking comes on top, for existing systems from 2 December 2026.
• Do you know who in your organization is accountable for this? That is part of the AI literacy obligation taking effect on the same day.

How we look at this at vragen.ai

Article 50 divides the work. Paragraph 1 puts the design obligation on the provider of the system, so on us: it has to be possible to get the notice in the right place and the right form. What actually appears in your environment is up to you. No supplier can tick this off for you, and we don't sell it that way either.

In vragen.ai you set the example questions and the text in the input field per environment, so that is where you can put the notice in your own words. Two things don't count as a notice: a placeholder like "Ask us anything" and a sender label above the answer don't say you are talking to AI. And a brand name containing "ai" certainly doesn't. That last point applies to us just as much.

As for the rest of the conversation: every answer comes from your own content and points to where it was found, so a visitor can check what it is based on. If the answer isn't there, the system says so and refers them on. That honesty is exactly where most chatbots come unstuck.

One thing to watch if you work on the design yourself: adjust the styling or the front end and the accessibility of your version is your work too. What is set out about hosting, data storage and anonymization is on the security page.

Short answers

Do I have to disclose that visitors are talking to an AI chatbot?

Yes. Article 50(1) requires providers to design AI systems that interact directly with people so that the user knows they are communicating with AI. That information comes at the latest at the first interaction, clear and distinguishable. It applies from 2 August 2026.

Does the AI Act postponement also cover the transparency obligation?

No. The Digital Omnibus pushed the high-risk obligations back to 2 December 2027 and 2 August 2028. Article 50 and the AI literacy obligation in article 4 stayed on 2 August 2026.

What should the notice look like?

The law prescribes no particular form and a popup isn't required. The information has to be clear and distinguishable at the latest at the first interaction, and meet the applicable accessibility requirements. Halfway through the conversation or in the terms and conditions is too late.

What is the fine?

Breaching article 50 carries a fine of up to 15 million euros or 3% of worldwide annual turnover, whichever is higher. In the Netherlands, the Dutch Data Protection Authority supervises the transparency obligations.

Can I rely on "it's obvious it's AI"?

That exception exists, but it is assessed from the point of view of a reasonably well-informed, observant and circumspect person in that specific context of use. A widget in your own brand styling without a label is exactly the case where it is not obvious.