EU AI Act transparency: what your website chatbot must tell visitors
By Miklós KazatsayPublished: 10 min read

That they are talking to an AI, clearly and by the first message. Article 50 of the EU AI Act is its transparency rule: since 2 August 2026, anyone interacting with an AI must be told so, unless it is obvious. For a website chatbot or AI phone assistant, that is one sentence, not a legal project.
- Since 2 August 2026, a chatbot or AI voice assistant has to say it is a machine by its first message.
- The Digital Omnibus did not delay chatbot disclosure: the 2 December 2026 grace period covers only machine-readable marking of generative AI output, and only for systems already on the market.
- Under the European Commission's guidelines, it is not “obvious” that a customer-service chatbot is a machine, so disclosure is required there.
- A line buried in your terms or privacy notice is not enough on its own: the disclosure belongs in the chat window or the first sentence of the call.
- AI-written ads and product descriptions need no label, but an AI-generated product photo that makes the product look better than it is can count as a deep fake.
What does Article 50 of the AI Act require, and since when?
Article 50 of the AI Act requires that anyone interacting directly with an AI system is told so. It is part of Regulation (EU) 2024/1689 on artificial intelligence, it has applied since 2 August 2026, and it covers website chatbots and AI phone assistants regardless of company size.
Two paragraphs usually matter for a small firm: paragraph 1 on the chatbot or voice assistant visitors talk to, and paragraph 4 on AI-made content you publish. Paragraph 5 sets the manner: the information has to be clear and distinguishable, given no later than the first interaction, and it has to meet accessibility requirements.
Plenty of articles this summer said the EU had postponed the AI Act. But the amending regulation published in July 2026, the Digital Omnibus (Regulation (EU) 2026/1744), pushed back the high-risk deadlines. Chatbot disclosure stayed put.
| Obligation | Who carries it | Applies from |
|---|---|---|
| Telling people they are talking to an AI (Art. 50(1)) | The provider of the chatbot or voice assistant | 2 August 2026 |
| Labelling deep fakes and public-interest AI text (Art. 50(4)) | Whoever uses the system and publishes the content | 2 August 2026 |
| Machine-readable marking of AI output (Art. 50(2)) | The provider of the generative AI | 2 August 2026; 2 December 2026 for systems already on the market |
| High-risk systems, such as recruitment or credit scoring (Annex III) | Providers and users | 2 December 2027 |
So the 2 December grace period is for the makers of generators; there's no extra time for the disclosure in your chat window.
When is it “obvious” that a machine is answering?
When a reasonably well-informed, observant visitor would have almost no doubt that they are dealing with a machine. In that case the regulation asks for no separate disclosure. But the European Commission's Article 50 guidelines, adopted on 20 July 2026, say this exception must be read narrowly.
The guidelines' examples are clear enough. A coding assistant for professional developers is obviously AI; so is an internal HR assistant for trained staff. A chatbot built into a customer-service or helpdesk tool, whose replies users may take for a person's, is not. That is exactly the “How can I help?” bubble on a physio clinic's or a lettings agency's website.
According to the Commission, a human name and photo in the chat window, a natural-sounding voice and a general-public audience, including older people and children, all make the machine less obvious. A smiling “Sophie from customer care” bot is about the worst starting point.
I wouldn't build anything on the obviousness exception. One sentence in the first message costs nothing; relying on the exception means having to justify it if anyone ever asks.
What should you write in the chat window, the phone greeting and the privacy notice?
The chatbot's first message has to say that an AI is answering. For chatbots, the Commission's guidelines suggest a first-turn disclosure plus, where appropriate, a label that stays visible throughout the conversation, placed close to the input field and worded plainly. Some wording to adapt:
- First message: “Hello, I'm [company]'s AI assistant, not a person. I can book appointments and answer common questions. If you'd rather talk to someone on the team, just say so and I'll pass it on.”
- Persistent label in the chat header: “AI assistant”.
- First sentence on the phone: “Good morning, you're speaking with [company]'s AI assistant. This call is recorded and transcribed.”
Under the guidelines, it is not enough on its own if the disclosure sits only in your terms, if the bot just calls itself an “assistant”, or if a generic “we use AI on this site” notice hangs somewhere. On the phone, a tone doesn't replace a spoken sentence.
The AI Act doesn't change your privacy notice, but the GDPR already requires the chatbot in it: what data it asks for, why, how long you keep it, and which vendor processes it. That complements the chat-window disclosure; it doesn't replace it.
With our own AI phone assistant, Echo, the first sentence of every call says that an AI assistant is speaking and that the call is recorded and transcribed. It's on by default and can't be switched off.
„An AI assistant's first sentence tells the caller who is speaking and what happens to what they say. It's worth rereading that sentence after every settings change, because it's only true for as long as the system actually works that way.”
Who is responsible for the chatbot: the vendor or you?
Article 50(1) puts the duty on the provider: whoever develops the chatbot and offers or runs it under their own name. The deployer is the business using it under its own authority, say by adding an off-the-shelf chatbot to its website. With a packaged chatbot, building in the disclosure is mainly the vendor's job.
It comes back to you in two ways. Build the chatbot in-house, or substantially modify one and run it under your own name, and the guidelines make you the provider. And visitors see your logo, not the vendor's: if there's an AI-label switch and it's off, the lost trust is yours.
Under Article 99, breaching Article 50 can mean fines of up to EUR 15 million or 3% of worldwide annual turnover; for SMEs the cap is whichever is lower. Enforcement sits with national market surveillance authorities, and each member state designates its own, so check who that is in the countries where you operate.
In Hungary, our home market, the decree implementing the Act (344/2025) gives the market surveillance role to a ministerial portfolio, not an agency. Portfolios were reshuffled after the 2026 change of government, and as of October 2026 we found no official statement on which ministry holds it, so check directly.
Do AI-generated text and images on your website need a label?
Mostly not, with one important exception. Article 50(4) requires a label from whoever uses the AI in two cases: realistic, deceptive images, audio or video made with it (deep fakes), and AI-written text published to inform the public on matters of public interest without human editorial review.
The guidelines cover the business edge cases. AI-written ads and product descriptions are outside the rule unless they make health, consumer-safety or sustainability claims. An AI product photo that makes the product look better or different than it is counts as a deep fake, as does a realistic AI double of your CEO giving a speech. A real product on an AI-generated background is fine if it doesn't mislead.
Every post on the brizz.hu blog carries the line “Drafted with AI assistance, edited and fact-checked by a human.” A person edits our articles, so paragraph 4 probably wouldn't require it. We show it anyway, because a reader wants to know the same thing as a chatbot user: who, or what, is talking to them. More in our answer on what AI-enabled means.
Under the guidelines, content made and published before 2 August 2026 needs no retroactive label.
What should you check on your own website this week?
The order follows risk: what a visitor notices first comes first. The first four steps take about an hour.
- 1Open your website in a private window, start the chatbot and read its first message. If it doesn't say clearly that an AI is answering, fix that first.
- 2Check the chatbot's name and profile picture. If it shows a human name and photo, switch to a neutral icon or add “AI assistant” next to it.
- 3Ask your chatbot or phone-assistant vendor in writing how they meet Article 50(1), and whether there is a setting for it. If there is, make sure it's switched on.
- 4If you use an AI phone assistant, call your own number and listen to the first sentence. It must say that a machine is speaking, and that the call is transcribed or recorded if it is.
- 5Update your privacy notice: what data the chatbot collects, why, how long you keep it, and which vendor processes it.
- 6Review the AI-made images on your site. Anything that shows a product or a real person differently from reality should be labelled or replaced with a real photo.
- 7Write down when you checked and what you found. When the vendor updates the system, repeat the first four steps.
If you're still planning a chatbot or other AI tool, ask for the disclosure sentence in the first draft. Our System service starts chatbot work there too.
What doesn't change
- A booking or customer-service chatbot is not a high-risk system: no permit, registration or conformity assessment. It is different if the system screens job applicants or assesses creditworthiness; stricter rules apply to those from 2 December 2027.
- Your GDPR obligations stay the same. The AI Act sits on top of them.
- Consumer protection law still applies: according to the Commission, if AI is a main characteristic of your service, you may have to say so before customers sign up, whether or not the interaction is obvious.
Who this is not for: if your firm builds and sells AI products, the provider obligations (machine-readable marking, documentation) mean considerably more work than this. This article is general information, not legal advice, and reflects the rules as of October 2026; for your own situation, ask a lawyer or data protection professional.
What does your chatbot's first message say today, and would a visitor know from it that no person is answering?
Frequently asked questions
Does a chatbot that only picks from pre-written answers need a disclosure?
If the bot is rule-based, choosing from pre-written replies rather than generating them with an AI model, it probably isn't an AI system under Article 3 of the regulation. The line isn't always sharp, and visitors still benefit from knowing the replies are automatic, so we'd label it anyway.
Is it enough to say it once, at the start?
According to the Commission's guidelines, a single prominent notice before the first interaction is enough in most cases. In riskier contexts such as health, financial or legal advice and complaints handling, reminders during the conversation are advisable.
Does the 2 December 2026 grace period apply to me?
Not if you only use a chatbot on your website. It is for providers of generative AI systems placed on the market before 2 August 2026, and it concerns machine-readable marking. Chatbot disclosure has applied since 2 August.
Miklós KazatsayCo-founder – Operations & AICo-founder of Brizz. He runs how projects work and how AI tools are put to practical use, from brief to handover.
Drafted with AI assistance, edited and fact-checked by a human.
Sources
- EUR-Lex: Regulation (EU) 2024/1689 on artificial intelligence, Articles 50, 99, 111 and 113 (2024)
- EUR-Lex: Regulation (EU) 2026/1744, Digital Omnibus on AI (2026)
- European Commission: Guidelines on the transparency obligations under Article 50, C(2026) 5054 (2026)
- European Commission: Code of Practice on marking and labelling AI-generated content (2026)
- Future of Life Institute: AI Act Explorer, Article 50 consolidated (accessed 2026)
- Mondaq: The Digital Omnibus on AI enters into force today (2026)
- Nemzeti Jogszabálytár: Act LXXV of 2025, Hungarian implementation of the AI Act (accessed October 2026)
- Nemzeti Jogszabálytár: Government Decree 344/2025, as of 2 October 2026