Since 2 August 2026, a chatbot that talks to people in the EU has to tell them it is an AI. The notice must come no later than the first interaction and must be clear, easy to notice and accessible. That is the core of Article 50(1) of the EU AI Act (Regulation (EU) 2024/1689). The only exception is when it is already obvious to a reasonably well-informed person that they are talking to a machine, and experts advise against relying on that exception.
For a website chat widget, this means four things. People need to see an “AI” label. The opening message should say what they are talking to. The bot must answer honestly when someone asks whether they are talking to a real person. And the bot must not use a human persona that suggests otherwise. The rest of this guide covers who carries the legal duty, what the 2026 “digital omnibus” did and didn’t change, five disclosure messages you can copy, and the most common mistakes.
This article is general information for small businesses, not legal advice. It reflects the regulation, official guidance and law-firm commentary available on 13 September 2026. If your chatbot handles regulated topics such as health, finance or legal services, talk to a qualified lawyer.
What Article 50(1) actually requires
Article 50(1) puts the duty on providers. They must design AI systems “intended to interact directly with natural persons” so that the people concerned are told they are interacting with an AI system. Article 50(5) sets the manner: the information must be given “in a clear and distinguishable manner at the latest at the time of the first interaction”, and it must meet applicable accessibility requirements.
The European Commission’s FAQ on Article 50 lists four conditions for the obligation to apply:
- the system qualifies as an AI system;
- it is designed for a genuine two-way exchange, not just collecting data;
- the interaction is direct, meaning the AI itself talks to the person rather than through a human intermediary;
- the other party is a natural person.
An AI chat assistant on your website that answers customers in real time meets all four. One consequence of the “direct” condition: when an agent asks an AI to draft a reply, reads it, edits it and sends it, the AI is not talking to the customer directly. Many helpdesks offer this workflow. It still helps to be open about it, but it is a different situation from a bot answering on its own.
The “obvious” exception
No disclosure is needed where it is obvious to “an average person, who is reasonably well-informed, circumspect, and observant” that they are dealing with AI (Commission FAQ). The Future of Life Institute’s practical guide to Article 50 warns that the exception “should not be over-relied on”. A chat bubble in the corner of a shop, with a friendly name and a typing indicator, is not obviously a machine to most customers. Assume you need to disclose.
When does it apply, and did the 2026 omnibus delay it?
Article 50 has applied since 2 August 2026. In May 2026, EU lawmakers agreed the “digital omnibus” amendments to the AI Act: the provisional agreement came on 6 May and the Council confirmed it on 13 May. The omnibus pushed back the high-risk system deadlines. It did not delay the chatbot disclosure duty in Article 50(1).
The one Article 50 change was to the watermarking duty in Article 50(2). Systems already on the market before 2 August 2026 have until 2 December 2026 to comply with it (Gibson Dunn). The Commission adopted its guidelines on Article 50 transparency obligations on 20 July 2026. They are not binding, but national authorities are expected to use them as their reference.
Who is responsible: the provider or the deployer?
Article 3 defines the two roles. A provider develops an AI system, or has one developed, and places it on the market or puts it into service “under its own name or trademark”. A deployer uses an AI system under its own authority in a professional context.
| Your situation | Likely role | What that means in practice |
|---|---|---|
| You use a chat widget from a software vendor as it comes | Deployer | The vendor must build disclosure in. Don’t undo it with your name, avatar or greeting settings. |
| You build your own bot on a language-model API and run it on your site under your brand | Provider | The Article 50(1) duty is yours. Design disclosure into the product. |
| You white-label a vendor’s bot and sell it to clients as your product | Likely provider | You are placing it on the market under your own name. Take advice. |
Location doesn’t get you out of it. Article 2(1)(c) covers providers and deployers based outside the EU where the AI system’s output is used in the Union. A UK or US shop whose chatbot answers customers in France is in scope.
Even as a deployer, you usually control the parts customers notice: the bot’s name, avatar and first message. Legal responsibility may sit with your vendor, but customers see a business that hid the AI.
What good disclosure looks like in a chat widget
- A visible “AI” label in the widget header that stays for the whole conversation, not only on the first screen.
- An opening message that says it plainly: the bot’s name, the fact that it is an AI, what it can help with, and how to reach a person.
- Text, not just an icon. A robot icon means nothing to a screen reader. Article 50(5) requires disclosure that meets accessibility requirements.
- An honest answer to “are you a bot?” The bot should answer that directly every time, whatever the rest of its instructions say.
- A clear notice when a human takes over, and another if the conversation goes back to the AI.
- No human photograph as the bot’s avatar. A name is fine. A face that implies a person is not.
- Disclosure in the language of the chat. If the bot answers in German, the notice should be in German.
Five disclosure messages you can copy
Replace the details in square brackets. Each one names the AI, says what it does and gives the customer a way to reach a person.
1. Standard greeting
Hi, I’m [Bot name], an AI assistant for [Business]. I answer from our own policies and guides. If you’d rather talk to a person, type “talk to someone” at any time.
2. Short version for small widgets
You’re chatting with an AI assistant. Ask about [orders, bookings, opening hours]. Want a human? Just say so.
3. Answer to “Am I talking to a real person?”
No, I’m an AI assistant, not a person. I can keep helping, or I can pass this conversation to our team. They usually reply within [time] during [office hours].
4. Handover to a human
I’m passing you to [Agent name] from our team. They’re a person, and they can see our conversation so far, so you won’t need to repeat anything.
5. Outside office hours
You’re talking to our AI assistant. The team is offline until [time and time zone]. I can try to help now, or take your email so a person can reply when they’re back.
Common mistakes
- A bot posing as a named human. “Sarah, Customer Care Specialist” with a stock photo, answering instantly at 3am, is exactly what Article 50 targets.
- Scripted dodges. Replies like “I’m here to help!” to the question “are you a bot?” are evasive. Some older scripts even said “No, I’m real”. Both need to go.
- Disclosure only in the terms or privacy policy. Law firm Travers Smith notes that information easily overlooked, such as text buried in terms and conditions, doesn’t meet the “clear and distinguishable” standard.
- Disclosing once, then hiding it. A first-message notice that scrolls away, followed by a human-looking avatar for the rest of the chat, weakens the disclosure.
- Blending AI and human replies. If customers can’t tell which messages came from a person, the handover isn’t transparent.
- Assuming the vendor has it covered after you have renamed the bot, changed its avatar and rewritten its greeting.
What are the penalties?
Under Article 99(4), breaching the transparency obligations in Article 50 can bring fines of up to €15 million or, for companies, 3% of total worldwide annual turnover for the previous financial year, whichever is higher. Article 99(6) protects small businesses: for SMEs and start-ups, the cap is whichever of the two figures is lower. Enforcement is handled by national authorities, and for most small businesses customer trust is at least as important as the fine.
A quick compliance checklist
- The widget header shows an “AI” label in text, readable by screen readers.
- The first message says the assistant is an AI and explains how to reach a person.
- Asking “are you a real person?” gets a direct, truthful answer.
- The bot’s avatar is not a photo of a person.
- Handovers to and from humans are announced in the chat.
- Disclosure appears in every language the bot speaks.
- You know whether you are the provider or the deployer, and you have kept a note of why.
- Your privacy notice mentions the AI assistant, in addition to the in-chat disclosure.
How KuraChat handles it
We built KuraChat so disclosure can’t be switched off. The widget always shows an AI label, the assistant uses your chosen name but says it is an AI when asked, and there is no stealth mode. Our AI transparency page explains the design choices. The features overview shows how handoff to a person works, so customers always have a route to a human.
Frequently asked questions
Does Article 50 apply to businesses outside the EU?
Yes, if the chatbot's output is used in the EU. Article 2(1)(c) covers providers and deployers established in third countries where the AI system's output is used in the Union, so a UK or US shop serving EU customers is in scope.
Can my chatbot have a name?
Yes. A name is fine as long as it is clearly presented as an AI from the first message and the bot answers truthfully when asked. What causes problems is a human-sounding persona with a photo that suggests a real employee.
Is it enough to mention the AI in my privacy policy?
No. Article 50(5) requires the information to be clear and distinguishable at the latest at the first interaction. A mention buried in terms or a privacy policy is easy to overlook, so put the disclosure in the chat itself.
Do I have to disclose AI-drafted replies that a human agent sends?
The Commission FAQ says Article 50(1) covers direct interaction, where the AI itself communicates with the person. A reply an agent reviews and sends is not direct AI interaction, but being open about AI assistance is still good practice.
What is the maximum fine for breaking Article 50?
Up to €15 million or 3% of worldwide annual turnover, whichever is higher, under Article 99(4). For SMEs and start-ups, Article 99(6) applies whichever of the two is lower.