Does Your Website Chatbot Need an AI Disclosure? What the EU AI Act Requires Since August 2, 2026

Yes, in most cases. Since August 2, 2026, Article 50 of the EU AI Act requires that people who chat with an AI system are told it's an AI, clearly and no later than their first message. If your website chatbot is open to visitors in the EU, it should say "AI" somewhere they can't miss before the conversation starts.

I build ChatBento, an AI support chatbot for websites, so I had to read the rule closely for our own widget. This post covers what the text actually says, who it applies to, what a compliant chat widget looks like, and the mistakes I'd avoid. It's a practical reading, not legal advice. If you're in a regulated industry or unsure, ask a lawyer.

What does Article 50 actually say?

The core of it is one sentence in Article 50(1): providers must make sure AI systems "intended to interact directly with natural persons" are designed so those people "are informed that they are interacting with an AI system."

Two more details matter for website chat:

• Timing and format. Article 50(5) says the information must be given "in a clear and distinguishable manner at the latest at the time of the first interaction." It also has to meet applicable accessibility requirements.
• The "obvious" exception. The duty doesn't apply where it's obvious to a "reasonably well-informed, observant and circumspect" person that they're dealing with AI.

The article has applied since August 2, 2026.

Does it apply to my small business?

Probably, if any of your visitors are in the EU. The AI Act covers providers and deployers inside the EU, and also those outside it when the AI system's output is used in the EU. A US shop with European customers chatting to its bot is in scope for those conversations.

The duty in Article 50(1) is formally on the provider: whoever builds the AI system and puts it on the market. If you use a chatbot product from a vendor, the vendor is responsible for designing the disclosure in. If you built your own bot on top of a model API, you're the provider.

In practice that distinction doesn't help you much. It's your website, your brand and your customer. If the widget you installed doesn't disclose, the customer's experience is your problem, and you're the one who chose the tool. Check it either way.

Can I rely on the "obvious" exception?

I wouldn't. "Obvious" is judged from the visitor's side, and a chat bubble in the corner of a website looks exactly like the live chat many businesses run with real people. A friendly name and a photo avatar make it look even more human.

The exception was written for cases where nobody could be confused. A support widget called "Sarah" with a stock photo is the opposite of that. A two-word label costs nothing, so there's no reason to argue about whether you needed it.

What does a compliant chat widget look like?

The rule says "clear and distinguishable" and "at the latest at the time of the first interaction." For a website chat widget, I read that as:

1. A visible label before the first message. The visitor should see that it's AI when the window opens, not after they've typed. A label in the chat header works well because it stays on screen for the whole conversation.
2. Plain words. "AI assistant" says it. "Powered by advanced language models" in grey 9-pixel text at the bottom doesn't.
3. Not buried in your terms or privacy policy. A sentence in a document nobody opens during a chat isn't given "at the time of the first interaction."
4. Readable by assistive technology. Article 50(5) says the information must conform to accessibility requirements. An "AI" badge that exists only as an image, or only as a color, fails screen reader users. Use real text.
5. Consistent with the greeting. If the opening message says "Hi, I'm Sam from the support team!", you've undone the label above it. Something like "Hi, I'm the AI assistant for Acme. Ask me about orders, shipping or returns" is honest and still friendly.

In ChatBento, the chat header reads "AI assistant · replies instantly" under your bot's name, as real text, from the moment the window opens. You choose the name and the greeting, so write the greeting to match.

Can my bot still have a name and personality?

Yes. Nothing in Article 50 bans a name, an avatar or a warm tone. The rule is about not misleading people about what they're talking to. "Penny, the AI assistant for Penny's Bakery" is fine. "Penny" with a human headshot and no mention of AI, who claims to be "checking with the kitchen", is the kind of thing the rule exists to stop.

A good test: if a visitor asked "am I talking to a real person?", would they be surprised by the honest answer? If yes, add the label.

What about handing off to a human?

Disclosure matters most at the seam between the AI and your team. When a conversation moves from the bot to a person, say so. When a person's reply comes back by email, it should be clear it's from a person.

The simplest pattern, and the one ChatBento uses, keeps the two separate: the AI answers in the chat, and when it can't help, it offers a contact form that sends the visitor's name, email and question to your inbox. Your reply then comes from you, by email. There's no point in the chat where a human quietly takes over the AI's identity, or the other way round.

What are the fines?

Under Article 99(4), breaching the Article 50 transparency obligations can mean fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs and start-ups, Article 99(6) flips that to whichever is lower.

Nobody expects a regulator's first move to be a seven-figure fine against a bakery's chat widget. But the fix is a label and an honest greeting, and leaving it out has no upside.

Mistakes I'd avoid

• Disclosing only in the privacy policy. It isn't clear or given at the first interaction.
• Hiding the label until the visitor sends a message. Show it when the window opens.
• Using a human stock photo as the avatar. Use your logo, an illustration, or a photo of your actual team with the AI label right next to it.
• A greeting that claims to be a person. "I'm Jess from customer care" contradicts any label above it.
• An image-only "AI" badge. Screen readers can't read it unless it has proper alt text, and real text is simpler.
• Assuming non-EU means exempt. If EU visitors use the chat, the rule reaches those conversations.

A quick compliance checklist

• The chat window says "AI" in plain text before the first message
• The label stays visible during the conversation
• The greeting doesn't claim the bot is a person
• The avatar doesn't look like a real human, or is clearly labelled
• The label is real text that screen readers can read
• Handoff to a human is clearly marked as a handoff
• You checked your vendor's widget yourself, rather than assuming

The short version

Tell people it's an AI, in plain text, before they start typing, and keep the greeting honest. That covers the core of Article 50 for a website chatbot, and it also happens to build trust: visitors forgive an AI that says "I don't know, let me pass this to the team" far more readily than a "person" who turns out not to be one.

If you're choosing a chatbot now, check how each one labels itself before you install it. ChatBento's widget shows the AI label by default, installs with one script tag on any platform (see the install guides), and has a free plan so you can see exactly how it looks on your site before paying anything.