Short answer: a chatbot on your website is compliant with the EU AI Act if a visitor can tell, from the very first message, that they are talking to an AI. That duty sits in Article 50(1) and has applied since 2 August 2026. It is one sentence of law, but most website chatbots still fail it, usually because of a friendly human name and a greeting nobody revisited.
This guide covers who the rule applies to, when the “obvious from context” exception helps you, what a compliant disclosure looks like, and the other laws that still apply to the same chat window.
What Article 50(1) actually requires
Article 50(1) says providers must design and develop AI systems that interact directly with people so that those people are informed they are interacting with an AI system. There is one exception: when that is already obvious to a reasonably well-informed, observant and circumspect person, taking into account the circumstances and the context of use.
Article 50(5) then adds how the information must be given:
- in a clear and distinguishable manner,
- at the latest at the time of the first interaction,
- in line with applicable accessibility requirements.
Breaching it falls into the AI Act’s middle fine tier: up to 15 million euros or 3% of worldwide annual turnover, with the lower of the two applying to SMEs. Our guide to EU AI Act fines and penalties explains how authorities set the actual amount.
Provider or deployer: who is responsible for your chatbot?
On paper, Article 50(1) is a provider obligation. In practice, the line is blurrier than many site owners assume.
- You use an off-the-shelf widget (Tidio, Intercom, Crisp, Zendesk and similar). The vendor is the provider of the AI system and should build disclosure into the product. But most of these widgets let you rename the bot, upload a face and write the greeting. If you configure “Hi, I’m Laura from the team” and remove any AI hint, the visitor is misled by your configuration, not the vendor’s code.
- You built your own bot on a model API (for example a custom assistant on top of OpenAI, Anthropic or Mistral, wired into your site by a developer or plugin). Here you, or your agency, can become the provider of that AI system, because you put it into service under your own name. The model maker is the provider of the general-purpose model, not of your chatbot. The disclosure duty then lands on you directly.
- An agency built it for a client. Who counts as provider depends on whose name the system is put into service under. Our guide on what agencies must do under the EU AI Act goes through the common setups.
Either way, the practical conclusion is the same: whoever controls the chat window’s first message should make sure it says the visitor is talking to an AI. Waiting for the vendor to fix it is not a defence if you were the one who wrote the greeting. The same logic is covered more broadly in are you responsible for third-party AI tools on your website.
When is it “obvious” that it’s a bot?
The exception is narrower than it sounds. The test is a reasonably well-informed, observant and circumspect person, and the Commission’s guidance points out that you should also consider vulnerable groups who might be using the system, such as children or older users.
| Setup | Obvious it’s AI? |
|---|---|
| Widget titled “AI Assistant” with a robot icon and the first message “I’m an AI assistant” | Yes, and you’ve disclosed anyway |
| Bot called “Laura” with a stock photo, typing indicator and “Hi, how can I help?” | No |
| Live chat that is a person during office hours and an AI after hours | No, the switch is invisible |
| Phone line with a natural-sounding AI voice | No |
| Product finder quiz with fixed buttons and no free text | Often not an AI system at all; check whether anything is generated |
If you have to argue about whether it is obvious, it isn’t. A short line costs nothing.
What a compliant disclosure looks like
Put it in the first thing the visitor sees from the bot: the greeting bubble, the widget header, or both. Examples that work:
- “Hi, I’m the Acme AI assistant. I can answer questions about orders and returns. Type ‘human’ to reach our team.”
- Widget header: “AI assistant” with the bot’s name underneath.
- For voice: “You’re speaking with an automated AI assistant” before the first question.
Things that don’t work on their own:
- A line in your privacy policy or terms. That is not “at the latest at first interaction”.
- A tiny grey “Powered by AI” in the widget footer that most people never notice.
- A disclosure only in the image of an avatar, with no text a screen reader can pick up.
- Telling people only when they ask “are you a bot?”
A persona name is fine. “Laura, our AI assistant” is compliant. “Laura from support” is not. For placement patterns across the rest of your site, see where to place AI disclosures on your website.
Hand-offs between bot and human
Many support tools mix AI and human agents in the same thread. Tell the visitor when the switch happens in both directions: “You’re now chatting with Mark from our support team” and “Our team is offline, you’re now chatting with our AI assistant.” Without that, a visitor who was told “this is a person” at 16:59 has no way to know it changed at 17:00.
Other rules that still apply to the same chat window
The AI Act is not the only law in play. A compliant chatbot also needs to respect:
- GDPR. Chat transcripts are personal data. Your privacy notice needs to mention the chatbot, its vendor (usually a processor), where data goes, and how long transcripts are kept. If the vendor uses conversations to train its models, you need a legal basis for that, or switch it off. Our comparison of the EU AI Act and GDPR explains how the two fit together.
- Consumer protection law. A bot that invents a refund policy or a discount is your statement to a customer. Courts and regulators have already held businesses to what their chatbots said. Limit the bot to your real policies and test its answers.
- Cookie and tracking rules. Many widgets set cookies or load scripts on every page. That falls under the ePrivacy rules and your consent banner, not the AI Act.
- Article 4 AI literacy. Since 2 February 2025, organisations using AI must take measures to support the AI literacy of staff who run it. For a chatbot, that means the person who configures it understands what it can get wrong and how to check it.
Checklist: is your website chatbot compliant?
- Open your site in a private window and start a chat. Does the first message or the header say it’s an AI?
- Check the bot’s name and avatar. Could a visitor reasonably think it’s a person?
- If humans and AI share the chat, is every switch announced?
- Can a screen reader read the disclosure? Test with your browser’s built-in reader or check the widget markup.
- Do you know whether you are the provider (custom bot) or a deployer (configured widget)? Write it down.
- Is the chatbot listed in your privacy notice, with vendor and retention period?
- Have you tested ten realistic questions, including ones about prices, refunds and legal topics, and checked the answers?
- Keep a dated note of what you checked. Our website AI audit guide has a template.
Adding the notice in WordPress
Most widgets let you change the greeting in their own dashboard, and that is the best place to fix it. If you want a consistent notice across widgets without editing each one, the free EU AI Act Ready plugin adds an AI disclosure badge to 15 common chat widgets, logs the setting in its compliance report, and runs on your own server with no account. The same plugin handles post-level disclosure for AI-generated content, so you can keep all your Article 50 decisions in one place.
Frequently asked questions
Does a rule-based chatbot need an AI disclosure?
Article 50(1) covers AI systems. A bot that only follows fixed decision trees with pre-written answers may not be an AI system under the Act’s definition. Most modern widgets now generate answers with a language model, though, and those clearly are. If you’re not sure, disclose. It costs one line.
Is “Powered by AI” in the footer of the widget enough?
Rarely. The disclosure has to be clear and distinguishable at the first interaction. A small footer line that most people never notice is hard to defend. Put it in the greeting or header.
Can my chatbot have a human name?
Yes, as long as it also says it’s an AI. “Mia, your AI shopping assistant” is fine.
Does this apply if my business is outside the EU?
If the chatbot is offered to people in the EU, or its output is used in the EU, assume Article 50 applies. Being based elsewhere doesn’t exempt you.
When did the chatbot rule start to apply?
2 August 2026. The Digital Omnibus did not delay it. The only Article 50 delay, to 2 December 2026, concerns machine-readable marking of AI-generated content by providers of generative systems already on the market, not chatbot disclosure.
Practical guide: AI chatbot disclosure: copy-ready wording and setup.