Adding an AI chat widget takes about five minutes these days. A few clicks in Tidio, Intercom, Crisp or a WordPress AI plugin, and your site answers questions at 3 a.m. The EU AI Act has a short rule for this, and it’s pretty reasonable: people should know when they’re talking to a machine.
Let’s go through who the rule is really aimed at, what a good notice looks like according to the European Commission, and the mistakes that make a notice pointless.
A tool, not legal advice. This guide explains public documents in plain language. Your exact role depends on how your chatbot is built and branded. If it matters for your business, ask a lawyer.
The rule
Article 50(1) of the AI Act says AI systems meant to interact directly with people have to be designed so those people know they’re dealing with AI. The only way out is if it’s obvious to a reasonably well-informed, observant and careful person.
Article 50(5) adds that the info has to be clear, easy to spot, and given no later than the first interaction.
This has applied since 2 August 2026. You might have heard that some providers got until 2 December 2026. That extra time is only for machine-readable marking. It doesn’t cover chatbot notices.
Whose job is it, yours or the vendor’s?
Article 50(1) is aimed at providers, the companies that build an AI system and put it out under their own name. The Commission gives two examples:
- A company that sells a ready-made chatbot that businesses plug into their sites as is. That company is the provider.
- An organization that builds its own chatbot in-house and runs it under its own name. That organization is the provider too.
For a typical website, it shakes out like this:
- You embed a vendor’s chat widget as it comes. The vendor is normally the provider and carries most of the responsibility.
- You build your own bot on top of an AI model, or heavily customize one and run it under your brand. You may well be the provider.
Even when it’s mainly on the vendor, it’s still your site and your visitors. Vendor notices are all over the place. Some are tiny, some hide behind a menu, and some you can switch off in the widget settings. Checking what your visitors actually see takes a minute.
Is your chatbot even AI?
The rule is about AI systems that can hold a real back-and-forth. The guidelines point out that simple automated replies, like rule-based canned answers or old-school out-of-office messages, aren’t covered.
So a widget with three fixed buttons (“Shipping”, “Returns”, “Talk to us”) and pre-written answers probably isn’t in scope. A widget that answers whatever people type using a language model definitely is.
And make sure the thing is actually live. A chat plugin that’s installed but never connected to an account shows nothing to visitors. Nothing to disclose yet.
“Everyone knows it’s a bot” rarely flies
The guidelines say the “obvious” exception should be kept for cases where there’s almost no doubt for the average person in the audience. Their examples of obvious cases are narrow: coding assistants used only by professional developers, internal tools for trained staff, characters in a single-player video game.
Modern chatbots write like people. So a support chat on a public website will rarely count as “obvious”. Plenty of sites use the same chat window for both a bot and human agents, which muddies things even more.
What a good notice looks like
The guidelines list what works and what doesn’t work on its own.
What works
- A plain-language label or banner, like “You are interacting with an AI system.”
- A first message where the bot says it’s an AI.
- A badge or label that stays put for the whole conversation.
- Putting it right by the chat, near the input field, in simple words.
- A recognizable “AI” symbol alongside the text.
According to the guidelines, one clear notice before the first interaction is likely enough in most cases. In riskier situations, or if kids or other vulnerable people are likely to use the chat, you may need reminders along the way.
What doesn’t cut it alone
- A line in your terms and conditions, a URL or the docs. Fine as an extra, not as a replacement.
- Vague names like just “Assistant”, or human-looking avatars that suggest a real person.
- Generic statements like “Services on this website use AI.”
- Tech talk like “This chat uses LLMs” that never actually says the visitor is talking to an AI.
- Hidden metadata or watermarks visitors can’t see.
Don’t promise a human you don’t have
A notice shouldn’t promise what you can’t deliver. “Ask for a person at any time” is only true if someone’s actually there to take over. If the chat runs on AI alone after hours, say so, or stick to what’s always true:
You are chatting with an AI assistant, not a person.
A two-minute test
Open your site in a private window, on your phone if you can, and start a chat like a visitor would.
- Before typing anything, can you tell it’s AI? Check the chat header, the welcome message and the area around the input box.
- Does the first reply say it’s AI? Look at the greeting in your chatbot settings.
- Does the notice stay visible as the conversation scrolls?
- Is the wording plain? “AI assistant” is clear. “Your friendly helper Emma” isn’t.
- Does it work in every language your site speaks?
If the answer to number 1 is no, add a notice.
How AuthorLabel helps
It finds your chat widget. The first-run check scans your installed plugins and page scripts for common chat widgets, including Tidio, Tawk.to, Crisp, Intercom, HubSpot and AI Engine. It also knows the difference between a widget that’s live and one that’s just installed. Say Tidio is installed but has no public key set. The check tells you visitors can’t see it, instead of asking you to label a chatbot that doesn’t exist yet.
It shows the notice. You’ve got three ways to do it:
- a small badge next to the chat widget, so the notice sits right where the conversation happens
- a footer bar across the bottom of the page
- a shortcode or block you place yourself, for example on a custom chat page
The default text is “You are chatting with an AI assistant, not a person.” You can change it, and developers can tweak it with the authorlabel_notice_text filter. Of the three, the badge lines up best with the guidelines’ advice to put the notice next to the chat. A footer bar shows on every page, but it sits further from the conversation.
EU visitors only, if you want. You can show the notice only to visitors from EU countries. AuthorLabel uses the country header your CDN already sends (Cloudflare’s, for example) and never calls an outside service. No header? Then everyone sees the notice.
Your “How we use AI” page. Once you confirm in the first-run check that you use a chatbot, the generated transparency page lists your chat tool along with the other AI tools you use. It starts as a draft you can edit before publishing.
AuthorLabel doesn’t change the chatbot itself and doesn’t touch other plugins. If your vendor has its own AI notice or greeting setting, switch that on too. They work fine together.
The short version
- The chatbot rule is mostly on the provider, but if you build or brand the bot yourself, that might be you.
- The notice belongs in the chat, at the start, not buried in your terms.
- “Obvious” is a narrow exception. Public customer chats rarely qualify.
- Keep the words plain, keep the notice visible, and don’t promise a human you don’t have.
Sources
- Regulation (EU) 2024/1689 (AI Act), Article 50(1) and (5).
- European Commission, Guidelines on the transparency obligations under Article 50 of the AI Act, C(2026) 7702 final, Sections 2 and 3. digital-strategy.ec.europa.eu
The Commission’s guidelines aren’t legally binding.
Last reviewed: November 2026. AuthorLabel is a tool, not legal advice.