Tawk.to and the EU AI Act: does your chat widget tell people it's AI?
Tawk.to is the internet's default free live-chat widget, and since AI Assist shipped, the same green bubble can be a human agent, Apollo AI Bot answering 24/7, or an AI that hands off mid-conversation. Visitors can't tell which from the outside: the widget looks identical either way. That ambiguity is the Article 50(1) problem in miniature, if Apollo answers first and nothing says so, your free chat widget quietly became an undisclosed AI system on an EU-facing site.
The rule itself is short. Article 50(1) of the EU AI Act requires AI systems that interact directly with people to be designed so users know they're dealing with AI, no later than the first interaction, in a clear and distinguishable way. The exception for cases where it's "obvious" is narrow: a natural-language customer-service bot doesn't qualify just because it has a robot icon. A line in your terms of service doesn't satisfy it either. And it's easy to enforce, because a regulator can open your site and start a chat.
Where the disclosure lives in Tawk.to
- The chat bubble and maximized widget, identical whether a human or Apollo AI is answering
- The greeting/first reply, when AI Assist is on, Apollo AI typically answers before any human does
- The bot's display identity in conversations (name shown to visitors when the AI responds)
- Human-takeover moments, tawk.to lets agents monitor and take over bot conversations in real time
Your Tawk.to disclosure checklist
- Open your site in a private window as an EU visitor and start a chat: is it explicit you're talking to AI before or in the very first exchange?
- In the tawk.to dashboard, open AI Assist and review the Instructions page, it controls how Apollo greets visitors; put an explicit AI line in that greeting (e.g. "You're chatting with our AI assistant, a human can take over anytime")
- Check what name visitors see when Apollo answers: a teammate-style name with a human photo is the combination to avoid (TODO: verify the exact property/alias setting path for your workspace version, tawk.to has changed this UI)
- If agents take over from Apollo mid-chat, make the participant change visible in both directions
- Verify the disclosure survives on mobile and in the minimized-to-maximized transition
- Screenshot the first-interaction state with Apollo enabled and save your AI Assist settings with a date
Watch out for
The classic tawk.to failure mode is inheritance: the widget predates AI on most sites, so nobody revisits its greeting when AI Assist gets switched on. The bubble that honestly meant 'chat with our team' for years starts meaning 'chat with an LLM' with zero copy changes, and because AI Assist is a per-workspace add-on any admin can enable, the gap can appear overnight without a deploy.
Common questions
We use tawk.to with human agents only, does Article 50(1) apply to us?
If no AI system interacts with your visitors, the chatbot-disclosure duty doesn't bite, human live chat isn't in scope. The catch is drift: the moment someone enables AI Assist (it's a dashboard add-on), the same widget becomes an AI interaction. Worth a periodic check rather than a one-time answer.
Where exactly do I add the disclosure in tawk.to?
Two layers: the AI Assist Instructions page (tells Apollo how to greet and behave, put the explicit AI line in the greeting behaviour there), and your widget's pre-chat/greeting copy. Belt and braces: if Apollo answers first, its first message should say what it is regardless of what the widget frame shows.