Generative AI
Article 50 asks that visitors know they are talking to an AI. Scan your site free, and read below exactly where that notice goes in Chatbase.
Free · No account · Result in ~30s
Chatbase bots are generative by definition: you upload documents or a sitemap and the model answers questions from them. There is no scenario where a Chatbase widget is "just a form". That makes the notice easy to argue for here, and it makes the system prompt a place where you have to record something the bot must never do.
We have not yet scanned enough sites where Chatbase was the first widget we detected to say anything meaningful here. As soon as the sample is large enough it will appear, counts included. We would rather publish nothing than a percentage based on two sites.
Our scanner looks for these signals on your public page. Two independent signals are required for a match, so a blog post that merely names the vendor does not count.
www.chatbase.cochatbase.cowindow.chatbase#chatbase-bubble-button#chatbase-bubble-windowThe initial message field in the chatbot settings, plus the first line of your system prompt.
Open your chatbot in Chatbase and go to Settings, Chat Interface.
Find the initial message field.
Put the AI notice there as the first line.
Go to the AI settings and add to the system prompt that the bot never denies being an AI.
Save and reload your site; the embed script picks up the new settings.
The sentence you paste
You are chatting with an AI assistant, not a person.
The welcome message only appears after opening. The banner snippet from the Compliance Kit covers the moment before that with a single script tag, and your AI transparency statement covers the question that comes after.
Watch out with Chatbase: Chatbase bots are generative by definition. Also instruct the system prompt to confirm it is an AI when asked.
Chatbase documentation: www.chatbase.co
Three mistakes we run into on sites that did set the notice. Each one costs points in the scan, and each one takes minutes to fix.
Give the bot a name and a human role in the system prompt and, asked "are you a robot?", it will fairly regularly say no. State explicitly that it never does, and test with exactly that question.
Chatbase shows suggested questions the visitor can click. Clicking one jumps past the initial message straight to an answer. Make sure the notice is visible alongside the suggestions, not only above them.
A new Chatbase bot starts with empty fields. We regularly see sites where the notice stood correct for months and vanished on a retrain against a new source. Put the field in your release checklist.
Open your site in an incognito window and read the initial message. Then click one of the suggested questions instead of typing, because that is the route by which visitors skip the initial message. Next, ask literally whether you are talking to a person and read the answer: on a generative bot that is the only real test of your system prompt. Repeat that test after every retrain, because a new bot starts with empty fields.
After that you can scan again. Our scan looks at the same public page a visitor does, so what you see in an incognito window is roughly what we see.
If your system prompt gives it a human persona, yes. Instruct it explicitly never to do that, and test by asking literally.
No, the visitor never sees it. The notice has to be visible in the interface. The prompt is an addition, not a replacement.
Training on your own content does not by itself make you a provider under the AI Act; you are deploying an existing system. Placing it on the market as a product changes that picture. Put it to a lawyer.
The Compliance Kit contains a copy-paste disclosure banner (NL/EN/DE), an AI transparency statement, an AI system inventory and a step-by-step guide per widget.
See the Compliance Kit (€149)Technical compliance tooling and templates. Not legal advice.