Skip to content

AI Act, Art. 50: checklist for your chatbot

  • The provider (us, Botshido) must design the system so that the disclosure exists and cannot be bypassed: the widget must be able to tell the visitor, in the interface, that they are talking to an AI.
  • The deployer (you, the site owner) must make sure it is actually shown, correctly worded, and not hidden, removed or contradicted by the way your chatbot presents itself.
  1. Say it’s AI, not a person. The visitor must be explicitly informed about the artificial, non-human nature of the counterpart. “Powered by AI” alone is a technical description; write it as a statement about the interaction.
  2. At the first interaction, at the latest. Not in the footer of the site, not in the terms, not in a document. In the chat, before or at the moment the conversation starts.
  3. Clearly and distinguishably. Visible text near the input or as the first message; not a grey 9px line, not a link labelled “Disclaimer” that the visitor has to click to discover they are talking to a machine.
  4. Consistently, for the whole conversation. If your persona has a human name and a human face, the disclosure is what makes it honest — keep it visible (a badge or a persistent label for sensitive sectors: health, legal, finance, insurance, complaints).
  5. Never misleading. If a human takes over, don’t let the bot keep claiming to be the same entity; if the bot is asked “are you a human?”, it must answer that it is an AI.
  • ❌ A disclosure only in the privacy policy, the terms of service or the documentation.
  • ❌ A generic line such as “services on this site use AI”.
  • ❌ A purely technical statement (“this chatbot uses an LLM”) with no explanation of what it means for the visitor.
  • ❌ A footer link labelled “Disclaimer” that opens a modal: it can complement the in-context disclosure, not replace it.
  • ❌ A human-like avatar or name with no visible AI label. The Commission explicitly lists misleading human-like representations as insufficient.
  • The AI disclosure is enabled and visible before the first message (in Botshido: the chatbot name and the opening message: see the note at the top).
  • The text says you are interacting with an AI, mentions the operator’s name, and is written in the visitor’s language.
  • The same notice links to your privacy policy.
  • The full AI Disclaimer stays in the widget footer as the detailed version. Do not remove it.
  • Your chatbot’s opening message and persona do not contradict the disclosure (no “I’m Sara from customer care” without the AI label).
  • If a human takes over, the conversation says so.
  • For sensitive sectors: the disclosure remains visible during the conversation, not only at the start.
  • Evidence filed: screenshot of the notice, date, and the exact text used. If someone asks, you show it.
  • Staff trained on what the chatbot can and cannot do (AI literacy, art. 4 AI Act — already applicable).
  • The AI provider’s DPA/terms accepted and filed; your privacy notice names the provider.

Art. 50 sits among the obligations whose breach is sanctioned under art. 99(4) AI Act: up to €15 million or 3% of total worldwide annual turnover, whichever is higher. The realistic risk for a small site is a complaint or a customer audit, not a fine of that size — but the exposure is not theoretical.