What should the disclosure actually say?

Article 50(1) prescribes an outcome, not a magic sentence: people must be informed they're interacting with AI, at the latest at first interaction, clearly and distinguishably. But the Commission's draft guidelines include example wording, and the patterns below follow it. Three surfaces matter.

Persistent badge (chat header or beside the input field): short and unambiguous. English: "AI assistant" or "Powered by AI." German: "KI-Assistent" / "Mit KI betrieben." French: "Assistant IA" / "Optimisé par l'IA." The Code of Practice's harmonised icon convention ("AI" / "KI" / "IA" in a fixed position) works alongside this.

First-message notice, the workhorse: "Hi! I'm {Name}, an AI assistant. You are chatting with an artificial intelligence system, not a human. I can help with {topics}. Type 'agent' at any time to reach a human colleague." The elements doing legal work: naming the system as AI explicitly ("artificial intelligence," not just a cute bot name), the "not a human" contrast, and delivery before or with the first AI output. The human-handoff line is good practice rather than a requirement.

Voice opening, spoken before any substantive content: "This call is handled by an AI-powered virtual assistant." For long calls, repeat mid-call; if a human takes over, saying so keeps the caller accurately informed.

What disqualifies otherwise-fine wording: burying it in terms & conditions (expressly insufficient in the draft guidelines), machine-readable-only signals, an audio earcon with no words, or colour-only cues that screen readers can't convey. Make the badge a real, labelled element. And match your audience: if children or low-literacy users are foreseeably in it, simplify. The free generator assembles these patterns for your bot's name and languages; wording in German and French should get a native legal read before you rely on it commercially.