If your chatbot talks to anyone in the EU, you have a hard deadline: August 2, 2026 — eight days from now. On that date, Article 50 of the EU AI Act (Regulation (EU) 2024/1689) starts applying, and one of its plainest rules is that people must be told when they're talking to a machine. The good news: the minimal correct fix is genuinely small, and you can ship it this week.
Short answer up front: add a clear, visible "you're talking to an AI" notice at the start of every conversation — a banner, a first message, or a spoken line for voice — unless it's already obvious. No specific wording is required; clarity and accessibility are.
What Article 50 actually requires#
Article 50(1) puts the duty on providers of AI systems "intended to interact directly with natural persons." You must design and build the system so those people are informed that they are interacting with an AI system — clearly, and at the latest by the time of the first interaction. The European Commission's Article 50 guidance adds that the information must be clear, distinguishable, and conform to accessibility requirements.
There's one carve-out: the obligation doesn't apply where it is obvious to a reasonably well-informed, observant and circumspect person that they're dealing with AI. Treat that exception as narrow. A bot on a bare-bones support page might qualify; a fluent, human-named voice assistant almost never will.
The rule isn't "watermark your prose." It's "don't let a human think a human is on the other end." Disclosure is the whole ask.
Separately, Article 50(2) covers generative systems — outputs like synthetic text, images, audio and video must be marked in a machine-readable format as artificially generated. That's a different obligation from the "I am an AI" conversational notice, and the two are easy to conflate. For a text chatbot, your first job is the 50(1) disclosure.
Where it bites#
The scope is broader than "the chat widget on our marketing site."
- Voice agents. The "obvious" exception is weakest here, because a natural-sounding voice is precisely what fools people. Say it out loud at the start of the call: "Hi, I'm an automated assistant."
- Embedded and white-label widgets. If you provide the AI system, the duty is yours as provider — even when it's dropped into someone else's site under their brand. Don't assume the deployer handles it.
- Agent-to-agent. When your agent is reached by another agent rather than a person, a visible banner does nothing. Expose the disclosure as structured data (a response field, a metadata header) so the calling system can surface it if a human is ultimately on the other end.
This is the same regulatory turn we've been tracking all year — China regulated the persona itself, and it's threaded through this week's founders' wire. The direction is unmistakable: tell people when it's a bot.
The minimal correct implementation, step by step#
You are not building a compliance platform. You are adding one honest sentence in the right places.
- Inventory your surfaces. List every place your AI talks to a person: web widget, in-app chat, WhatsApp/SMS, phone/voice, and any embedded copies.
- Add a first-contact disclosure on each surface — visible before or at the first exchange, not buried in a privacy policy.
- Make it accessible. Real text (not an image), sufficient contrast, screen-reader friendly. For voice, speak it.
- Handle agent-to-agent by returning the disclosure as a field, not just UI.
- Keep a dated record of what you shipped and where, in case a market-surveillance authority asks.
A plain, illustrative banner — conceptual, not a real API:
<!-- Visible at the top of the chat, before the first user turn -->
<div class="ai-disclosure" role="note">
You're chatting with an AI assistant, not a human.
</div>For a system-prompt / first-message pattern:
[first assistant message, every new session]
Hi — just so you know, I'm an AI assistant. How can I help?And for an agent-to-agent response, an illustrative field:
{ "actor": "ai", "ai_disclosure": "This response was generated by an AI system." }That's the core of it. No mandated phrasing means you can keep your brand voice — you just can't be coy about the machine.
How it lines up with China and the US#
You're probably not shipping to only one jurisdiction, so build once to the strictest bar that applies:
- China — the Measures for Labeling AI-Generated Synthetic Content, effective September 1, 2025, require both explicit (visible) and implicit (metadata/watermark) labels, with a visible AI symbol specifically called out for chatbot-type output. That's stricter than the EU's conversational notice because it reaches into the content itself.
- California — SB 243, effective January 1, 2026, targets "companion" chatbots and requires a clear and conspicuous notice that the bot is AI where a reasonable person could be misled, plus safety protocols and a private right of action (statutory damages of $1,000 per violation). It's narrower in scope but sharper in enforcement.
A single, unambiguous "I am an AI" disclosure is the common denominator. Where China wants watermarks or California wants safety protocols, layer those on — but the conversational notice satisfies the shared core.
What to do this week#
Ship the disclosure now; refine later. Today, list your surfaces. Tomorrow, add a first-contact notice to each — text, voice, and any embedded copies. Before August 2, add the agent-to-agent field and write down what you did and when. It's one honest sentence, repeated in the right places — and at up to EUR 15,000,000 or 3% of turnover for getting it wrong, it's the cheapest compliance work on your list.
