Open a few small-business websites and look at the chat bubble. A friendly name, a smiling stock photo, "Hi, I'm Sam, how can I help?" Sam is a language model. Since 2 August 2026, in the EU, that greeting is a legal question and not just a branding choice.
That is the whole story of AI chatbot disclosure in one example. The rule is short, it already applies, and most of the confusion around it comes from people mixing it up with the parts of the AI Act that really were postponed. We build an open-source AI support agent, so we have a stake here. We'll say so where it matters. We are also not lawyers, and nothing below is legal advice.
What Article 50 actually says
Article 50(1) of the EU AI Act says providers must design AI systems that interact directly with people so that those people are informed they are talking to an AI. There is one carve-out: if it is obvious to a reasonably well-informed person from the context, you don't need to say it. Paragraph 5 adds the timing. The information has to arrive "in a clear and distinguishable manner" at the latest at the first interaction, and it has to meet applicable accessibility requirements. You can read the full text of Article 50 in a few minutes.
Notice what is missing. There is no list of approved wording. No required badge. No minimum font size.
So the law tells you the outcome (the person knows) and leaves the method to you. That is both the good news and the trap. A disclosure buried in a terms page does not meet "at the latest at the time of the first interaction". A one-line grey caption in a 9px font probably does not meet "clear and distinguishable" either, though nobody has tested that in front of a regulator yet.
What the Digital Omnibus delayed, and what it didn't
Here is where the muddle starts. You may have read that the EU "delayed the AI Act." That is true of one part and false of another.
According to Usercentrics' summary of the Digital Omnibus deal, the high-risk system deadline moved from 2 August 2026 to 2 December 2027 for Annex III systems. Separately, Article 50(2), the machine-readable marking of AI-generated content, got a four-month grace period to 2 December 2026, but only for generative systems already on the market before 2 August. The same summary is explicit that the other transparency obligations, including the Article 50(1) duty to tell people they are talking to an AI, were left untouched.
A second legal write-up, Praxikon's analysis of the Article 50 deadline, reaches the same conclusion: Article 50(1) was not postponed and has applied since 2 August. It also notes that the Code of Practice on transparency is still being finalised and is voluntary, but that following it will be the practical evidence a marking approach is adequate. That code is about marking generated content. It does not change the chatbot duty.
So if your support bot is live and talking to EU visitors, the clock is not about to start. It started.
Who owes the disclosure: you or your vendor?
Read the text again. The obligation in 50(1) sits with the provider, the party who designs and develops the system. A business that drops a third-party chat widget on its site is usually the deployer, not the provider.
That sounds like a way out. It isn't a very comfortable one. Praxikon's reading is that using a third-party model does not by itself make you the provider, and that a deployer relying on the vendor should confirm the vendor has built the notice in. It also flags that heavy branding or modification can shift your role. If you take an open-source system, rename it, change what it does, and ship it under your own name, a lawyer could reasonably ask whether you became the provider.
Two practical consequences.
If you buy a hosted chatbot, ask the vendor, in writing, where the AI notice lives and whether you can turn it off. A vendor that lets you delete it is handing you the liability.
If you self-host, you are closer to the provider end of the spectrum. We went through this in our piece on self-hosting an AI chatbot: running the software on your own server buys you control over where the data lives. It does not exempt you from Article 50.
The fine, and why small teams shouldn't relax
Article 99(4) puts breaches of the Article 50 transparency obligations in the middle tier: administrative fines up to €15,000,000 or 3% of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises, Article 99(6) flips it: the fine is capped at whichever of the two figures is lower.
For a five-person company with €800,000 in turnover, 3% is €24,000. That is a ceiling, not a prediction, and a regulator would weigh the circumstances. But it is not nothing, and it is far above what a one-sentence welcome message costs.
We have not found a published fine under Article 50 so far. Treat that as "early", not as "safe". National authorities are still building their enforcement habits, and the first cases tend to be the clearest ones. A bot that says "Hi, I'm Sam" and denies being a machine when asked would be about as clear as it gets.
The US contrast: Utah asks only when you ask
Not every jurisdiction went the EU way, and the differences are instructive. Utah passed an AI disclosure law for consumer interactions, and its 2025 amendments narrowed it. According to Ashurst's write-up of the Utah changes, a business has to disclose when a consumer makes a "clear and unambiguous request" to know whether they are dealing with AI. For regulated occupations the duty applies to high-risk interactions: collecting sensitive health, financial or biometric data, or giving the kind of advice people rely on for significant decisions.
There is also a safe harbour. A provider avoids enforcement if the AI itself clearly and conspicuously says it is non-human at the outset and throughout the interaction. The law runs until 1 July 2027 under SB 332, so it is a moving target.
Put the two side by side. Utah: tell people if they ask. EU: tell people before they have to ask. And notice that Utah's safe harbour is, in effect, the EU rule. A disclosure at the start that stays visible is the one design that satisfies both. You don't need to track every state to pick it.
What a disclosure that works looks like
We have opinions here, and they are blunt. Hiding the bot is a bad bet, legally and commercially. We covered the commercial side in AI customer service frustration is real, and the data says it isn't about the AI: customers get angry at bots that trap them and stonewall them, not at bots that are labelled. Disclosure costs you very little. Getting caught pretending costs a lot.
A version that holds up:
- Say it in the first message. "I'm an AI assistant for Harbor Dental. I can answer questions from our docs, and I can bring in a person." One sentence, first thing the visitor reads.
- Keep a label in the widget header. The first message scrolls away. A persistent "AI assistant" tag doesn't.
- Answer the question honestly every time. If a visitor types "are you a real person?", the bot says no. Put that in the instructions. We wrote about how instructions behave, and where they fail, in a chatbot system prompt is a job description, not a law, so test it instead of trusting it.
- Don't give it a human face. A human name is a judgment call. A stock photo of a smiling person is much harder to defend as "obvious from context".
- Mark the handoff. When a person takes over, say so: "You're now talking to Priya on our team." The reverse matters just as much, and it is the part of human handoff people forget.
- Check accessibility. The accessibility language in paragraph 5 means a disclosure that a screen reader skips is a weak disclosure. Put it in real text, not in an image.
That list isn't long. Nothing on it needs engineering. Mostly it needs someone to decide, once, what the first sentence says.
Where ChatterMate fits, and where it doesn't
In ChatterMate the agent's name and welcome message are configurable, so the disclosure is something you write and own, not a toggle we hide. It also means we can't make you compliant. We can make it easy to say the right thing in the first message, and we would rather you did.
If you're evaluating any support bot, vendor or open source, ask three questions. Can I see the exact text shown on first contact? Can the bot be configured to deny being an AI? Where do the conversation logs live if a regulator or a customer asks what was said? The answers tell you more than any feature table.
The rest of what we've written on grounding, handoff and testing is in the ChatterMate blog hub.
Written by the ChatterMate team. We build an open-source AI support agent. This post summarises public sources as of 5 October 2026 and is not legal advice. If your exposure is real, ask a lawyer who handles AI regulation.
ChatterMate is open source and free to start, with 50 free AI replies a month, and you can self-host it. Try it at chattermate.chat.



