AI chatbot disclosure examples: 15 clear, weak and risky patterns
Article 50(1) of the EU AI Act says an AI system that interacts with people must make that clear, and Article 50(5) says the information has to arrive at the latest at the time of the first interaction, in a clear and distinguishable manner. Neither sentence tells you what to put in the chat window. This page does: fifteen real-world wordings, sorted into the three buckets a reviewer actually uses — clear, weak, and risky — with the reason each one lands where it does.
The sort order is not arbitrary. When we swept 1,088 EU-facing sites eight days after Article 50 began to apply, our automated visitor could read the first-interaction text on 174 of them. A disclosure was detected on 19. On 82 — 47% of the readable surfaces — the wording was inconclusive: something in the greeting gestured at automation without ever saying it. That middle bucket is the biggest one, it is where most teams currently sit, and it is entirely made of the "weak" patterns below.
What "clear and distinguishable" is measured against#
Four things decide whether a disclosure works, and none of them is how much effort went into it.
- Timing
- The disclosure has to be present at or before the first interaction. Text that appears after the visitor types is late.
- Location
- It has to be where the interaction happens. The Commission's guidelines of 20 July 2026 are explicit that a line in the terms and conditions is normally not enough.
- Wording
- It has to state that the counterpart is an AI system, in terms a reasonably well-informed person would not have to infer.
- Persistence
- A notice that scrolls out of the conversation after two messages was there at first interaction, but nothing about the transcript afterwards shows it.
The narrow exemption in Article 50(1) — where it is obvious to a reasonably well-informed, observant and circumspect person — is doing much less work than teams hope. A fluent natural-language support bot is the paradigm case of something that is not obvious, which is why naming it "Assistant" and adding a robot avatar does not reach the bar.
Five clear patterns#
These state the fact, in the conversation, before anyone types.
1. The plain opener.
"Hi — you're chatting with an AI assistant. I can help with orders, returns and delivery. Ask for a human any time and I'll pass you over."
Says what it is, at first interaction, in the surface the visitor is looking at. The handover sentence is not required by Article 50, but it removes the main reason teams soften the wording.
2. The persistent header label.
An "AI assistant" label in the chat window's title bar, present for the whole conversation, next to a greeting that also says it. The header alone is weaker than it looks — see pattern 9 — but paired with an opening line it survives a reviewer scrolling back through a transcript.
3. The pre-chat notice on the launcher itself.
Text on the bubble before it is opened: "Chat with our AI assistant." This discloses before the first interaction rather than at it, which is the safest side of the line, and it is the only pattern that still works if the panel fails to open.
4. The named-bot-plus-nature line.
"You're speaking with Ada, our AI support assistant."
A persona name is fine as long as the sentence containing it also carries the nature of the counterpart. A persona is fine. The failure is when it replaces the disclosure.
5. The AI-nature answer to the direct question.
Visitor: "Am I talking to a person?" Bot: "No — I'm an AI assistant. I can hand you to a human colleague if you'd prefer."
Worth saying plainly: this one is not sufficient on its own. Article 50(1) requires disclosure at the first interaction. Answering when asked is already too late. An assistant that owns up only when asked has not met the duty. It belongs on the clear list because it is a necessary complement — a bot that denies being AI when asked directly undoes every other pattern on this page — and because it is the single easiest thing to test yourself.
Five weak patterns#
Each of these is something. None of them is a statement that the counterpart is an AI system, and each is the kind of wording that lands in the "could not verify" bucket rather than the "detected" one.
6. "I'm your virtual assistant."
"Virtual" describes where it is, not what it is. Human agents work virtually. This is the single most common near-miss we see, and the reason the inconclusive bucket is 47% rather than 5%.
7. "You're chatting with our automated helper."
Better than "virtual" — "automated" at least excludes a person. It still leaves a reader guessing between a decision-tree macro and an AI system, which matters because Article 50(1) attaches to AI systems specifically. Where the thing behind the widget is an AI system, say so.
8. "Powered by AI" in small print under the composer.
Describes the vendor's technology stack, not the identity of the counterpart, and sits outside the conversation in the chrome most people never read. A reviewer reads this as branding.
9. A header label with no opening line.
"AI Assistant" in the title bar and a greeting that says only "Hi! How can I help?" Whether the label registers before the first exchange is a question about one visitor's eye path, not a property of the page — and it is exactly the kind of question you do not want decided after the fact from a screenshot.
10. The disclosure inside a cookie banner or the terms link.
The Commission's guidelines name this one directly. Consent text and T&Cs are not the first-interaction surface, and putting the AI disclosure there is the pattern the guidelines were written to rule out.
Five risky patterns#
These do not merely fail to disclose; they push in the opposite direction.
11. A human first name and a stock human photo.
"Sarah" with a headshot, no AI mention anywhere. The avatar is an affirmative representation that the counterpart is a person. This is the worst pattern on the page, and it is common.
12. "One of our agents will be right with you" — from the bot.
Implies a queue for a human that the visitor is already past. Reviewers treat a misleading statement about the counterpart as materially worse than silence.
13. The disclosure that only appears on desktop.
Compact mobile layouts routinely drop the header label, the pre-chat notice, or both, while the desktop view a team screenshots for its file keeps them. The duty follows the visitor, and most of your EU visitors are on a phone. Check the two views separately — see how to audit a chatbot disclosure.
14. The disclosure that appears only after a form gate.
Name-and-email first, disclosure after. The visitor has already interacted, which puts the notice on the wrong side of "no later than the first interaction."
15. The bot that denies being AI.
Asked directly, it answers "Yes, I'm a real person" or deflects. Usually nobody configured this — it is a system prompt written for warmth, generalising badly. It is worth testing precisely because no one intends it, so no one finds it.
What our own measurement says about the middle bucket#
On the 174 first-interaction surfaces our sweep could read on 10 August 2026:
| Outcome | Sites | Share of readable |
|---|---|---|
| Disclosure detected | 19 | 11% (7–16) |
| No disclosure detected | 73 | 42% (35–49) |
| Could not verify (wording inconclusive) | 82 | 47% (40–55) |
Read that third row carefully, because the second and third rows do not add together. "Could not verify" is not a finding of no disclosure: nearly half of the readable surfaces carried wording that an automated check could not resolve either way. Those sites are not silent; they are ambiguous, and ambiguity is a wording problem with a one-line fix. Full method, exclusions and confidence intervals are in The State of AI Disclosure 2026.
Two cautions on those numbers, stated the way we state them everywhere: 174 readable surfaces is a small denominator carved out of 794 confirmed widgets, and our sweep egressed from the United States, so a widget that behaves differently for EU visitors would look different from inside the EU. This is a snapshot of what an external visitor can observe. It is not a verdict on anyone.
Rewriting a weak pattern into a clear one#
The fix is almost always one sentence, and almost always in a field you already control:
- "I'm your virtual assistant" → "You're chatting with an AI assistant."
- "Powered by AI" in the footer → move the same claim into the greeting message.
- Header label only → keep the label, add the greeting line. Both, not either.
- Desktop-only → set the greeting in the widget's message settings, which render in both layouts, rather than in a desktop-only theme block.
Where that field lives differs per product. The per-widget configuration guides show the exact setting for Intercom, Zendesk, HubSpot and 17 others.
One thing a rewrite does not do by itself: create a record. If you change the greeting today, nothing on your site shows what it said last month — which is the subject of proving a disclosure was live on a given date.
Common questions
Is "virtual assistant" enough to satisfy Article 50(1)?
It is the most common near-miss and we would not rely on it. "Virtual" describes where something works, not what it is — human agents work virtually too. Article 50(1) is about informing people they are interacting with an AI system, so the safest wording says exactly that: "You're chatting with an AI assistant." On the first-interaction surfaces our 2026 sweep could read, 47% carried wording of this kind that an automated check could not resolve either way.
Does an "AI Assistant" label in the chat header count on its own?
It is real reinforcement but weak as the only disclosure. Whether a header label registers before the first exchange depends on one visitor's eye path, which is not a property you can evidence afterwards. Use a line in the first bot message as the baseline and keep the header label as well — both, not either.
Can we use a human name and photo for our AI chatbot?
A persona name is fine as long as the sentence containing it also states the nature of the counterpart: "You're speaking with Ada, our AI support assistant." A human first name plus a stock human headshot and no AI mention anywhere is the riskiest pattern on this page, because the avatar is an affirmative representation that the counterpart is a person.
Our bot admits it is AI when asked. Is that sufficient?
No. Article 50(1) requires disclosure no later than the first interaction, not disclosure on request, so an assistant that owns up only when questioned has not met the duty. Test it anyway: a bot that denies being AI or deflects is a materially worse problem, and it usually comes from a system prompt written for warmth rather than from a deliberate decision.
Does the "obvious" exception in Article 50(1) cover a support bot?
The exception applies where it is already obvious to a reasonably well-informed, observant and circumspect person. A fluent natural-language support bot is the paradigm case of something that is not obvious, which is why naming it "Assistant" and adding a robot icon does not reach the bar. The Commission's guidelines of 20 July 2026 say a bot name or a robot icon alone is normally not enough.
Sources and further reading
- Article 50 — Transparency obligations (EU AI Act)
- Commission guidelines on transparency obligations, 20 July 2026
- The State of AI Disclosure 2026 — full method and exclusions
Last updated September 2026. Informational only, not legal advice: this page describes what the text of the EU AI Act says and what an external check can observe, not whether any particular site complies. Corrections welcome at hello@disclosureproof.com.