Telling people they are talking to an AI sounds simple until you have to write the actual sentence. This guide gives you practical AI disclosure notice examples for a support chatbot, a voice assistant, and an AI writing helper, plus guidance on placement, timing, tone, and accessibility. It also draws the line around what "obvious from context" really covers under Article 50(1) of the EU AI Act.
This is not legal advice. The examples below are drafting starting points to review with your own counsel, not approved legal text. Sufficiency is determined by the AI Act and official guidance, not by this article.
What Article 50(1) actually requires
Article 50(1) is the "you are dealing with a machine" rule. Where an AI system is intended to interact directly with people, the provider must design and build it so that the affected people are informed they are interacting with an AI system — unless that is obvious to a reasonably well-informed, observant and circumspect person, taking account of the circumstances and context.
A few things follow from the wording:
- The obligation is about being informed, not about consent. You are giving notice, not asking permission.
- It applies to systems intended to interact with people — chatbots, voice agents, AI-driven avatars, in-app assistants.
- The duty sits with the provider (who builds or places the system on the market), though deployers who present the system to their own users usually surface the notice in practice. Assess this per product, because the same company can ship one feature that clearly needs a notice and another that does not.
These transparency obligations start to apply on 2 August 2026. If you are shipping AI interaction features into the EU market — or your output is used in the EU — the rule can reach you even if your team sits outside the EU. Using a third-party model API such as OpenAI or Anthropic does not remove the duty attached to the system you provide or deploy.
What "obvious from context" does and does not cover
The exception is narrower than teams hope. "Obvious" is judged from the perspective of a reasonably observant person in the actual context — not from the perspective of your product team, who already know it is an AI.
Reasonable arguments that context makes it obvious:
- A feature literally branded and named as an AI (for example, a panel titled "AI Assistant") that the user opened deliberately.
- A clearly labelled AI image generator where the entire purpose of the tool is generation.
Weak or risky assumptions that it is "obvious":
- "Our brand is known for AI, so everyone knows." Recognition among your existing fans is not the reasonable-observer test.
- A chat widget with a human-sounding name and a stock avatar photo. That tends to imply a person, not an AI.
- Voice assistants with natural, human-like speech. Realistic voices push *toward* needing a clear notice, not away from it.
- A support handoff where an AI answers first and a human joins later, with no signal about which is which.
When in doubt, disclose. A short, plain notice costs you little; a missed one is hard to defend as "obvious." Treat the exception as something you document a reasoned case for with counsel, not a default you lean on.
Where and when to place the notice
Getting the words right matters less than putting them where people will actually notice them, at the moment they start interacting.
First message / first contact. The notice should be present before or at the very start of the interaction, not buried in a terms-of-service link. For chat, that usually means the opening message or a header on the chat window.
Persistent label. A single opening line is easy to scroll past. A persistent, visible label on the interface — a header, a badge next to the message stream, or a name that includes "AI" — keeps the notice available throughout a long session.
Before a voice call. For voice, disclose audibly at the start of the call, before substantive conversation, because there is no screen to carry a persistent badge. If the user can miss the opening line because they were not yet listening, consider a brief repeat or an option to hear it again.
On resumption. If a session is picked up days later, or a human hands back to the bot, re-assert who the user is now talking to.
Timing and tone
Timing rule of thumb: the person should know before they invest anything — before they type a real question, share personal details, or act on advice. A notice that appears only after the conversation is over does not meet the point of the rule.
On tone: the notice must be clear and not deceptive, but it does not have to be cold or bureaucratic. Avoid two failure modes:
- Over-softening that hides the fact. "Hi, I'm Alex from the team!" with no indication Alex is software is misleading.
- Over-legalising that no one reads. A dense paragraph of defined terms buries the one fact people need.
Aim for a plain sentence a hurried person understands on first read.
Accessibility
A disclosure only works if everyone can perceive it.
- Do not rely on an icon or colour alone (for example, a small robot glyph). Pair any icon with text.
- Screen readers: make sure the notice is real text in the reading order, with appropriate labelling, not baked into an image or hidden as a decoration.
- Voice / audio: the spoken disclosure must be intelligible at normal playback, and a text equivalent should exist where a transcript or caption is offered.
- Contrast and size: a persistent label set in tiny low-contrast grey is technically present and practically invisible. Meet your normal accessibility contrast targets.
AI disclosure notice examples to adapt
The wording below is illustrative. Adapt it to your product, then review with counsel before shipping. These are examples, not approved legal text, and they do not by themselves make anything "compliant."
Support chatbot
Opening message, shown with a persistent header label alongside it:
> AI Assistant — You are chatting with an automated AI assistant, not a human. It can help with common questions and can connect you to a person at any time. Please do not share sensitive personal or payment details here.
A shorter persistent badge that stays visible above the message stream:
> You are talking to an AI. Type "agent" to reach a human.
For the moment a human takes over, so the change is explicit:
> You are now connected to Priya, a member of our support team. You are no longer chatting with the AI assistant.
Voice assistant
Spoken at the very start of the call, before any substantive exchange:
> Hello. Before we start, please note this call is handled by an automated AI voice assistant, not a person. You can ask to speak to a human at any time, or say "repeat that" to hear this again.
A shorter variant for a returning caller who has already heard the full notice:
> Hi again. You're speaking with our AI voice assistant. Say "agent" for a human.
If the assistant also appears in an app, pair the audio notice with an on-screen label such as "AI voice assistant — automated" so the disclosure is not audio-only.
AI writing helper
A persistent label on the feature panel, shown wherever the tool is used:
> AI writing assistant. Suggestions below are generated by AI and may be inaccurate or biased. Review and edit before you use or publish them.
A first-run notice the user sees the first time they open the feature:
> This assistant uses AI to draft and rewrite text. It is a tool to help you, not a substitute for your own judgement. You are responsible for what you publish.
Note that a writing helper can also touch other parts of Article 50. If it produces synthetic text, image, audio or video, Article 50(2) requires providers to mark that output as artificially generated in a machine-readable format, and Article 50(4) adds deployer duties for deepfakes and for AI-generated text published to inform the public on matters of public interest (with narrow exceptions for editorial control and artistic or satirical work). The interaction notice under 50(1) does not discharge those separate marking duties.
A quick worked example
Say you run a SaaS help centre. A user clicks a chat bubble labelled only "Help." An AI answers first.
- First contact: the opening message names it as an AI assistant and offers a human handoff.
- Persistent label: a header badge reads "AI Assistant" for the whole session.
- Timing: the notice appears before the user types their real question.
- Handoff: when a human joins, a line states the change.
- Accessibility: the badge is real, readable text, not just an icon.
- Documentation: you keep a record of the exact wording, where it appears, and the date it went live, so you can show your reasoning later.
That last point matters. Being able to evidence what you disclose, where, and when is part of taking the obligation seriously — and far easier to assemble as you go than to reconstruct afterwards.
A note on machine-readable marking
If your feature generates content, remember that the machine-readable marking under Article 50(2) is advisory metadata — HTML attributes, JSON-LD, or header metadata — not signed provenance. It is not C2PA certification and should not be treated as sufficient on its own. There is a voluntary EU Code of Practice on AI-generated content that providers can look to. Provider-side machine-readable marking has a limited transition that can extend to 2 December 2026 for eligible systems already on the market before the general application date. For a plain-language walkthrough of the whole article, the Article 50 explainer and the Commission's GPAI guidance are useful reference points.
Next step
Work through your AI features one product at a time: where do people interact directly with a machine, is a notice present before they invest anything, and can you show where and when it appears? If you want a structured starting point, run DiscloseKit's free readiness check to map which Article 50 obligations touch each of your products — then take the specifics to your counsel for a final view.