What Article 50(1) actually requires
Article 50(1) obliges providers to ensure that AI systems intended to interact directly with natural persons are designed and developed so that the persons concerned are informed they are interacting with an AI system. The obligation is disapplied where this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use. There is a further carve-out for systems authorised by law to detect, prevent, investigate or prosecute criminal offences, subject to safeguards, unless the system is available for the public to report an offence.
Three things follow, and each one catches teams out.
It is a design duty. The words are "designed and developed in such a way that". That places the obligation in the product, not in a policy document, and it places it on whoever designs and develops the system rather than on whoever happens to be operating it. The line between those two parties is set out on the provider and deployer roles page, and it moves more often than buyers expect.
It attaches to direct interaction, not to the presence of AI. A recommendation engine that reranks a product grid does not interact directly with anyone in the Article 50(1) sense. A support assistant, a voice agent on a telephone line and an in-product onboarding bot all do.
It is tier-independent. Nothing about Annex III or the high-risk regime is relevant here. A system can be comfortably outside Article 6 and still owe this duty, which is why the classifier runs transparency as its own stage. The full map of who owes which limb sits on the Article 50 transparency page.
The obvious carve-out, and why it is narrow
The exception is the part teams reach for first and the part that will not bear the weight. It is not satisfied because your team knows the assistant is a bot, or because a technically literate user would guess. The reference person is "reasonably well-informed, observant and circumspect", a general standard borrowed from consumer law, and the assessment runs on the circumstances and context of use.
In practice the facts push in two directions. Pushing towards obvious: a widget explicitly labelled as an assistant, a synthetic voice that nobody would mistake for a person, a context where the user arrived by clicking something that said "ask our AI". Pushing away from obvious: a human first name, a photographic avatar, a conversational style that mirrors a support agent, a channel such as WhatsApp or SMS where a human sender is the default expectation, and any handover flow where a human agent and a bot share the same thread.
The practical rule is to decide it per surface rather than per product, because the same underlying system can be obvious in one place and not in another. Write down the reasoning for each surface at the time you make the call. If a market surveillance authority asks later, a contemporaneous note that engages with the standard is worth considerably more than a reconstruction.
What the disclosure has to say
No wording is prescribed. Article 50(1) fixes the outcome and Article 50(5) fixes the manner: the information must be provided clearly and distinguishably at the latest at the time of the first interaction or exposure, and must conform to the applicable accessibility requirements. Within those constraints the drafting is yours.
What works is short, specific and unhedged. "You are chatting with an AI assistant." "This is an automated assistant, not a human agent." "You are speaking to an AI voice assistant. Say 'agent' at any time to reach a person." Naming the escape hatch is not required by Article 50(1), but it is the single thing that most reduces the chance a user feels deceived, which is the harm the article exists to prevent.
What fails is wording that is accurate to a lawyer and misleading to a user. "Powered by AI" describes the technology stack without telling anyone who they are talking to, and sits on plenty of products that do have human agents behind them. "Smart assistant" and "virtual assistant" are marketing terms that predate the current generation of systems and do not reliably convey that no person is present. If the sentence would survive being read aloud to a user who then said "so is this a real person?", it is not doing the job.
Accessibility is a live constraint rather than a footnote. A disclosure carried only in a colour change, only in an icon, or only in placeholder text inside an input field will not reach a screen-reader user at first interaction. Put it in text, in the accessible name of the component, or in the first message of the transcript.
Where it goes, surface by surface
The trigger in Article 50(5) is the first interaction, so the placement question is always "where does this conversation actually begin?" That answer differs by channel.
| Surface | Where the first interaction happens | Where the disclosure belongs |
|---|---|---|
| Web chat widget | The moment the panel opens | Header label plus the opening message, not the launcher tooltip alone |
| Voice agent on a phone line | First audio the caller hears | Spoken in the greeting, before any question is asked |
| WhatsApp, SMS and messaging | The first inbound message | In the first message body, because there is no surrounding interface |
| In-product assistant | Opening the assistant pane | Persistent label on the pane, plus the first response |
| Email auto-responder | The first automated reply | In the visible body, not only in the signature block |
| Human handover flows | Every switch between bot and agent | At the start, and again whenever control returns to the AI |
The handover row is the one most often missed. A thread that starts with a bot, passes to a human and then returns to the bot has three first interactions from the user's point of view, not one, and a disclosure delivered only at the top of the session has stopped being clear and distinguishable by the time control comes back.
Five patterns that fail Article 50(5)
- The terms-of-service disclosure. Accurate, findable, and delivered months before the interaction it is meant to govern. It fails the timing limb.
- The tooltip on the launcher. Visible only on hover, so it never reaches touch users at all, and it disappears the moment the conversation starts.
- The disclosure below the fold. Present in the transcript but scrolled out of view by the time the user reads a response. Clear and distinguishable is judged from where the user is looking.
- The human name with no qualifier.An assistant presented as "Anna from support" with a portrait actively builds the impression the article requires you to dispel.
- The one-time disclosure in a persistent thread. Common in messaging channels, where a conversation resumed weeks later reads as a continuation to the system and as a fresh interaction to the person.
Who owes it, and what to do this week
Article 50(1) binds the provider, which is the party that develops the system and places it on the market or puts it into service under its own name or trademark. If you bought a chat product and deployed it as shipped, the design duty sits with the vendor. That is comfortable until you change something: putting your own name on it, making a substantial modification, or changing the intended purpose can move provider status to you under Article 25, and the duty travels with it. The three triggers are worked through on the provider, deployer and Article 25 flip page.
Buying rather than building does not remove the exposure either, because the deployer chose the surface, the persona and the placement. A vendor cannot design a compliant first interaction for a channel it does not control.
Four things worth doing now, given the duty has already applied:
- List every surface where an AI system talks to a person, including voice lines, messaging channels and internal tools that external people touch. The inventory is usually longer than the product team expects.
- For each one, record the obviousness call and the reasoning behind it. Where the answer is not clearly yes, add the disclosure rather than argue the exception.
- Check the placement against the first-interaction test, and check that it survives a screen reader and a handover.
- Keep the note. Article 50 carries no documentation obligation of its own, but a dated record of how you assessed it is the cheapest evidence you will ever produce, and the page on what has applied since 2 August 2026 sets out the rest of what landed on the same date.
If you have not established which Article 50 limbs reach your system, the free triage classifier runs transparency as its own stage and names the article each answer rests on.