Skip to content

How to tell people
they are talking to an AI

From 2 August 2026, an AI that talks with people must make clear it is an AI, at the latest at the first interaction, unless that is obvious. Put the disclosure in the first message or the first seconds of a call. Deepfakes and AI text on public matters need visible labels.

What does Article 50 require?

Article 50 of the EU AI Act sets transparency duties that apply from 2 August 2026. The Digital Omnibus amendment of July 2026 did not move this date. The duties split between providers, who build the system, and deployers, who use it:

Article 50 duties by paragraph
ParagraphWhoDuty
50(1)ProviderDesign an AI system that interacts directly with people so they are told they are interacting with an AI, unless this is obvious.
50(2)ProviderMark synthetic audio, image, video and text output in a machine-readable format so it can be detected as AI-generated.
50(3)DeployerTell people when they are exposed to emotion recognition or biometric categorisation.
50(4)DeployerDisclose deepfakes, and AI-generated text published to inform the public on matters of public interest.
50(5)BothGive the information in a clear and distinguishable way, at the latest at the first interaction or exposure, and meet accessibility requirements.

Paragraph 1 is written for providers. In practice, the business that configures the agent controls what it says. If you have a custom agent built and put it into service under your own name, you can be its provider. If you buy a product, check that you cannot switch the disclosure off or rename the bot as a person. Either way, make the disclosure part of your launch checklist.

What the Commission guidance adds

The Commission published guidelines on Article 50 on 20 July 2026, and a voluntary Code of Practice on marking and labelling AI-generated content on 10 June 2026. The guidelines are not binding, but regulators will read the law through them. The points that matter most for a business agent:

  • Four tests. The duty applies when the system is an AI system, is designed for a genuine two-way exchange, communicates directly with the person, and the other party is a natural person. Background tools and machine-to-machine traffic are outside it.
  • Timing. Tell people from the start of the first interaction.
  • Form. A first-turn greeting, a label near the chat box or, on the phone, a spoken statement at the start of the call (“This is an AI-powered assistant”). A tone or sound alone is not enough.
  • What does not count on its own. Disclosure only in terms and conditions, a vague word such as “assistant”, a human-like avatar, a general line such as “Services on this website use AI”, or “this system uses LLMs”.
  • Who the agent acts for. An AI agent that makes bookings or handles correspondence must disclose that it is an AI and the person on whose behalf it acts.
  • Honest answers. The system must disclose its nature whenever someone asks about it.
  • Reminders. One clear notice is likely to be enough in most cases. Periodic reminders are likely to be needed for vulnerable people and in health, financial, insurance, legal and complaints contexts, and in long voice calls after an interruption.

The obvious exception is narrow. The guidelines read it restrictively. A consumer helpdesk chatbot is not obvious, because users may think a person wrote the replies. An internal assistant used only by trained staff can be obvious.

Example wording and placement by channel

These examples are short on purpose. Adapt the business name, the tasks and the escape route. Keep the AI statement in the first message or the first sentence, before any question.

Disclosure wording and placement
ChannelExample wordingWhere and when
WhatsApp agent“Hi, this is the AI assistant for [business name]. I can book, move or cancel appointments. Type PERSON at any time to reach our team.”First message of every new conversation. Repeat it when a conversation restarts after a long gap. Also state it in the business profile description.
Phone or voice agent“Hello, you’ve reached [business name]. I’m an automated AI assistant. Calls are recorded. Say ‘reception’ at any time to speak to a person.”The first sentence of the call, before any question. Speak it clearly, at normal speed, with no music over it.
Website chatHeader: “AI assistant”. First message: “I’m an AI assistant for [business name]. I can answer questions and book appointments. Ask for a person at any time.”In the chat window header and the first message. Do not use a human name and photo for the bot.
Automated email repliesSignature line: “This reply was written and sent by an AI assistant. Reply HUMAN to reach our team.”In every email the agent sends without human review.
AI images or video of real people or placesVisible label: “AI-generated image” or the EU icon for AI content.On the image or video itself, where people see it first, not only in the caption.
AI text on public mattersNote: “This article was generated with AI.”At the top of the text, unless a person reviewed it with editorial responsibility.

For WhatsApp agents, Meta’s own rules apply too. See our guide to WhatsApp message templates and the WhatsApp agents page.

Deepfakes and AI-generated content

A deepfake, in the Commission’s reading, is AI-generated or manipulated content that resembles real people, objects, places or events and would falsely appear authentic to a person. If your marketing uses an AI video of a real staff member, or a realistic AI photo of your shop, label it. The deployer cannot rely on hidden watermarks from the provider. The label must be visible or audible, no later than first exposure.

For clearly artistic, satirical or fictional work, the label can be placed so that it does not spoil the work, for example in the end credits. For AI-generated text on matters of public interest, the label is not needed if a person reviewed the text and someone holds editorial responsibility. A spelling or grammar check is not a review.

  • Content made and published before 2 August 2026 does not need a label after the event, although the Commission encourages it.
  • Providers of generative AI systems already on the market before 2 August 2026 have until 2 December 2026 to meet the machine-readable marking duty in Article 50(2).

Common mistakes we see

  1. A human name and photo for the bot. “Hi, I’m Sarah” with a stock photo suggests a person. Use a name only together with the words “AI assistant”.
  2. Disclosure after the first question. “What’s your date of birth?” before the AI statement is too late.
  3. Disclosure only in the privacy policy. A link in the footer is not a clear statement at the first interaction.
  4. No route to a person. The Act does not require one in every case, but people who cannot reach a person complain. Our human-in-the-loop guide shows how to hand over.
  5. A dishonest answer to “Are you a real person?” Test this question before launch. The agent must say it is an AI.

In Ireland, Coimisiún na Meán supervises Article 50 for online platforms, hosting services and broadcasters. For other businesses, see who regulates AI in Ireland. GDPR transparency is a separate duty: your privacy notice must also explain the AI processing.

Common questions

Do I need to disclose AI if the customer will probably guess?

Usually yes. The exception for interactions where AI use is obvious is read narrowly. The Commission says a consumer helpdesk chatbot does not meet it, because users may think a person wrote the replies. A short statement in the first message costs nothing and removes the question.

Can our AI assistant have a name?

Yes, if it is clearly presented as an AI. “I’m Aoife, the AI assistant for [business name]” is clear. “Hi, I’m Aoife from reception” is not. Avoid a human photo as the avatar, and make sure the agent answers truthfully when someone asks if it is a person.

Does Article 50 apply to emails our staff write with AI help?

Not when a person reviews and sends the email. Article 50(1) is about AI systems that interact directly with people. An email drafted by AI and sent by a member of staff is a human message. An agent that replies automatically without review is different, and its emails should say they come from an AI.

When did the AI disclosure rules start?

The Article 50 transparency duties apply from 2 August 2026. The July 2026 Digital Omnibus amendment kept that date. It only gave providers of generative AI systems already on the market until 2 December 2026 for the machine-readable marking duty in Article 50(2).

Sources

Bring us the task.

Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.

Start a project.

Two people on a sofa with laptops, planning a project beside a whiteboard of sticky notes

Book a call or send us a message. We read everything.

tash@neu.ie
Working with teams in Ireland and worldwide.

Get in touch

Free 30-minute AI review.