Article50.io
Question · Art. 50(1)

Do AI Phone and Voice Agents Need a Disclosure Under Article 50?

· Updated · 12 min read

Yes. Article 50(1) of the EU AI Act covers AI systems intended to interact directly with people, and an AI agent that answers or makes phone calls does exactly that. Callers must be told they are talking to an AI, at the latest at the start of their first interaction. On a call, that means a spoken statement at the beginning of the conversation, not a notice on your website.

Key point: An AI phone or voice agent is covered by Article 50(1) in the same way as a text chatbot. The disclosure has to reach the caller in a clear and distinguishable way at the latest at the first interaction (Art. 50(5)), which on a phone line means saying it at the start of the call. If the voice agent also sits on your website as a widget, a scanner such as Article50.io can check whether the page tells visitors it's AI.

Why voice agents are squarely in scope

The test in Article 50(1) is functional: is the AI system intended to interact directly with natural persons? It doesn't matter whether the interaction happens by text, voice or video, or whether the system is high-risk. AI receptionists, booking lines, support hotlines, outbound sales callers and in-app voice assistants all interact directly with people.

The only way out is the "obvious" exception. Article 50(1) doesn't require a disclosure where it is obvious "from the point of view of a natural person who is reasonably well-informed, observant and circumspect" that they are dealing with an AI. Voice agents are usually built to make that less obvious. They use natural-sounding voices, pause and say "mm-hm", and often introduce themselves with a human name. For most voice agents, rely on the disclosure, not the exception.

What the disclosure should sound like

The Commission's Article 50 guidelines, adopted on 5 November 2026 as C(2026) 7702 final and non-binding (paragraph 5), give a spoken statement at the beginning of a voice session as an example of a disclosure (paragraphs 36 and 37). A generic reference to an "assistant", used on its own, is not enough (paragraph 38). The 20 July 2026 text (C(2026) 5054 final) was a draft submitted for approval and said the same. This is not a new November obligation. Guideline paragraph numbers on this page refer to Commission Guidelines C/2026/6005 (OJ C, 12.11.2026), adopted as C(2026) 7702 final on 5 Nov 2026, non-binding (point 5).

In practice:

  • Say it first. Put the disclosure in the opening line of every call, before you ask for a name, account number or reason for calling. Keep that opening spoken disclosure.
  • On a longer call, repeat it as appropriate. Paragraph 37 gives periodic reminders in longer interactions as an example, in particular after an interruption or when the AI system's role changes. The same idea covers a session that resumes. There is no fixed interval.
  • Say "AI" plainly. For example: "Hi, you've reached Acme. I'm an AI assistant and I can help with bookings, or put you through to a person." The words "virtual assistant" or "assistant" alone are not enough (paragraph 38).
  • Don't bury it after hold music or a long menu. The caller should hear it before they start talking to the AI.
  • Keep it accessible. Article 50(5) requires the information to meet applicable accessibility requirements. Speak at a normal pace, and don't rely on a text message or web page as the only disclosure.

For written disclosure wording you can adapt, see the AI chatbot disclaimer template.

Inbound and outbound calls

  • Inbound (AI receptionists, support lines). The caller chose to ring you, but they may not expect an AI. Disclose in the greeting.
  • Outbound (AI sales or reminder calls). The person didn't start the conversation, so the disclosure matters even more. State that the caller is an AI, and which business it's calling for, at the start.
  • Transfers. If the call moves from the AI to a person, or back, make the switch clear.
  • Voicemail. Whether an AI-generated voicemail counts as an interaction isn't settled. The cautious approach is to include the disclosure in the message anyway.

Who owns the duty: you or the voice platform?

Article 50(1) is a provider duty. Who the provider is depends on how your voice agent was built.

  • You use an off-the-shelf AI phone product (an AI receptionist or call-centre platform). The platform that builds the system and places it on the market under its own name is most likely the provider. You are most likely a deployer. But you usually write the greeting script, pick the voice and name the agent, so in practice you control whether callers hear a disclosure.
  • You or an agency build your own voice agent on speech and language model APIs, and run it under your own name. You are likely the provider of that voice agent, and the Article 50(1) duty is then directly yours.

The same reasoning is set out in more detail, for text chatbots, in do I need Article 50 compliance if I only use a chatbot?

Cloned and synthetic voices

Voice agents raise two further Article 50 questions that a text chatbot doesn't.

Machine-readable marking (Art. 50(2)). Providers of AI systems that generate synthetic audio must mark the output in a machine-readable format so it can be detected as artificially generated. That is the voice platform's duty, not yours, if you use someone else's. For generative systems already on the market before 2 August 2026, the transition for the Article 50(2) machine-readable marking duty ends on 2 December 2026 (Article 111(4) on the AI Act Service Desk); see the 2 December 2026 deadline.

Deep fakes (Art. 50(4)). A deep fake is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic (Art. 3(60)). A cloned voice of a real person, such as your founder or a well-known presenter, used in recorded audio can be a deep fake, and deployers must disclose deep fakes under Article 50(4).

Whether a live AI call in a cloned real voice also falls under Article 50(4) is not settled. A practical approach: if your voice agent uses a clone of a real person's voice, say so in the opening disclosure as well, for example "This is an AI assistant using [name]'s voice." For recorded audio, see deepfake disclosure and Article 50 for podcasters.

Do advertising voiceovers need an AI label?

A recorded advertising voiceover usually does not need an Article 50(1) disclosure. Article 50(1) covers AI systems intended to interact directly with people, and a recorded voiceover that does not exchange with the listener is not that kind of interaction (Commission guidelines, C(2026) 7702 final, paragraph 30; non-binding). This is about recorded production audio, not a phone or voice agent.

Two other duties can still arise.

Machine-readable marking (Article 50(2)) sits with the provider of the tool that generated the audio. If you only buy a synthetic voice from a platform, that marking duty is usually the platform's. It applies to synthetic audio generally, not only to clones of famous people. You can still ask the vendor how the file is marked, and check that your edit or upload path does not strip it.

A visible or audible disclosure (Article 50(4)) sits with you, as the deployer, only if the audio is a deep fake. The Commission's Article 50 guidelines, adopted 5 November 2026 (C(2026) 7702 final; non-binding), set a cumulative test in paragraph 113: the audio must resemble an existing subject and would falsely appear authentic or truthful. A clone of a real person, such as a presenter or a founder, used as if it were their voice, is the clearest case — the guidelines' examples after paragraph 116 include voice cloning of a podcast's regular presenters. A generic realistic voice from a production tool is not automatically in scope, and "it does not sound like a famous person" is not the whole test. The same examples box treats AI voice replication for fictional characters (for example in audiobooks, games or animation), with no deception as to the identity of the narrators, as not constituting a deep fake. False appearance is judged in context (paragraph 114). Advertising that is exclusively commercial, and recognisable as such, is excluded from the lighter artistic regime; advertisements are assessed case by case (paragraph 122).

If Article 50(4) applies, paragraph 117 says the disclosure should be perceivable without a special tool — an audible line in the ad, or a clear on-screen label, not only a mark in the file. This is general information, not legal advice.

Other rules still apply

Article 50 sits alongside other laws rather than replacing them (Art. 50(6)). Phone lines often raise separate obligations under data protection, call recording, telemarketing and consumer protection rules, which vary by country. An Article 50 disclosure doesn't satisfy those, and they don't replace it.

Checklist for an AI voice agent

  • The opening line of every inbound call says the caller is speaking to an AI, in plain words.
  • Outbound AI calls say they're from an AI, and on whose behalf, at the start.
  • The disclosure comes before the AI asks for personal details.
  • Transfers between the AI and a person are announced.
  • The agent doesn't use a human name without also saying it's an AI.
  • If the voice is a clone of a real person, that is disclosed too.
  • You've asked your voice platform how it handles Article 50(2) marking of synthetic audio.
  • You've called your own number, as a customer would, and heard the disclosure.

Which scanner checks voice agents for Article 50?

Article50.io's scan checks web pages for EU AI Act Article 50: it loads your page in a real browser and recognises the ElevenLabs voice agent, Vapi voice agent and Telnyx AI agent widgets by their scripts and embed markup. It reads the text visitors see, plus the text the widget shows as it first loads, inside chat-like frames among the first 25 frames on the page and in open chat-like shadow roots on those elements, within a short time limit. A frame is read only when its labels or address match that chat-like filter. ElevenLabs, Vapi and Telnyx are custom elements, so their first-load text is read from an open shadow root on the element when the browser exposes one. Their own sites are not on the frame-address list, so a disclosure that exists only inside a cross-origin frame from those sites may not be read. Then it checks whether that text tells visitors they're dealing with an AI. It recognises that wording in seven languages: English, German, French, Spanish, Italian, Dutch and Portuguese.

What it can't see:

  • Voice widgets from other vendors: only those three vendors' voice widgets are recognised.
  • How a disclosure inside a widget served as a frame from another site is presented (size, contrast, position): the scan reads the text but can't measure how it is shown.
  • Text in closed shadow roots or drawn on a canvas.
  • Widgets that load only after cookie consent. The scan sees the page as a first-time visitor does before accepting cookies, so a voice widget that loads only after consent won't be detected. Check it yourself.
  • Machine-readable marking under Article 50(2) of synthetic voice output, which is the voice vendor's job as provider.
  • Anything said on a call, or audio of any kind. The scan does not hear calls. Nothing is clicked or opened, so a disclosure spoken only once a conversation begins isn't heard, and reminders later in a call can't be checked.
  • Phone-line agents with no widget on the page. Use the checklist above for those.
  • Pages other than the one address you scan, and anything behind a login.

Key point: the scan can tell you whether a voice widget's visible text says it's AI. It can't hear what the agent says, so test your own phone line too.

Each scan is capped at 45 seconds. A clean result means the scan found no obvious issue on that page; it isn't a compliance clearance or legal advice. See what the website scan can and cannot check for the full method and vendor list. The duty itself is in Article 50 on the Commission's AI Act Service Desk.

Frequently asked questions

Does an AI phone receptionist need to say it's an AI?

Yes, unless it's obvious. Article 50(1) covers AI systems that interact directly with people, including by voice, and Article 50(5) requires the notice by the first interaction: on a call, the greeting. The Commission's Article 50 guidelines (C(2026) 7702 final, paragraphs 36–37; non-binding) give a spoken opening statement as an example, and reminders on longer calls as appropriate.

When during a call should the AI disclosure be made?

At the start, before the caller is asked for anything. Article 50(5) requires the information at the latest at the time of the first interaction, and in a clear and distinguishable manner. A disclosure after the caller has already given personal details, or only on your website, is unlikely to be enough.

Is a realistic AI voice "obviously" an AI?

Usually not. The exception in Article 50(1) only applies where a reasonably well-informed, observant and circumspect person would already know they are dealing with an AI. The Commission's guidelines (C(2026) 7702 final, paragraph 45) read it restrictively and list vocal tone among the factors, so natural-sounding voices rarely benefit from it.

Who is responsible if I use an AI receptionist platform?

The platform that builds the AI system and places it on the market under its own name is most likely the provider, which is who Article 50(1) addresses. You are most likely a deployer, but you usually control the greeting script, so the disclosure callers hear is in your hands. This is a legal judgement that depends on your setup.

Does using a cloned voice of a real person change anything?

It can. Audio that resembles an existing person and would falsely appear authentic is a deep fake under Article 3(60), and deployers must disclose deep fakes under Article 50(4). Whether a live AI call counts as deep fake content is not settled, so the cautious approach is to disclose the cloned voice in the opening line too.

Which scanner checks AI voice agents for EU AI Act Article 50?

Article50.io scans one web page in a real browser and recognises voice-agent widgets from ElevenLabs, Vapi and Telnyx by their scripts and embed markup, then checks whether visible text tells visitors it's AI. It can't hear calls or audio, recognises those three voice vendors only, and doesn't click anything. A clean result isn't a compliance clearance.

Does an advertising voiceover made with a synthetic voice need an AI label?

Usually not under Article 50(1), which covers AI systems that interact directly with people, not recorded audio. Article 50(4) applies only to a deep fake under the cumulative test: resemblance of an existing subject and a false appearance of authenticity (guidelines C(2026) 7702 final, paragraph 113; non-binding), judged in context (paragraph 114). A generic voice is not automatically in scope.

This article is general information, not legal advice.

Check your site automatically

Article50.io is an automated scanner for the Article 50 transparency rules of the EU AI Act. It scans a website for potential Article 50 transparency gaps and generates a report of what the scan found and what it could not check, with remediation guidance, implementation instructions and disclosure wording you can paste in.

The free scan checks one public page, including ElevenLabs, Vapi and Telnyx voice widgets, and shows your most severe finding. Each scan is capped at 45 seconds, with no signup. The full report, €499, lists every finding with the Article it relates to, what couldn't be checked, and wording you can paste in.

More from the blog

Automated technical guidance, not legal advice. Citations refer to Regulation (EU) 2024/1689.