AI Voices in Advertising:
What the EU AI Act Now Requires of You
On 2 August 2026, Article 50 of the EU AI Act became applicable. Since then the rule is straightforward: if your ad, your podcast, or your e-learning module uses an AI voice that could pass for a human, you have to say so. And not in the fine print.
The short version
Synthetic voices that sound deceptively real need a disclosure. The duty applies not only to tool providers but to you as the deployer. Fines reach up to EUR 15 million or 3 percent of worldwide annual turnover. And even with perfect labelling, GDPR and personality rights are still waiting for you.
Contents
1. What has applied since 2 August 2026
The EU AI Act is not one date but a staged rollout. The transparency rules in Article 50 are the stage that marketing, communications, and HR now actually feel. The core is pleasantly simple: if you publish artificially generated or manipulated audio that closely resembles a real person, you must disclose that the material is artificial.
What is interesting is what the text does not say. There is no prescribed wording, no mandatory icon, no minimum volume for the notice. The legislator requires the information to arrive clearly, distinguishably, and in time. Everything else is craft. And that is exactly where it gets messy in practice, because “it is in there somewhere” is not an implementation.
2. Who is affected, and who thinks they are not
The most common misunderstanding we hear in briefings: “We are not building AI, we are only using it.” That is precisely the point. The AI Act distinguishes providers from deployers, and you become a deployer the moment you use a voice tool for something you publish. In concrete terms this hits:
- Marketing and advertising, from radio spots to social video to product demos
- Corporate communications, explainer videos, annual reports, internal townhall recordings
- HR and learning, onboarding modules, compliance training, instructional video
- Service and sales, phone greetings, hold messages, voicebots, outbound campaigns
- Product, in-app announcements, navigation voices, voice assistants
The uncomfortable part: you are also responsible when an agency or freelancer slipped the AI voice in without telling you. If your logo is at the end of the spot, it is your spot. Which is a good reason to add the question to your standard briefing rather than playing detective later.
Practical tip: Make it a standing question on every audio delivery whether synthetic voices are included, and get the answer in writing. One line in the sign-off document costs nothing and saves you an unpleasant reconstruction exercise later.
3. The three legal layers nobody separates
This is where most of the confusion lives, because three entirely different rulebooks apply simultaneously and all of them sound like they are about “AI voices”. They do not replace each other. They stack.
| Layer | The question it asks | What you need |
|---|---|---|
| EU AI Act | Does the listener know the voice is artificial? | Clear, timely disclosure |
| GDPR | Was the voice recording lawful to process at all? | A legal basis, usually explicit consent |
| Personality rights | May this specific voice sound like this? | A contract with the person, for this exact purpose |
The most expensive mistake is treating the disclosure as a free pass. It is not. If you clone a well-known narrator's voice without permission and dutifully print “AI-generated” underneath, you have satisfied the AI Act and still violated their personality rights. Honestly labelled theft is still theft.
One more detail that tends to get lost: if voice is used for identification or authentication, for example in customer service, the stricter rules for biometric data come into play as well. At that point it stops being a marketing topic and becomes one for your data protection team.
4. What a disclosure actually sounds like
The good news: it sounds far less awkward than everyone fears. We have tested several variants with clients over the past months, and the finding is consistent. A confident disclosure lands better than a sheepish one.
Works well
- A short spoken pre-roll: “This voice was generated with AI.” Two seconds, done.
- Combining the audio notice with an on-screen caption, so it also reads when muted
- One line in the podcast description plus a mention in the episode intro
- For voicebots: the first spoken line, not the end of the call
Does not work
- A note only in the privacy policy or legal imprint
- The disclosure at the end of a 45-minute training module
- An asterisk at the edge of frame that scales away on mobile
- “AI-enhanced audio” as the wording. Nobody understands it, which is precisely the problem
From the production side: teams that build the disclosure into the creative early get it integrated cleanly, tonally. Teams that bolt it on three days before air date can hear the seam. That is a craft argument rather than a legal one, but it decides whether your label sounds like confidence or like a warning sticker.
Not sure what is inside your audio files?
We review your live audio assets and tell you where synthetic voices are in play, where a disclosure is missing, and where a real recording would actually be cheaper than the AI workflow.
Request an audio check →5. Checklist for your next audio production
Six questions to run before go-live. If you cannot answer one of them, that is your answer.
- Is there a synthetic voice in the material? Partially counts. One line counts. A placeholder that accidentally stayed in counts.
- Does it sound human? If you hesitate, the answer is yes.
- Is the disclosure there on first listen? Not merely present somewhere, but placed where it actually lands.
- Is there consent covering the training data? In writing, not as a memory of a phone call.
- Does the talent contract cover this exact use? A radio buyout does not cover AI cloning for voicebots.
- Is it documented who decided? When it matters, what counts is not the truth but the paper trail.
6. When the real voice is the cheaper decision
Now the part where we as a voice-over agency are biased, so let us do the maths openly. AI voices are unbeatable at high volume with a short shelf life: 300 product variants, weekly price updates, twelve languages for internal training. The workflow wins there, and we build those workflows for clients ourselves.
It looks different as soon as one of these is true:
- The asset has a long run. A brand film with a five-year life justifies a real recording, if only for legal clarity over those years.
- The brand sells trust. Banking, insurance, healthcare. Here the disclosure is an active signal, and not always a flattering one.
- It is about emotion, not information. A breath in the right place is still hard to synthesise.
- Compliance effort exceeds the talent fee. This happens more often than people expect. A professional narrator plus clean rights clearance is frequently cheaper than three internal review rounds about whether the labelling is sufficient.
That is exactly why we keep both on the shelf: voice casting in over 40 languages with remote direction and studio recording, and documented AI workflows for the cases where volume beats nuance. The right answer is almost never ideological. It follows from run time, reach, and risk profile.
The disclaimer that belongs here: this article is a professional assessment from a production perspective, not legal advice. For a binding evaluation of your specific case, talk to a lawyer specialising in IT and media law. We are happy to work alongside your counsel and supply the production-side documentation.
Research, source selection, and editorial responsibility: Rahim Erbil. We used AI while drafting and researching. Every sentence was checked by hand before it landed here. Made with Herzblut and AI.
Related reading:
The same name on every line.
Concept, composition, sound design, mix, master: at Campera Media your project passes through one pair of hands, from the first briefing to the last second of the fade-out.
Meet Rahim →Common questions about
AI voices and the AI Act.
The transparency duty targets synthetic audio that appears deceptively real, meaning it could be mistaken for a genuine person. An obviously artificial robot tone is a different case from a perfectly cloned narrator. Because the line keeps moving as systems improve, the practical advice is simple: disclose as soon as a listener might hesitate.
It must be clear, unambiguous, and easily accessible, and it has to reach the listener on first exposure. One sentence at the start of the audio is often enough. A footnote on page four of your legal notice is not.
No. Disclosure and personality rights are two separate things. Without documented consent from that person for that specific use, you have a problem no notice text will solve.
Legally it is considerably simpler, because you work with a talent contract and a clearly defined buyout. Economically it depends on volume. For 200 product variants in twelve languages a properly licensed AI workflow usually wins. For a brand film, the real voice almost always does.
Ready for sound
that impacts?
Tell us about your project, free and without obligation. We will get back to you within 24 hours with concrete ideas for your audio identity.
