AI voices in advertising: what is allowed and what is not

An agency clones the voice of a presenter for a campaign, the client is happy, and three weeks later the presenter finds out from a friend who heard the radio. That is the situation nobody wants, and it is almost always avoidable with an hour of preparation. What follows is a practical overview, not legal advice. Rules differ per country and per situation, so run your own case past a lawyer.

Start with the voice actor

A voice is closely tied to a person. In practice that means you need clear, written permission before you make a model of someone, and permission to record is not the same as permission to clone. Put the specifics on paper:

What the model may be used for. Which brand, which products, which media, which countries.
For how long. A fixed term with an explicit end date works better than open ended wording.
What is off limits. Most actors want to exclude political messages, gambling, alcohol or anything that puts words about a competitor in their mouth.
What it pays. Pay for use, not only for the recording day. A model that runs for two years is worth more than one session.
What happens at the end. Agree that the model and the recordings get deleted, and say who confirms that this happened.

Treat the recordings as personal data, because that is what they usually are. Store them with access control, do not pass them to a third party without the actor knowing, and be able to say where they sit.

Be straight with your audience

Audiences accept synthetic voices better than most marketers expect, as long as they are not being tricked. The problem is never that a voice was generated. The problem is the impression that a real named person said something they never said.

The safe line: never suggest a specific identifiable person endorsed your product unless they really did. If the voice is unmistakably a real person, say it is synthetic. If the voice belongs to no one in particular, most people neither notice nor mind. Advertising codes in many markets already require that ads are not misleading, and a voice that implies a false endorsement fits that description regardless of the technology behind it.

The EU AI Act in outline

The European AI Act entered into force in August 2024 and applies in stages. The bans on certain uses came first, in February 2025, along with the duty to make sure staff working with these systems know what they are doing. Obligations for general purpose models followed in August 2025, and the transparency rules relevant to most marketing teams apply from 2 August 2026.

The part that matters here is straightforward in principle: content that is artificially generated or manipulated has to be recognisable as such, and material that convincingly imitates a real person needs to be disclosed. How you disclose it depends on the medium and can be modest. A line in the credits, a note in the description, a spoken mention. The point is that a reasonable listener can find out. Details are still being worked out in guidance, so check the current state before you build a campaign around a borderline case.

Known voices and voices of the dead

Two situations deserve extra caution. Imitating a celebrity, even without cloning their actual recordings, can suggest an endorsement that does not exist and has led to disputes in several markets. And using the voice of someone who has died needs agreement with the family or estate, plus a serious conversation about whether it is the right thing to do at all. Legally possible and publicly acceptable are not the same question.

Put it in your own process

Keep a short file per voice model: who gave permission, what for, until when, who signed, and where the recordings live. Add one line to your production checklist asking whether the voice in this asset is synthetic and whether disclosure is needed. Two minutes of work per campaign, and it is the difference between a calm conversation and a nasty one.

A useful test before anything goes out: would you be comfortable if the voice actor heard this ad on the radio tomorrow without warning? If the answer is no, the paperwork is not the problem. The plan is.