Whose face and voice is it? Likeness and voice rights in AI video

There are two compliance questions hiding inside every AI video, and most teams only ask one of them. The first (do we have to disclose that this is AI-generated) gets the attention, because regulation like the EU AI Act made it concrete. The second is quieter and older, and it does not go away because a model generated the footage: did we have the right to use this face and this voice? Disclosure governs what you tell the audience. Rights govern whether you were allowed to create the thing in the first place. A perfectly labelled ad built on a face you had no permission to use is still a problem.

This question used to be simple because the answer was a contract. You hired an actor, signed a release, and the scope of use was spelled out on paper. AI scrambles that clarity in three different ways, namely synthetic avatars, cloned likenesses, and accidental resemblance, and each carries a distinct kind of risk that a single "we used AI" disclosure does nothing to cover.

The right that does not disappear

Most jurisdictions protect a person's control over the commercial use of their own image and voice. In the United States it is the right of publicity; in Germany it is the *Recht am eigenen Bild* under the Kunsturhebergesetz, reinforced by general personality rights. The mechanism varies, but the principle is consistent: you generally cannot use an identifiable person's likeness or voice to sell something without their permission, and a synthetic reproduction does not escape that simply because no camera was involved.

This is the conceptual mistake that gets teams into trouble. They reason that because the output was generated rather than filmed, no real person is depicted, so no release is needed. But the law cares about identifiability, not technique. If a viewer can recognise a specific person, or believe they are seeing one, the protection attaches regardless of how the pixels were made. The generator is just a new way to create a likeness; it is not a way to avoid the rules that govern likenesses.

Three sources of risk, three different cures

The practical work is recognising which situation you are in, because the remedy is different for each.

• Licensed synthetic avatars. Stock-avatar platforms sell access to faces, but those faces usually belong to real actors who licensed their image for specific uses. The risk is not invisible, it is contractual: the licence has a scope. Many forbid political, regulated-industry, or sensitive-claims use, restrict the territories and channels, and cap the term. Using an avatar outside that scope breaches the actor's agreement, and you inherit the liability.
• Cloned likeness and voice. Recreating a specific real person (a founder, an employee, a paid creator) needs explicit, informed, written consent that names what you are cloning, where it will run, and for how long. A general talent contract from the pre-AI era almost never covers it. Voice deserves its own clause: voice is independently protected, and a convincing clone of a recognisable voice is its own violation even if no face appears.
• Accidental resemblance. This is the new one, and the easiest to miss. A model asked for a generic spokesperson can output a face that closely resembles a real, identifiable person you never intended to depict. There is no licence to check because you did not knowingly use anyone. That is exactly why it slips through. It calls for review, not paperwork: a human looking at the output and asking whether this reads as anyone in particular.

A disclosure label answers a question the viewer is asking. A likeness release answers a question the depicted person is entitled to ask. Different people, different rights. Only one of them is satisfied by a line of small print.

The paper trail that protects you

Rights compliance is ultimately an evidence problem. If a claim ever arrives, what matters is whether you can show what you were permitted to do. That means keeping the avatar licence and its scope on file, retaining signed consent for anyone whose likeness or voice you cloned with the specific use named, and logging which assets were used in which campaign so scope can be checked against reality. None of this is exotic; it is the same diligence a careful team always applied to talent, extended to cover faces and voices that were generated rather than filmed.

The reason it gets skipped is that generation feels frictionless. You type a prompt and a person appears, with none of the signing and scoping that a shoot forces on you. That missing friction is precisely the trap: the ease of creating a likeness has raced ahead of the habit of clearing one.

A workable default

You do not need a legal team to stay out of trouble; you need a few defaults applied consistently. Treat every face and voice as belonging to someone until proven otherwise. Use licensed avatars and stay inside the licence scope. Clone a specific person only with written, use-specific consent that covers voice as its own item. Add a resemblance check to your review so an accidental likeness gets caught before it ships. And keep the records, because the value of permission is only as good as your ability to prove you had it.

Disclosure and rights are not competing obligations; they are the two halves of using a synthetic person responsibly. One tells your audience the truth about what they are watching. The other respects the truth that every face and voice, even a generated one, may belong to a real person who never agreed to sell you their image.

• European Parliament and Council, Regulation (EU) 2024/1689 (the AI Act), transparency provisions, 2024.
• German Kunsturhebergesetz (KUG), §§ 22–23, on the right to one's own image (Recht am eigenen Bild).
• WIPO, "Artificial intelligence and the right of publicity / personality rights," 2024.
• IAB, "Synthetic media, likeness licensing, and talent rights for advertisers," 2025.