Right of publicity and AI voices: a separate right from copyright

Two rights live in the same recording

Copyright and publicity protect different things in the same audio file. Copyright covers the recording as a work: who may copy it, distribute it, adapt it, train on it. Publicity law covers the person: their name, their likeness, and in a growing number of states their voice as well.

A buyer who takes a licence to the recording has cleared the first and not the second. The seller, if it is a production company or a studio, can grant rights in its own work and cannot grant the performer's interest in how they are used in trade. Those are two different assets owned by two different parties, and the deal documents often describe only one of them.

This is the most common structural gap in voice data transactions, and it is invisible in a contract that grants "all rights" without saying which rights it is talking about.

Publicity law in the US is state law, and the variation between states is wide enough to change the answer to a simple question:

• Some states name voice expressly in their statute. California, New York and Tennessee are the ones cited most often, and each reaches voice in its own way and to its own extent.
• Other states protect voice through case law rather than a statute, which means the boundary is described by past decisions rather than by a text you can read.
• Post-mortem rights differ sharply. Some states recognise none at all; others recognise one lasting decades, which decides whether archival material needs an estate consent in addition to the performer's.
• Damages and remedies differ too, which is why the same corpus carries different exposure depending on where the speakers were.

The rule that governs is where the performer was

When a corpus is drawn from performers in more than one state, every one of those regimes travels with it into a contract that may be signed somewhere else entirely, because the injury is treated as occurring where the person is protected rather than where the company operates. That is why the speaker's location belongs in the manifest as a field, not in a legal memo as an afterthought.

The practical consequences: a corpus from one state may need a document that a corpus from another does not, and a supplier that records location only for logistics cannot tell you which. Where a performer cannot be located at all — archival material, an agency that has closed — the honest position is that the corpus cannot be cleared for publicity, and it should be priced or excluded accordingly.

Where the imitation line sits

The classic disputes involve advertisers who hired a sound-alike after failing to book the original performer. Courts have found liability where the imitation was deliberate and the result was recognisable as a particular person, particularly in a commercial context where it implied an endorsement.

The line that emerges is about identification and context rather than similarity in the abstract:

• A voice that a listener identifies as a specific person, used to sell something or to imply that person's involvement, is the core case and the one most likely to end badly.
• A generic voice that resembles many speakers is not, which is why synthetic voices are often designed to sit deliberately away from any identifiable person.
• A performer imitating a public figure as part of a performance has been treated differently from an advertiser doing the same thing — a distinction that maps badly onto a model, which has no performance and no audience of its own.

Why the trained model is the hard case

A recording captures a fixed performance. A model trained on a performer's sessions captures a capacity, and it is the capacity that the law has not settled.

Three consequences follow:

• The harm is not one use but an unbounded set. The argument that a use was brief, incidental or part of a larger work weakens when the asset can generate new utterances indefinitely.
• The model can be pointed at a purpose nobody contemplated at the session, which is where the scope language in the consent has to carry the weight.
• Disclosure becomes a publicity question as well as a consumer-protection one. An audience that does not know a voice is synthetic may be misled, and that engages the performer's interest separately from any copyright question.

What a publicity-clean corpus looks like

The paperwork to look for, per performer:

• A consent that names the person and grants use in trade in the performer's own name, not only a production-side release signed by a company.
• A scope that lists the products, the term, the territory, and the categories of content that are off limits.
• An explicit statement about synthetic voice: whether a replica may be made, and whether the performer's name may be associated with a model.
• For archival material, a chain to whoever holds the post-mortem right, or a written statement that the material has been excluded from the delivery.
• For minors, a guardian consent that names the permitted use rather than consenting generally.

The test to apply before signature

Take one performer from the corpus at random and ask the supplier to produce the document that lets you use their voice in a synthetic product. If it takes a week and produces a document that names training and replica use, the corpus is licensable. If it produces a recording release that mentions only the production, the buyer is holding the wrong paper.

Legislatures have been moving toward treating digital replicas as a subject of their own, which is a signal that existing statutes are being read as incomplete. That direction argues for collecting the broadest consent available now, since a corpus assembled under the next regime will be easier to license than one assembled under the last.

This is an operational overview rather than legal advice. Publicity law is state-specific and fact-specific, and applying it to a particular corpus is work for counsel.