🎙️ AI Voice Actors: What Consent Now Legally Requires

I spent a career on the other side of the glass from voice actors, and the thing that strikes me about the current argument is how little of it is about audio quality. The models sound fine. Whether you are allowed to use what they produce is the question that actually decides budgets, and it is the one almost nobody checks before generating.

By the numbers#

• SAG-AFTRA's ratified 2026 TV/Theatrical Agreement includes 12 AI provisions covering consent, digital replicas and royalties
• Synthetic voice is a separate consent category — it is not covered by a standard performance agreement
• Over 16,000 signatures on the June 2026 open letter urging Congress to pass the NO FAKES Act
• Disclosure requirements for AI-generated people in ads and listings expanded through 2026

The consent gap that catches producers#

Here is the mistake I watch people make. They hire a voice actor, pay properly, get a clean session, and then treat those recordings as raw material — training a model on them, or generating additional lines later when the script changes. It feels reasonable. You paid for the voice, after all.

Under the current framework it is not reasonable, because synthetic use is carved out as its own thing needing its own written consent. The recording you commissioned and a model trained on that recording are two separate grants, and the second one does not come free with the first. Studios wanting a synthetic version of a performer's voice have to ask for it explicitly, in writing.

For anyone producing outside union agreements, the practical read is the same even where the letter does not bind you: the industry standard is converging on explicit per-use consent, and "the old contract did not say I couldn't" is a defence nobody wants to test.

So what: if a human recorded it, get separate written permission before any model touches it.

What is actually being displaced#

The honest version is narrower than either side's framing. The work under real pressure is the functional tier: e-learning narration, phone systems, product explainers, short corporate reads. High volume, low interpretation, tight budgets. That work is genuinely moving, and pretending otherwise does the affected performers no favours.

What is not moving nearly as fast is anything requiring direction. A performance that changes because a director asked for it differently, character work with sustained identity, comedic timing, or an emotional read that has to land a specific way — these remain hard, and the tools are not close in the way a demo reel suggests. Our dubbing tools breakdown covers the adjacent case where a real performance is translated rather than replaced, which is a meaningfully different proposition.

Voice is not copyright, and that keeps surprising people#

A voice sits outside copyright almost entirely. What protects it is right of publicity and performers' rights, which protect a person's identity rather than a work. That distinction is why clearing a licence never clears a voice, exactly as it does not for AI cover songs.

It is also why the NO FAKES Act matters as a direction of travel: it would establish a federal property right in voice and likeness, which is currently patchy state-by-state. Until something like it passes, you are operating under an inconsistent patchwork, and the conservative reading is the one that survives a challenge.

So what: treat a voice as belonging to a person, not as an asset you licensed.

What to do if you generate voices commercially#

Use a fully synthetic voice not modelled on an identifiable person, or a licensed voice from a provider that can document consent, or your own. Those three are defensible. Cloning a real person because the tool made it easy is the one that is not, and the tools rarely ask.

Keep records. Which voice, which provider, which plan tier, which date, and what the licence said at the time. Terms in this category change often enough that "it was allowed when I generated it" is a claim you will want to be able to evidence. The same discipline applies to the model side covered in our ElevenLabs guide, where plan tier determines commercial rights.

And build disclosure in from the start. Retrofitting a label across a published campaign is far more expensive than including it in the first cut.

What this means if you hire talent#

Three practical changes to how you contract, and none of them are expensive to adopt now versus untangling later.

Separate the grants explicitly. One clause covering the recording, a distinct one covering any synthetic use, with the second left unsigned unless you genuinely need it. Bundling them to save a conversation is what creates the dispute.

Scope the synthetic permission. Consent for a specific campaign, a specific duration and a specific territory is defensible. Perpetual worldwide rights to a person voice, obtained cheaply in a session agreement, is the shape of deal that gets challenged and that performers are increasingly advised to refuse.

Keep the training data question separate again. Permission to generate lines in someone voice is not the same as permission to retain and train on their recordings indefinitely. If your provider trains on what you upload, that is a third grant you may not have.

For smaller productions none of this requires a lawyer on retainer. It requires a two-page rider and the discipline to use it, which is considerably cheaper than discovering mid-campaign that your lead voice never agreed to be synthesised.

So what: three separate permissions, scoped and dated, before anything is generated.

Where I land#

The measurement I care about has moved. It used to be whether a synthetic read survives a listener on headphones, and increasingly it does. Now it is whether you can produce, on request, the written consent covering that voice. A studio can fix a mediocre read in a pickup session. It cannot retroactively obtain permission it never got.

None of this is legal advice, and anything with real commercial exposure deserves an actual entertainment lawyer rather than a production blog. But the operating habit is simple enough to adopt today: before you generate a voice, be able to say whose it is and who agreed.