The short version
- Copyright does not protect a voice. It protects specific recordings, so a convincing imitation that copies no recording is not copyright infringement.
- The working claims are the right of publicity and false endorsement, which US courts have applied to deliberate voice imitation since the late 1980s.
- Tennessee's ELVIS Act, effective 1 July 2024, made voice a protected property right covering simulations, and reaches those distributing the tools as well as the output.
- California AB 2602 attacks the contract instead: from 1 January 2025, a digital replica clause is unenforceable unless it describes the intended uses specifically and the performer had counsel or a union.
Your voice is not copyrighted, and that is the problem
Copyright protects original works of authorship fixed in a tangible medium. A recording of you speaking is such a work. The voice itself is not — it is a physical characteristic, like a face or a gait, and no one holds copyright in it.
That distinction decides most cases. A model trained on your recordings that produces new audio you never spoke has not reproduced any particular recording, so the classic copying analysis has nothing to grip. There may be a separate argument about whether ingesting the training recordings was itself infringement, but that claim belongs to whoever owns those masters, which for most working voice artists is a client rather than themselves.
What does bite is the right of publicity: the interest a person has in their own identity not being used to sell or endorse. US courts reached voice through that door long before generative audio existed. In *Midler v. Ford Motor Co.* (9th Cir. 1988), an agency that could not license Bette Midler's record hired a former backing singer to imitate her; the court held that deliberately imitating a distinctive, widely known voice to sell a product was actionable. *Waits v. Frito-Lay* (9th Cir. 1992) applied the same reasoning and added a federal false-endorsement claim under the Lanham Act, on the theory that the imitation falsely suggested the artist had endorsed the product.
What changed when the machines got good
The sound-alike cases assumed friction. Imitating someone meant finding a person who could do it, paying them, and directing a session — expensive, slow, and leaving a trail of witnesses. Cloning removes all of that. A few minutes of reference audio, which most working voice artists have published publicly as a showreel, is now enough.
Three years of law arriving at the same problem from different directions
Jul 2024
ELVIS Act in force
Tennessee adds voice to its right of publicity and defines it to include a simulation of the voice.
Jan 2025
California AB 2602 bites
Labor Code §927 makes vague digital replica clauses unenforceable for performances after this date.
Jun 2026
NO FAKES clears committee
The Senate Judiciary Committee advanced S. 4591 unanimously on 18 June 2026. It is not law.
Now
Patchwork governs
Which state's law applies, and whether your contract gave the right away, decide most real disputes.
The ELVIS Act, and why it is broader than it looks
Tennessee's Ensuring Likeness Voice and Image Security Act took effect on 1 July 2024, replacing a 1984 statute that covered name, photograph and likeness but not voice. Three features make it the most useful model to understand.
- Voice is a property right, and simulation counts. The statute protects both a person's actual voice and a simulation of it, which removes the argument that a synthetic performance is not really the person's voice at all.
- Fame is not required. The protection is not limited to celebrities, which is the gap the *Midler* line of cases left open. Session artists, narrators and podcasters are covered on the same terms as recording stars.
- Liability reaches the tools. Distributing an algorithm, software or service whose primary purpose is producing an unauthorised replica can itself be actionable where the distributor knows the use is unauthorised — a materially wider net than suing whoever published the output.
Enforcement runs through a civil claim, and a violation is also a Class A misdemeanour. Labels and other exclusive licensees can bring claims on an artist's behalf, which matters because individual artists rarely have the resources to litigate alone. Several other states have since added voice or digital-replica provisions to their own publicity statutes on comparable lines; the rules are converging in shape while remaining different in detail.
California went after the contract instead
AB 2602, chaptered on 17 September 2024, took a different route. Rather than expanding what counts as a violation, it made a category of contract clause unenforceable. From 1 January 2025, a provision permitting the use of a digital replica of a performer's voice or likeness in place of work they would otherwise have done in person is unenforceable if the clause lacks a reasonably specific description of the intended uses and the performer was not represented by counsel or a union when they agreed it.
The design is deliberate. Most voice cloning is not piracy; it is authorised, by a paragraph the performer signed without noticing, in a session agreement offered on a take-it-or-leave-it basis. A statute aimed only at unauthorised use would miss almost all of it. Note the limit too: only the offending clause falls away, not the rest of the contract, and only the vague version does — a specific, negotiated replica licence remains perfectly enforceable.
What a replica clause has to contain to survive
The digital replica clause
Voice-over contract template
Sets out the session, the deliverables and the usage grant, with space to state expressly whether synthetic replication is permitted, for what, and on what terms — rather than leaving it to a general rights clause.
Where federal law stands
The NO FAKES Act would create a federal property right in a person's voice and visual likeness against unauthorised digital replicas, enforceable nationwide, with liability extending to platforms that host replicas knowing the depicted person did not consent. The 2026 version pairs it with a DMCA-style notice-and-takedown process: a person can demand removal without suing first, the uploader may file a counter-notice, and the content is restored if no action follows within fourteen days. Knowingly false counter-notifications carry penalties.
If you record your voice for a living
The leverage is entirely at the point of booking. Once a clause granting replica rights is signed and the session is delivered, you are arguing about the meaning of words you agreed to — a far weaker position than declining to agree them.
Read for these before the session, in every agreement
- Any grant of rights in "synthetic", "digital replica", "voice model", "simulated" or "AI-generated" performances — the vocabulary is not standardised, so read the rights clause rather than searching for a word.
- Whether the grant covers creating a model as well as using its output, and whether the model must be deleted at the end of the term.
- Whether the licence is "in perpetuity" or "all media now known or hereafter devised", which for a replica means indefinitely rather than merely for a long time.
- Whether machine learning or training on the recordings is permitted, and by whom — the client, its affiliates, or its vendors.
- What content categories the replica may not be used for, and whether you have a refusal right rather than only a notification.
- Whether replica use triggers a separate fee, and how that fee is calculated per use rather than as a one-off buyout.
- Which state's law governs the agreement, since the protections differ materially between them.
None of this requires refusing synthetic work. Replica licensing is real, growing and legitimately valuable — pickups without a session, localisation into languages you do not speak, accessibility narration at scale. The point is that it is a separate product from a recording session, and it should be priced, bounded and terminable as one, in the same way that usage rights in creator deals are a separate product from the post.
If a clone of your voice is already out there
Four rungs, cheapest first
- Free, hours to days
Platform report
Every major platform now has a synthetic-media or impersonation policy separate from copyright. Report under that policy, not as a copyright claim, which will be rejected.
- Low, days
Letter to the publisher
Most business users of a cloned voice did not commission the clone and will stop immediately. Name the right of publicity and the false endorsement risk.
- Moderate, weeks
Contract claim against the client
If the voice came from a session you delivered, the real question is what your agreement permitted. This is usually the strongest and cheapest claim to run.
- High, months
Publicity and Lanham Act claim
A state right-of-publicity action, and false endorsement where the use implies you approved the product. Viable where the use is commercial and the voice identifiable.
Preserve evidence first: the audio, the URL, the date and the placement. Synthetic media is removed and re-uploaded constantly.
Where the material is defamatory or sexual rather than commercial, the analysis moves elsewhere entirely — several states now criminalise non-consensual synthetic intimate imagery, and the defamation route may matter more than publicity rights. Those are different claims with different evidence, and worth separating early.
If you are the one using synthetic voice
The exposure sits with whoever publishes, not with the tool. Three practices remove nearly all of it: licence the voice from a person by name with written consent covering your specific use; keep the consent and the provenance of the model with the asset, because in two years nobody will remember which vendor produced it; and do not use a synthetic voice that resembles an identifiable real person, even where no recording of theirs was used, since imitation is exactly what the sound-alike cases were about.
The through-line in all of this is that voice cloning is not primarily a copyright problem or a technology problem. It is a consent problem, and consent is recorded in contracts. The statutes arriving now are mostly attempts to fix contracts that were signed before anyone knew what they were granting — which is a good reason to read the ones you are signing today.
Sources
General information, not legal advice. This guide explains how these documents and rules generally work. Law varies by jurisdiction and changes, and none of it is applied to your circumstances here. For anything consequential, consult a licensed attorney where you are.
Frequently asked
Is it illegal to clone someone's voice with AI?
Cloning by itself is not a general offence in most of the United States. What creates liability is the use: putting a recognisable person's voice behind a product, an endorsement or a statement they never made. Tennessee's ELVIS Act goes further and treats the unauthorised simulation itself as a violation of a property right, and reaches those distributing tools for that purpose.
Can I copyright my voice to stop this?
No. Copyright protects fixed works of authorship, and a voice is a personal characteristic rather than a work. You can hold copyright in particular recordings, and that is worth having, but it does not stop a model producing new audio that never copies any of them. The protections that do apply are publicity, endorsement and contract law.
Does the ELVIS Act protect me if I do not live in Tennessee?
Not directly. Right of publicity is state law and its reach depends on where the person lives, where the use occurred and the choice-of-law rules that apply. A Tennessee resident has a strong claim there; someone elsewhere depends on their own state's statute or common law, which may or may not cover voice and may require the voice to be distinctive and known.
I signed a contract with an AI clause. Can I get out of it?
Possibly, if California law governs and the performance was after 1 January 2025 — AB 2602 makes a replica clause unenforceable where it lacks a reasonably specific description of uses and you had no counsel or union representation. Outside that, you are on ordinary contract ground: ambiguity, unconscionability and scope arguments. Read the actual clause before assuming either way.
What should a fair AI clause look like in a voice contract?
It should name the specific productions and media the replica may appear in, set a term and a territory, treat creating the model and using its output as separate consents, require deletion of the model when the term ends, pay per use rather than as a single buyout, and give the performer a right to refuse categories of content. Anything shorter is a buyout dressed as a formality.