Short answer. Three separate questions get collapsed into one, and they have different answers. Ownership: in the United States, purely AI-generated material is not copyrightable — the Copyright Office's January 2025 report holds that human authorship is required and that prompts alone do not supply it, so what you own is your human contribution, including creative selection, arrangement and modification. Permission to depict a person: governed by right-of-publicity law, which several states have expanded specifically to cover AI voice and likeness — Tennessee's ELVIS Act, effective 1 July 2024, made voice a protected property right. Permission to use a performance: governed by contract and, in union production, by collective agreements requiring separate written consent for digital replicas and synthetic voice. Clear all three, or the asset is not clear.
Key takeaways
- A model licence permitting commercial use of outputs answers only the ownership question; it says nothing about a performer's or depicted person's consent.
- The U.S. Copyright Office's January 2025 report treats human authorship as required for copyright and holds that prompts alone do not supply it.
- Tennessee's ELVIS Act, effective 1 July 2024, made voice a protected property right for every individual and reaches tools built to replicate it, not only the final output.
- The EU AI Act's Article 50 disclosure duty for deepfakes, applying from 2 August 2026, is separate from consent — obtaining one does not discharge the other.
- Union agreements increasingly require separate, specific, written consent for a digital replica or synthetic voice, distinct from consent to being recorded.
What three separate questions get asked here?
Ownership, permission to depict a person, and permission to use a performance are governed by three different bodies of law, and clearing one does not clear the others. A right of publicity is a person's legal control over commercial use of their identity — name, image, voice and likeness. A digital replica is a synthetic reproduction of a specific real person's appearance or voice, generated rather than recorded.
| Question | Governed by | Evidence you need on file |
|---|---|---|
| Do we own what we made? | Copyright law, and the model provider's terms | Licence terms, plus a record of the human contribution — briefs, edits, selection decisions |
| May we depict this person? | Right of publicity and personality rights, by jurisdiction | Signed consent scoped to use, media, term and territory — from the person, not from a stock library |
| May we use this performance? | Contract, and collective agreements in union production | Separate written consent for digital replica or synthetic voice, with compensation terms |
The middle row carries most of the risk. It applies whether or not the depiction is flattering, whether or not the content is commercial in the advertising sense, and — under several recent statutes — whether or not the person is famous.
What did the Copyright Office actually decide?
Works generated entirely by AI are not copyrightable in the United States, but copyright can subsist in the human-authored contributions around and inside them. In January 2025 the U.S. Copyright Office published Part 2 of its Copyright and Artificial Intelligence report, on copyrightability, and its conclusions were narrower, and more usable, than the headlines suggested.
- Existing law is sufficient. The Office concluded that no new legislation is needed here, and declined to create a separate registration analysis for AI-assisted works.
- Human authorship is required. This is presented as a bedrock principle rather than a policy choice.
- Prompts alone are not authorship. On currently available technology, prompts do not give a user sufficient control over the expressive elements of the output. The Office left open that this could change if the technology does.
- Human contributions are protectable. Copyright can subsist in human-authored material perceptible in the output, in creative selection, coordination or arrangement of AI-generated material, and in creative modifications of outputs, case by case.
The operational consequence is specific. The protectable asset is the human work around and on top of the generation, and its existence has to be demonstrable, which is a discipline covered in more depth in how professional AIGC video production works. A pipeline that generates and publishes leaves nothing to point to; one where people write the brief, direct the shot list, select among takes, edit and revise — and record it — has an evidentiary basis. That is the United States position; other jurisdictions differ, and some provide for computer-generated works in ways US law does not, so a multinational publisher's practical approach is to build to the most demanding requirement.
How does the law protect someone's voice and likeness?
Right of publicity governs the commercial use of a person's identity, and several jurisdictions have expanded it specifically to reach AI-generated voice and likeness rather than only photographs. Generative voice cloning exposed a gap that older statutes, drafted around images, did not anticipate.
Tennessee closed that gap first. The Ensuring Likeness, Voice, and Image Security Act — the ELVIS Act — was signed on 21 March 2024 and took effect on 1 July 2024, replacing the state's Personal Rights Protection Act of 1984. It adds voice as a protected property right of every individual in any medium, expands liability to anyone who publishes, performs, distributes or otherwise makes available a person's voice or likeness without authorisation, and extends liability to those who distribute or make available a tool whose primary purpose is producing a particular individual's voice or likeness without authorisation.
Two features of that drafting matter for production planning. It protects every individual, not only performers or celebrities, and it reaches distribution of the means of replication, not only the output — which is why an internal voice-cloning capability built on contributor recordings needs the same consent discipline covered in building a licensed voice library for synthetic speech. Other jurisdictions have moved the same way with different mechanics, and the EU adds a disclosure layer: under Article 50 of the EU AI Act, applying from 2 August 2026, deployers generating deepfake content — AI-generated or manipulated image, audio or video resembling existing persons that would falsely appear authentic — must disclose that it is artificially generated, at first exposure at the latest, a requirement covered alongside other jurisdictions' rules in AI content labelling law. Consent and disclosure are separate obligations; obtaining one does not discharge the other.
What consent do performers need to give?
In union production a digital replica or synthetic voice needs its own written consent, describing the specific uses, with compensation attached, and that principle is worth adopting as policy whether or not an agreement compels it. SAG-AFTRA maintains a public resource on artificial intelligence covering its agreements and AI provisions, and its recent TV/Theatrical and Interactive Media agreements carry terms addressing digital replicas and synthetic performance.
- Consent is specific. Named project, defined use cases, media, term and territory. "Consent to AI use" in a standard release is not consent to anything identifiable.
- Consent is separate. Agreeing to be recorded is not agreeing to have a replica made from the recording. Two documents.
- Consent is compensated. Where a replica substitutes for work the performer would otherwise have done, structured payment is the norm in union agreements.
- Consent is bounded. Define what happens at term expiry — deletion, cessation of use, or renewal — before the model is built.
- Consent covers the training, not only the output. If a voice model is trained on a person's recordings, that training use has to sit within what they agreed.
What does a clearance checklist look like before a synthetic asset ships?
Clearance is a production gate with seven checks, run before an asset ships rather than as a review comment afterward. Each step produces a record that has to survive being asked about a year later, in a market the asset was not originally made for.
- Read the model's output terms, and record them per asset. Commercial use permitted? Attribution required? Any restriction on depicting real people or regulated categories? Terms change between versions, so record which version's terms applied.
- Identify every real person, place, brand or event depicted — including incidental ones. A generated street scene containing a recognisable building or a person resembling a specific individual raises the same questions as a deliberate depiction, and nobody finds those unless it is somebody's job — the kind of check covered in AI content governance: disclosure and provenance.
- Obtain scoped consent for each identified person. Written, naming the project, uses, media, term and territory, and covering both training on their material and use of the resulting replica.
- Check the collective agreement position. If any performer is covered by a union agreement, that agreement's AI provisions govern and typically require separate written consent with compensation.
- Document the human contribution. Brief, shot direction, selection among takes, edits and revisions, attributable to named people.
- Apply labelling and provenance at export. Machine-readable marking, plus visible disclosure where the asset depicts real persons or events or where a market requires it, an approach detailed in content provenance: C2PA, SynthID and what survives.
- Record all of it against the asset ID. Model and version, terms version, consents held with expiry dates, human contributors, labels applied, markets cleared.
Why is retrospective clearance so expensive?
The discipline above is mostly cheap, provided it is built in, but the retrospective version is expensive because the record it needs was never made and often cannot be reconstructed. A library whose model versions, consents and human contributions were never recorded, assessed for reuse in a new market under time pressure, usually produces an unknowable answer — and an unknowable answer resolves to do not use. Two design choices remove most of that burden: avoid depicting identifiable real people unless the brief requires it, and keep a single asset ledger, since rights metadata and provenance metadata are one record viewed from two directions.
How does Lifewood handle rights and consent on AIGC deliveries?
Lifewood records rights information at asset level on AIGC deliveries — model and version, consent scope and expiry for any synthetic voice, human contributors, labels applied, markets cleared — because delivery batches cross markets with different rules, and a per-market clearance model does not survive that. Human creative direction sits at the shot level rather than at final review, which is what produces the documented human contribution.
Delivery spans 100+ languages from 40+ delivery centres across 30+ countries, which is why the ledger is built once rather than per market. This asset-level discipline sits within Lifewood's wider AIGC video production work and its broader AIGC services.