When an AI-generated actor looks or sounds like an identifiable person, producers need a clear process—not an assumption that the prompt, model, or training data settles the question. This checklist separates what teams know from what they still need to review.
A production can face questions about the output it distributes
A studio may not know exactly which materials influenced a model, and a resemblance does not establish that the model included a particular person's image or voice in its training data. But production teams make separate decisions when they select, edit, advertise, and distribute an output. A recognizable face, voice, name, or surrounding context can cause viewers to associate a synthetic performance with a real person.
Depending on the facts and jurisdiction, a dispute may raise publicity, privacy, false-endorsement, defamation, copyright, trademark, contract, or advertising questions. Not every theory applies to every output, and not every similarity is legally actionable. The U.S. Copyright Office's digital-replicas report describes the existing U.S. framework as a mix of state and federal law and discusses the difficulty of applying it consistently.
Separate model training from likeness use
Ask two different questions: what is known about the source material or model, and whether the finished use appears to exploit a particular person's identity. The first may require provider documentation or model-specific information that a production cannot independently see. The second can be reviewed against the actual image, voice, script, audience, and campaign context.
Do not state that a lookalike came from a specific training example unless there is evidence. Do not treat a lack of evidence about training data as proof that the output presents no rights concern. Keep those conclusions separate in the production record, and use an authorized provider's terms and response where the workflow requires them.
Use a consent record that matches the real brief
For a licensed performer or contributor, connect the signed terms to the exact person, asset, client, use category, territory, duration, media, and version of the campaign. Explain whether the permission includes synthetic performances, voice generation, altered appearances, edits, localization, paid ads, and future reuse. If the script or brand association changes, check the scope again.
State laws illustrate why a generic release is not a universal answer. Tennessee's 2024 ELVIS Act amended state law to address simulated voices. California AB 2602 addresses certain digital-replica terms in personal or professional services contracts and includes specific conditions. The laws have different scopes; counsel should confirm which current rules apply to a production.
A practical review before release
Assign a reviewer to compare the final output and planned distribution with the brief, license, source records, and provider terms. Pay attention to recognizable facial or vocal traits, names and biographical details, apparent endorsement, product claims, and whether an audience could mistake a fictional character for a real person. If an output raises a credible concern, hold it while the team seeks permission, replaces the output, or gets legal guidance.
Preserve the relevant prompt and source records, versioned output, terms, reviewer decision, and distribution plan according to the team's retention policy. Do not overwrite an unresolved concern with a generic cleared status. A clear record helps the production explain what it checked without claiming that software or an internal checklist made a legal determination.
- Identify whether a real person could reasonably be recognized from the full context.
- Check for claims or scenes that imply the person endorsed a product or message.
- Confirm permission covers the actual medium, client, territory, term, and edits.
- Escalate uncertainty to counsel before public distribution.
CLARIFICATIONS
Questions & answers
Can a production be sued even if it did not name a real person in the prompt?
A person may bring a claim based on the finished output and its use, even where the prompt did not name them. Whether the claim succeeds depends on the evidence, jurisdiction, context, and applicable defenses.
Does a similar-looking output prove the model trained on that person?
No. A visual or vocal resemblance by itself does not establish what data a model used. Training provenance and the rights analysis for the output are distinct questions.
Will a disclaimer prevent a likeness claim?
A disclosure can help explain that media is synthetic, but it does not create permission or resolve every rights issue. Review the intended use and applicable law separately.
REFERENCES