Creator-style advertising works on a borrowed assumption. The shaky frame, the room lighting, the unpolished delivery all say: this is a person, not a campaign. Generative production can reproduce every one of those signals perfectly, and the moment it does, the assumption becomes something you are responsible for.
The obligations are not one set. They fork depending on what kind of face is on screen.
The three cases, and how they differ
| CASE | WHAT IT IS | CONSENT NEEDED | DISCLOSURE NEEDED |
|---|---|---|---|
| Wholly synthetic person | A face with no real-world referent | Model licence cleared for commercial use; no personal consent to obtain | Yes, wherever the format implies a real customer or user |
| Trained identity of a real person | A model built from a real performer’s stills or footage | Signed release covering likeness, voice, scope, duration, territory and derivative training | Yes, plus performer credit where their contract requires it |
| Real person, synthetic words | Real footage or voice, generated script or lip-sync | Explicit release for synthetic dialogue, separate from the original shoot release | Yes, and this is the highest-risk case of the three |
Consent: what a usable release actually covers
Most likeness releases in circulation were drafted for photography and film. They grant use of captured material. They do not grant the right to build a model that can generate new material the person never performed, which is the entire point of a trained identity.
A release that works for synthetic production names, at minimum:
- The specific permission to train a model on the supplied material, stated in those words rather than implied by a broad grant.
- Scope of output: what the generated likeness may be used to say and sell, and what it may not. A blanket grant is a liability for both sides.
- Territory and duration, with a stated end date rather than "in perpetuity", which is increasingly unenforceable and always a negotiation risk.
- What happens to the trained model at the end of the term. Deleted, escrowed, or retained. Say which.
- A withdrawal mechanism, and an honest statement of what withdrawal can and cannot undo for material already published.
- Categories the likeness will never be used for. Political content, adult content, and claims about health, finance or legal matters are the usual carve-outs and performers ask for them.
Likeness: the cases that are not obvious
Two failure modes recur, and neither involves anyone deliberately impersonating anybody.
The first is accidental resemblance. A wholly synthetic face generated from a prompt describing a demographic will sometimes land close to a real, identifiable person, particularly in a small market. There is no malice and the exposure is the same. The mitigation is dull and effective: reverse image search on the final identity before it enters production, and a written record that you did.
The second is voice. A synthetic voice trained on a real person is a likeness, whatever the contract calls it, and a voice that merely sounds like a recognisable performer is where the newest and least settled law sits. Treat voice with the same paperwork as face. A signed release for the voice goes in the folder before anything renders, and that applies to a founder’s own voice as much as to a hired performer’s.
Disclosure: what synthetic UGC specifically requires
The general disclosure position is covered elsewhere, but synthetic UGC has a particular problem: the format itself is the claim. A polished brand film that happens to be generative is not asserting that a person exists. A piece to camera in a kitchen is.
So the standard is higher. Where a viewer would reasonably read the format as a real customer, user or independent creator, disclose regardless of whether the underlying claims are substantiated. And put it on the asset. A caption disclosure disappears the moment the video is reposted, and creator-format content gets reposted.
Since 2 August 2026, Article 50 of the EU AI Act has required deployers of deepfake-producing systems to disclose that content is artificially generated, clearly and at first interaction, wherever the content reaches people in the EU. Synthetic UGC sits squarely inside that. In the UK, CAP applies the existing misleadingness rules, which for this format reach the same conclusion by a different route.
The pre-render checklist
Nothing on this list happens after generation. All of it happens before, because every item is cheaper to satisfy than to remediate.
- Model licence confirmed as cleared for commercial use, for the specific model and version being run.
- Signed release in the folder covering likeness, voice, derivative training, scope, territory, duration and end-of-term disposal.
- Reverse image search on any wholly synthetic identity, with the result recorded.
- Disclosure decision made and written into the brief, set to the stricter of platform policy, UK position and EU Article 50.
- Claim substantiation held for anything the presenter says about the product. A synthetic presenter cannot make a claim the brand could not make itself.
- A named person with authority to stop the run, and a defined point at which they see the output before it ships.
- Provenance metadata configured on export, plus the on-asset label.
Consent is not a formality. A signed release sits in the folder before anything renders, and that applies to a founder’s own voice as much as to anyone else’s.
A synthetic likeness and voice release covering derivative training, scope, territory, withdrawal and end-of-term disposal. Free, no email gate.
DOWNLOAD THE CONSENT TEMPLATE →