Synthetic Performer Clearance: What to Lock Before an AI Face Fronts a Brand

Why a Synthetic Performer Is Now a Named Object in Advertising Law

A synthetic performer is no longer an informal term for an AI face in an ad. New York amended section 396-b of its General Business Law to define it as a digitally created asset, created or modified by computer using generative artificial intelligence or a software algorithm, intended to create the impression that the asset is engaging in an audiovisual or visual performance. The statute carries a $1,000 civil penalty for a first violation and $5,000 for any subsequent violation. Once a term is defined in statute, it stops being a creative choice and becomes a production object with paperwork attached.

That shift matters because most AI video clearance work has been aimed at the wrong layer. Teams audit whether the model output can be used commercially, whether the training data is contested, and whether the platform tier covers paid media. Most teams already treat AI video commercial rights as a separate workstream, and the performer layer now needs the same discipline. A clean footage licence tells you nothing about whether the face reading your script is allowed to exist in that ad.

The practical test is simple. If a reasonable viewer could watch the cut and form a belief about a human being - that this person exists, that they work for the brand, that they tried the product - then a performer layer is present and needs clearing, regardless of how the pixels were made.

The Three Origins of an AI Face, and What Each One Owes

Clearance requirements diverge sharply depending on where the face came from, and the three origins are worth separating on the call sheet. The first is a fully invented persona with no real-world referent: a model-generated spokesperson who is not identifiable as any living individual. The second is a digital replica, meaning a real performer whose voice or likeness has been captured and reused to produce a performance they did not actually give. The third is the ambiguous middle, where a generated face lands close enough to a recognisable person that audiences make the connection anyway.

The invented persona is the cheapest to clear and the easiest to get wrong, because its obligations are about disclosure and truthfulness rather than permission. The digital replica is the opposite: permission is the whole job, and the contract language is load-bearing. The lookalike is the dangerous one, because nobody signs off on it and it usually surfaces after the client has already approved the cut.

Build the origin question into casting rather than into legal review. Decide which of the three you are commissioning before the first render and log the answer, because retrofitting the classification onto a finished cut means re-clearing everything downstream of it.

What Your Synthetic Performer Can Never Say

The Federal Trade Commission's rule on consumer reviews and testimonials draws a hard line that most AI spokesperson scripts walk straight across. Under 16 CFR 465.2(a), it is a deceptive act for a business to create a testimonial that materially misrepresents that the testimonialist exists, or that they used or otherwise had experience with the product. A generated persona fails both tests by construction: it does not exist, and it has never used anything.

This is a script constraint, not a disclaimer problem. An AI face can demonstrate a product, narrate a benefit, present a specification, or play an obviously fictional character. It cannot say the serum cleared its skin, that it switched from a competitor, or that the results shown are its own. Adding a small on-screen label does not cure a first-person experience claim, because the deception sits in the substance of the statement rather than in the medium.

The cleanest place to enforce that boundary is the creative brief for AI video, where the script rules get written before anyone opens a model. Give writers a two-column reference: permitted register on the left, covering demonstration, narration and fiction, and prohibited register on the right, covering personal results, ownership and comparative experience. Reviewing this at script stage costs minutes, while catching it at delivery costs a reshoot that cannot be reshot.

The Consent Chain Behind a Digital Replica

When the face belongs to a real person, California's Labor Code section 927 shows exactly how thin a consent chain can be before it stops holding. A digital replica provision is unenforceable for a new performance fixed on or after 1 January 2025 if it lacks a reasonably specific description of the intended uses, and the individual was not represented either by legal counsel negotiating the replica licence with clearly stated commercial terms, or by a union whose collective agreement expressly addresses digital replicas.

Read that as a drafting checklist rather than a legal curiosity. A blanket all-media, all-purposes, in-perpetuity clause buried in a standard talent release is precisely the shape the statute targets. What survives scrutiny is specific: which campaigns, which markets, which channels, which product categories, how long the replica may be used after the capture session, and whether the brand may generate new lines the performer never recorded.

Two clauses get missed almost every time. The first is a retraining and retention term, covering whether the captured data may be kept, reused to fine-tune a model, or must be destroyed at campaign end. The second is a revocation path that states what happens if the performer's reputation changes mid-flight, which is the scenario that turns a cost saving into a recall.

Disclosure Is Three Obligations, Not One

Teams routinely collapse three separate disclosures into a single label and assume they are covered. The first is the statutory notice that state advertising law attaches to the performer itself. The second is the platform requirement: YouTube requires creators to disclose generative AI that makes a real person appear to say or do something they did not, or that generates a realistic scene which never occurred, and it may apply labels automatically to uploads carrying C2PA metadata. The third is the ordinary sponsorship disclosure, which does not disappear because the endorser is synthetic.

They stack rather than substitute. A brand-operated AI persona in a paid placement can owe all three at once, and each has a different audience, a different placement and a different failure mode. Platform-side failure is the fastest to hurt, because consistent non-disclosure can lead to content removal or suspension from the YouTube Partner Program, which takes distribution away long before any regulator writes a letter.

Teams that already maintain an AI video disclosure checklist can extend it with a performer row rather than starting over. Add three columns - statutory notice, platform attribute, sponsorship tag - and require each to be marked applicable or not applicable with a named owner. Ambiguity here is what produces cuts that ship with a note in the description and nothing on screen.

The Clearance Packet a Producer Files Before the First Render

Clearance fails when it lives in email. The fix is a single file, assembled before generation starts, that a producer can hand to a client's legal team without reconstructing anything. Six items belong in it, and all six are cheap to collect up front and expensive to rebuild later.

One: the origin classification, invented persona, digital replica or lookalike risk, with the reasoning recorded. Two: the likeness provenance, meaning which reference images, avatar library or capture session produced the face, with the vendor terms attached. Three: the executed consent, carrying the specific use description, term, territory, media and retraining language. Four: the script register showing what the performer may and may not claim. Five: the disclosure matrix with named owners. Six: the provenance and export record for the delivered master.

In a producer-led AI video workflow, the producer owns this file the same way they own insurance certificates and location permits. It is not a compliance artefact filed after the fact, it is a precondition for booking render time. The packet also forces the origin question early, and the origin question is what prevents the expensive mistakes.

Where Performer Clearance Sits in the Pipeline

Put the clearance gate before generation, not before delivery. Every check described here is answerable at brief stage, and every one of them becomes irreversible once a face has been approved by a client and cut into a master. The cost curve is unusually steep: a wrong origin classification caught at brief costs a conversation, and the same error caught at delivery costs the entire performer layer.

Fold the performer checks into the same AI video QC checklist the team already runs before delivery, so nothing needs a separate meeting. Delivery-stage review still has a job, confirming that the on-screen notice survived the final export, that the platform attribute is set, and that the shipped cut matches the approved script register, but it should be verification rather than discovery.

The broader shift is worth naming. Generative tools made the face the cheapest element in a commercial to produce and the most expensive one to get wrong. Treating an AI cast member as talent with paperwork, rather than as an asset with a prompt, is what keeps that asymmetry from landing on the client.