Synthetic talent and the rights stack.

A synthetic spokesperson costs less than a celebrity, never ages out of the contract, and can shoot in nineteen languages on a Tuesday. That is the pitch. The reality is that the asset is only as defensible as the rights stack underneath it — and most of that stack was rewritten between 2024 and 2026.

What actually exists now

The term "AI influencer" covers three very different assets, with three very different rights profiles.

• Fully synthetic persona — a character that does not correspond to a real human (Lil Miquela, Aitana López, Imma). The likeness and voice are authored. Rights questions are mostly about training data and trademark, not personhood [1].
• Licensed digital double — a photoreal AI version of a real performer, used with consent for specific contexts (the SAG-AFTRA digital replica framework). Rights questions are about scope, duration, revocation, and per-use compensation [2].
• Unlicensed deepfake — a real person's likeness used without consent. Not a brand-safe asset under any current regime; the 2025 NO FAKES Act and the EU AI Act both make this materially riskier [3][4].

The first two are the live brand surface. The third is the legal cliff most brands do not realize they are walking near.

What changed in the last 18 months

Four shifts moved the ground.

1. The 2023 SAG-AFTRA contract. The TV/Theatrical agreement established the first industry-wide framework for AI digital replicas — informed consent, bargained-for compensation, and scope limits per project [2]. Every major brand campaign that uses a real performer's AI likeness now defaults to this language.
2. The EU AI Act (2024) and its 2026 enforcement milestones. Deepfake disclosure obligations, transparency labeling for AI-generated content, and provenance metadata requirements are phasing in across the EU through 2026–2027 [4]. A campaign that runs in Europe without that disclosure is exposed.
3. The NO FAKES Act (US, 2025 reintroduction). Federalizes a right of publicity over digital replicas and sets statutory damages for unlicensed use. Several states (Tennessee's ELVIS Act, California's AB 2602 and AB 1836) already enforce parallel rights [3].
4. C2PA provenance as a contract requirement. Brands and platforms increasingly require Content Credentials (C2PA) embedded in delivered assets, so the synthetic chain-of-custody is portable across distribution [5].

The rights stack a campaign actually needs

We build every synthetic-talent campaign against five contractual layers. Skip any one of them and the asset has a hole.

• Likeness rights. For digital doubles: SAG-AFTRA-compliant consent, scope (project, territory, duration), revocation triggers, and re-use fees. For synthetic personas: a clear authorship record and a trademark filing on the character.
• Voice rights. Treated separately from likeness in most jurisdictions post-2024. Cloned voice requires its own consent and its own kill switch.
• Performance rights. The moves, gestures, and mannerisms an AI model learned from a specific performer can themselves be protectable. Address this explicitly in the contract instead of assuming it falls under likeness.
• Training-data provenance. A warranty from the model vendor that the model was trained on licensed or permissible data, and an indemnity that survives delivery. The 2023–2025 wave of copyright suits against generative model providers makes this non-optional [6].
• Disclosure and provenance metadata. C2PA Content Credentials embedded in every delivered file, and a clear disclosure plan for any market with a labeling rule (EU AI Act, China's deep synthesis regulations, several US state laws) [4][5].

What works commercially

The defensible play is not to fake a celebrity. It is to author a character the brand actually owns.

Fully-authored synthetic personas — characters the brand owns the trademark to, voiced by consented voice talent, animated by consented motion talent, on a model with clean training data — clear every layer of the stack in one structure. They are also the only synthetic-talent asset that can be ported into agentic commerce, gaming, AR, and licensing without renegotiating the rights chain each time.

Licensed digital doubles work, with discipline. Unlicensed deepfakes do not work, full stop — the 2025–2026 enforcement environment is the loudest signal in the industry [3][4].

The takeaway

The model is the cheap part. The contract is the asset. Build the rights stack first, then build the character, then ship the campaign. In that order.