A person on screen can speak, perform and carry the memory of an era. For a brand buyer, however, the hardest question is not whether the likeness looks convincing. It is who authorized that identity to speak, where it may appear, who approves each variation, and who makes it disappear when the license ends.
On October 2, 2026, Reuters reported that Swedish bearing maker SKF had recreated Greta Garbo for an advertisement using AI. According to SKF, the visual double was generated from text prompts rather than archival photos or footage. Garbo's voice from one early film was used with permission for voice training, a human actor contributed pacing and cadence, and the production was developed with the Garbo estate and family.
That case confirms one specific workflow: a historical likeness can be assembled through generative tools, human performance and rights-holder collaboration. It does not prove that every likeness is available for commercial use, nor that one approval clears every market. The useful procurement lesson is narrower and more durable. A digital performer should not be bought as “one AI video.” It should be bought as an asset that can be licensed, traced, approved, labelled, updated and retired.
Synthetic identity is now a delivery problem
The European Commission's Article 50 quick facts state that the AI Act transparency rules apply from August 2, 2026. Providers must enable reliable identification of synthetic content, including machine-readable marking in scope. Deployers must clearly and perceptibly disclose deepfakes: AI-generated or manipulated image, audio or video that resembles an existing person, object, place, entity or event and could falsely appear authentic. The Commission also distinguishes standard assistive editing from operations that substantially alter supplied content or meaning.
In the United States, a New York law effective June 9, 2026 requires producers or creators of an advertisement to identify when it includes an AI-generated synthetic performer. The IAB AI Transparency & Disclosure Framework V2, published in August 2026, recommends a risk- and materiality-based approach. Its position is not that every routine AI assist needs the same label; the higher-risk cases are those in which AI materially changes authenticity, identity or representation.
Exact obligations vary by jurisdiction, placement and creative execution, so local counsel remains necessary. Production can still solve a practical part of the problem now: define who decides disclosure, what must be disclosed, and which final versions carry each human-visible and machine-readable signal.
The five-part digital-performer handoff
1. Identity scope card
Define what the performer is before producing the performer. Is it a licensed extension of a living person, a reconstruction of a historical figure, an authorized digital double, or a fictional character that represents no real individual? Record reference sources, permitted facial and vocal traits, prohibited situations, and relationships the audience must not infer.
A fictional character still needs a non-resemblance check and clean asset licensing. A folder of photographs is not evidence of commercial likeness rights.
2. Permission matrix
Break permission into actual use conditions: grantor, territory, channel, language, term, paid media, edits, voice replication, derivative generations, model training, archival use and end-of-term treatment. A worldwide social campaign and a film licensed for one market are different products even when the picture is identical.
For historical figures, identify who controls the name, likeness, voice and underlying performance material. Then specify whether the rights holder approves the concept, script, work-in-progress frames, or every final output. The matrix reveals the gap between “technically generatable” and “cleared to publish.”
3. Generation record
Record models and versions, input provenance, the contribution of any human actor or voice performer, key generation and compositing steps, manual retouching and material revisions. C2PA Content Credentials can bind provenance assertions to an asset and help detect modification. They do not decide whether permission was valid, the likeness is accurate, or the creative use is respectful.
Technical metadata therefore complements rather than replaces the production ledger. A new operator should be able to identify where a shot came from, which version was approved, and whether a revision requires fresh approval.
4. Final-output approval sheet
“Looks good on the monitor” is not acceptance. For every final version, verify appearance, voice, line reading, action, product claim, context, visible disclosure, machine-readable mark, market and channel. Any version that changes the performer's apparent intent or the brand's meaning returns to approval.
Approve final outputs, not only a master concept. A vertical crop can hide the label. A dubbed line can change meaning. An automated platform adaptation can move disclosure outside a safe area.
5. Disclosure and retirement ledger
For each deliverable, record the public label, platform setting, provenance credential, launch date, rights expiry, hosting location, takedown owner and archive rule. When a license ends, stopping new media spend is not enough. The team must also check brand websites, social accounts, distributor libraries, historical landing pages and active ad accounts.
A manageable digital performer must be findable—and removable.
Compare the full cost, not generated minutes
A practical project formula is:
Effective cost per approved use combination = rights + identity and voice preparation + human performance + generation and finishing + legal and market review + disclosure and provenance + renewal and retirement, divided by approved market × channel × language × term-version combinations.
This is not an industry benchmark. It is a way to compare completeness within one project. A low-cost generation package without permission scope, approval records or takedown responsibility may produce only a demo. A more expensive system that clears several markets, preserves an updateable master and transfers the rights records may become a durable brand asset.
Supplier quotes should separate licensing, human performance, voice work, generation rounds, manual finishing, version outputs, market review, labelling, provenance and ongoing maintenance. That distinction shows whether cost comes from craft, governance or broader commercial use.
When a real person's AI likeness is the wrong choice
If the identity itself does not add meaning, use a fictional character that does not resemble a real person, or let the product and visual idea carry the story. If the communication must prove function, craft, user experience or safe operation, live action and real product demonstration usually provide clearer evidence. ONCE's article on filming the work customers cannot prompt explains that distinction.
A real-person digital likeness is also a poor fit when the rights chain is uncertain, market rules diverge, the schedule cannot accommodate approvals, or the brand has no owner for labels and retirement. A live performer, stylized animation, or an AI-assisted background workflow may be the safer and more effective creative decision.
Digital performers become worth the additional structure when identity is the premise—historical memory, endorsement continuity or controlled multilingual extension. The brief should then say more than “make an AI character that looks like this person.” It should define who the performer is, who authorizes the use, how the asset is generated, who approves it, how audiences are informed and when the asset must come down.
ONCE's view is simple: AI makes a person generatable; brand operations must make that person traceable and retractable. Teams planning a hybrid China production for global commerce can pair this framework with ONCE's cross-border ecommerce video scope to separate product facts, market versions and live-action proof before deciding what generative tools should create.
