A brand can secure permission to use a person’s likeness and still be unready to make the ad. In synthetic talent production, the license answers only the first question: who may use the identity. It does not automatically define the approved voice source, performance, product claims, derivative cuts, disclosure treatment, or takedown process.
That distinction became unusually visible in SKF’s new campaign featuring an AI likeness of Greta Garbo. SKF said on September 30, 2026 that the initiative was developed with Garbo’s family and Harriet Brown & Company. Reuters reported on October 2 that the visuals were built from text prompts rather than archival photographs or footage; one early film was licensed as voice material, and a human actor helped shape pacing and cadence.
This is one company’s production choice, not evidence that the market has settled on a universal standard. Its value for buyers is more concrete: a synthetic likeness can require separate decisions for rights, inputs, performance, tools, approvals and release versions.
The license and the production brief do different jobs
The rights document should define who controls the name, face, voice, archival material and any other protected identity asset. The production brief must turn those permissions into shot-level instructions: what the person may say, which products may appear, which markets and languages are covered, who approves the voice, and whether a 15-second cut, vertical edit or campaign still counts as the same use.
When these documents are treated as interchangeable, three common gaps appear. A face is licensed but the voice source is not. A hero film is approved while social cutdowns are not. The contract permits use, but the creative team has no written limit on claims, gestures or historical context.
Before choosing an all-generative route, a live-action performance or a hybrid live-action and AIGC workflow, ONCE recommends building five controls that can be signed off independently.
Control 1: an identity authority map
Map each identity element to an approver and a scope. Cover territory, language, media, product category, term, sublicensing to production and media partners, derivative edits, still-frame extraction and reuse after the original campaign. “Likeness approved” is too broad to manage a real deliverable list.
The working output should be a matrix aligned to the shot list and version list. If any cell remains unresolved, those shots should not enter batch generation. This is a production gate, not a legal conclusion.
Control 2: a source and generation register
Log every material input: reference images, voice recordings, scripts, body-double performances, model names, tool versions, generation dates and exported files. Also record how each source was used—training, reference, transformation or inclusion in the final master. SKF’s different approaches to visual and voice inputs show why a single “AI source” label is not specific enough.
This register is what makes future revision possible. If a tool term changes, a permission expires or one source must be withdrawn, the team can identify affected outputs without reconstructing the entire project from memory. Prompts alone are not a provenance system.
Control 3: a performance bible with staged approvals
A synthetic person can resemble someone while behaving in a way that person never would. The performance bible should therefore define permitted product categories, prohibited scenes, allowed and disallowed claims, vocal cadence, gesture range, wardrobe, historical context, and whether the identity may be placed in invented events.
Approval should happen at separate stages: script, storyboard, key frame, voice sample and final edit. Reuters’ report that a human performer helped shape pacing and cadence is a useful reminder that direction remains a production cost. It does not prove the campaign worked better; it shows that convincing synthetic performance is not simply a model-output task.
Control 4: a disclosure matrix for every delivery
The European Commission says Article 50 transparency obligations have applied since August 2, 2026. Deployers of AI-generated or manipulated image, audio or video content that constitutes a deepfake must disclose that it was artificially generated or manipulated. For evidently artistic, creative, satirical, fictional or analogous works, the manner of disclosure may be adapted so it does not impede enjoyment of the work; the obligation is not simply erased. See the Commission’s Article 50 transparency guidance.
That means disclosure must travel with the asset. A label that survives in a 60-second horizontal film may disappear in a six-second cut, silent autoplay placement, crop or localized subtitle file. Add “disclosure remains visible or audible” to the acceptance criteria for every master and derivative.
For U.S. work, the FTC’s updated Endorsement Guides also explicitly address virtual influencers and emphasize clear and conspicuous disclosure and potential liability across advertisers, endorsers and intermediaries. Whether a particular synthetic performance is perceived as an endorsement requires case-specific review; production teams should flag that question rather than assume the audience understands the fiction. This article is production guidance, not legal advice.
Control 5: an approval and exit ledger
For each published file, record the version ID, approver, live channel, permission expiry, re-edit status, project-file custodian and takedown owner. Set stop triggers before launch: a product recall, brand crisis, objection from the rights holder or family, a compliance change, or a material change to the approved claim.
Without an exit ledger, a “reusable digital asset” can become an unmanaged liability. ONCE considers the delivery complete only when the brand can trace every live version to its authority, inputs, approval and removal path.
The true cost is larger than the generation line item
Total synthetic talent cost = rights and licensing + source preparation and provenance + generation and look development + performance direction + review, disclosure and localization + lifecycle and takedown reserve.
This is not an industry benchmark or a claim about average pricing. It is a scope model. It lets buyers compare proposals that otherwise use incompatible assumptions. A quote based only on duration, output count and revision rounds is incomplete if it excludes approval stages, source records, disclosure variants and removal responsibility.
Attach the five controls to the request for proposal. Ask each supplier to price the same responsibility set. The exercise often reveals whether the apparent saving comes from efficient production—or from leaving critical work with the client.
Three stop conditions
- The authority chain cannot be resolved. If voice, archive, territory or derivative rights remain uncertain, use contracted live talent or an original fictional character.
- The performance would create a high-risk endorsement impression. If viewers may believe the real person personally used, evaluated or guaranteed the product, obtain market-specific review. If the ambiguity cannot be designed out, do not proceed.
- The campaign needs constant real-time variation. If every cut must return to an estate, agent, legal team and brand approver, generation speed is not the bottleneck. A non-identifiable brand character may be more controllable.
Put the five controls in the first creative meeting
The important lesson from the SKF case is not that AI can bring a historical figure back to the screen. It is that identity collaboration, input selection, human performance direction, multi-tool production and public disclosure must coexist in one delivery system.
Teams evaluating synthetic talent, fictional digital characters and live action should begin with the person, markets, channels, term and version count—not with a demo. ONCE works with brands connected to China’s supply chain to define the hybrid route, approval structure and deliverables before production begins.
Sources
- SKF Group, “Greta Garbo is back - starring in SKF's initiative to overcome friction,” September 30, 2026.
- Reuters, “Swedish ball-bearing ad revives Greta Garbo with US, Chinese AI,” October 2, 2026.
- European Commission, “Guidelines on transparency obligations for providers and deployers of certain AI systems,” current guidance for Article 50.
- U.S. Federal Trade Commission, “Federal Trade Commission Announces Updated Advertising Guides to Combat Deceptive Reviews and Endorsements,” June 29, 2023.
