A standard talent release is designed to answer a familiar production question: may the brand use the performance that was recorded? Generative production introduces a different asset. The same face can deliver a new line, move in a scene that was never shot, speak another language, or endorse a claim the performer never recorded.
That changes the procurement decision. A brand is no longer clearing only a captured performance. It may be acquiring the ability to create future performances from a person's identity.
On September 23, 2026, YouTube announced that likeness detection was coming to mobile and that it would begin integrating speaking-voice detection with facial detection. The current YouTube Help page is more precise about today's boundary: the feature is experimental, currently looks for the enrolled creator's face, and is still being extended to audio. A match lets a creator review a video and consider a privacy or copyright request; it does not automatically decide whether a use was authorized.
The signal is not simply that synthetic ads may become easier to find. It is that faces and voices are becoming searchable, contestable production assets. Brands therefore need a rights record that can travel with every generated version.
Detection, disclosure and permission solve different problems
These three controls are often collapsed into one compliance checkbox. They should remain separate.
Detection asks whether a platform can find a possible identity match. It is a post-publication discovery mechanism. It can miss content, and the YouTube help documentation says it may also surface ordinary footage featuring the enrolled creator's actual face. Detection is useful monitoring, not proof of authorization.
Disclosure asks whether the audience is told that content was generated or materially altered. YouTube's May 27, 2026 update, “Improving AI labels for viewers and creators,” describes creator disclosure and automatic signals for its AI labels. The IAB AI Transparency & Disclosure Framework V2, announced August 18, 2026, applies a risk- and materiality-based approach to formats including synthetic voices and digital twins. A label can inform a viewer, but it cannot grant the brand a license.
Permission asks whether this person agreed to this brand, use, script, market, channel and term. It must be resolved before generation and distribution. A disclosed ad may still exceed its license. An undetected ad may still be unauthorized.
The operational order is therefore simple: clear the identity use first, decide the disclosure second, and treat platform detection as a monitoring layer after publication.
When the standard release stops being enough
Not every AI-assisted edit requires the same control. A practical procurement model separates three production levels.
| Production action | Procurement decision | Minimum control |
|---|---|---|
| Blemish removal, keying, grading, ordinary subtitles or conventional audio edits | If no new identifiable performance is created, review against the original production release | Retain source footage, edit records and the approved channel list |
| Lip-sync translation, generated extensions, new motion or scenes featuring the same identifiable person | Treat the output as a new digital performance | Specify face, body, movement, language, market, media and term |
| Voice cloning, new dialogue, new product claims or a synthetic endorsement | Use the highest approval level because the work changes what the person appears to say | Separate voice permission, bounded claims or scripts, version approval and a takedown owner |
This is a production control, not universal legal advice. Rules vary by market, contract and performer status. The underlying test is useful across workflows: if the output could reasonably look or sound like a person performing something they never performed, do not treat it as routine post-production.
SAG-AFTRA's public AI resources offer one concrete but limited industry reference. Covered arrangements emphasize informed consent and the intended use of digital voice replicas, and some commercial arrangements require additional consent for a new advertisement. Those terms do not govern every project. They do show why consent to be filmed and consent to generate a new performance are becoming separate transaction records.
The ONCE Likeness Ledger: six fields that travel with the asset
The answer does not have to begin with a sprawling policy document. Put one shared ledger behind the brief, contract schedule, production tracker and delivery manifest.
- Identity. Record the performer or talent ID, the party authorized to approve use, the release version, and whether face, body, movement and voice are each included.
- Generation scope. Distinguish retouching, frame extension, lip-sync, translated speech, voice cloning, face replacement, full digital doubles and model training. “AI post” is not a usable scope.
- Commercial use. Record the brand, product, channels, countries or territories, languages, organic versus paid media, and start and end dates. Worldwide, perpetual use should never be inferred from a production shorthand.
- Version authority. Define which scripts, claims, emotions, actions and personalization variables may be generated. Claims involving efficacy, price, guarantees or sensitive categories should retain version-level human approval.
- Evidence chain. Give every output an asset ID linked to the consent version, source materials, model or vendor, generation date, responsible producer, disclosure decision and final approver. Keep the record out of private chat threads.
- Expiry and withdrawal. Name the reminder date, takedown owner, platform contact, and what must happen to source material, trained models, embeddings, caches and archives when the permission ends.
The ledger is valuable because the creative version and its permission share an identifier. When an English master becomes a Spanish adaptation, the producer can immediately see that translation is allowed but cloned voice is not. When a talent term expires, the media team can identify the live ads without reverse-engineering old campaign folders.
Put the ledger into the production workflow
At concept approval, ask whether an identifiable person is necessary. Product-only demonstrations, hands, abstract characters or fully fictional talent may communicate the idea without introducing a reusable identity asset. Synthetic realism is not automatically a creative advantage.
At casting and bidding, separate the rights variables. A useful estimate distinguishes live capture, creation of the digital double, voice rights, languages, version expansion, territories and term. Two quotes with the same total may purchase radically different permissions.
Before generation, gate the source materials. Only files cleared for that person and purpose should enter the selected tool. Permission to generate campaign outputs is also different from permission to train or improve a model. Track the two separately.
During review, run an identity pass in addition to creative and platform review. Check whether a script changes the original meaning, whether the tone creates a new promise, whether face, voice and action remain inside the authorized scope, and whether the disclosure decision is appropriate for the market and platform.
At delivery, ship more than the MP4. The minimum usable package should include the asset ID, linked permission version, approved territories and term, disposition of source files, disclosure decision and takedown contact. More versions should increase useful output, not multiply undocumented liability.
What this changes for brands and production partners
For a cross-border brand, the buying question changes from “How many videos can this system generate?” to “Which identity capabilities are licensed, for how long, and who approves each new performance?”
For a production partner, this changes the margin model. Pricing a single deliverable while silently absorbing unlimited languages, scripts and expiry management turns generative scale into unbounded service work. A defined likeness ledger, approval workflow and expiry service make the operational work visible and billable.
For global marketing teams working with Chinese manufacturing and content partners, the central judgment is straightforward: stronger platform detection will not complete the permission process for you. It will make undocumented decisions easier to surface after the campaign is live.
The next time a brief asks for ten language versions from one model's face, do not begin with the generation tool. Fill the six fields first. In a digital-replica project, that ledger is the first production asset.
Sources checked October 4, 2026: YouTube's September 23 announcement, YouTube Help on likeness detection, YouTube's AI-label update, IAB Framework V2, and SAG-AFTRA AI resources. This article separates verified platform facts from ONCE's procurement analysis and is not legal advice.
