Direct answer: An AI digital human or virtual anchor is not a video template. It combines portrait/voice rights, deep-synthesis labelling, advertising and endorsement rules, training data and a vendor contract. Before launch, settle who owns the face and voice, the scope and term of consent, whether the output is labelled, whether it counts as advertising, where data lives and who is liable—then discuss looks and lip-sync.
This guide is for marketing/brand, e-commerce/live commerce, customer service, legal/compliance, IT, procurement and vendors. It covers how to launch and accept AI digital humans, virtual anchors, digital twins and AI voice-overs. Information is current to 31 October 2026. It is public regulation, judicial interpretation and project method, not legal advice; portrait/voice consent, labelling, advertising review, platform rules and contractual liability are governed by the applicable rules, platform policies, signed authorisations and legal counsel.
Check the rights chain before model quality; a real person's face/voice needs explicit authorisation and separate consent; AI-generated/deep-synthesised content needs labelling; advertising, endorsement and live selling must be true and verifiable—never fake an expert, customer or medical result.
What kinds of AI digital human exist, and where do the risks differ?
Real-person clones, licensed digital twins, purely virtual personas and voice clones carry different risks. A real face/voice involves portrait and voice rights plus sensitive personal information; a virtual persona still involves training data, copyright and advertising.
"Digital human" is often sold as one product, but there are at least four types: a digital twin trained on a real person's photos/video; a virtual persona generated under a real person's authorisation; a fully fictional 3D/2D avatar; and an AI voice clone used only for audio. Real-person and voice clones involve portrait rights, voice interests and sensitive personal information; purely virtual personas involve copyright, trademarks, training data and advertising; live commerce adds platform rules and consumer protection. Classify first, then decide authorisation, labelling and contract.
| Type | Typical use | Main rights/risks | Confirm first |
|---|---|---|---|
| Real-person digital twin | Founder/expert/anchor presenting, customer service | Portrait/voice, personal information, advertising endorsement | Written authorisation/term/channels |
| Licensed virtual persona | Brand virtual spokesperson, IP character | Copyright/trademark/scope of licence | IP ownership/use scope |
| Purely virtual persona | Brand-built virtual anchor/service agent | Training data/copyright/advertising | Material sources/platform rules |
| AI voice clone | Voice-over, audio content, customer service | Voice interests/sensitive personal information | Separate consent from the voice owner |
| Digital-human live commerce | 24-hour streaming, product explanation | Advertising/consumer rights/platform rules | Ad review/platform filing |
The classification is editorial; specific rights, labelling, advertising and platform requirements depend on the use case, authorisation documents, platform rules and legal counsel.
How do you check the portrait/voice rights chain?
Portrait/voice authorisation must state who authorises, for what, on which channels, for how long, whether exclusive, whether training/adaptation is allowed, and how to withdraw/delete. A real face/voice is sensitive personal information and usually needs separate consent.
The most common failure is "we have the founder's/employee's/influencer's photos and recordings, so we can start training". Portrait rights and voice interests are not company assets just because the material is internal. The authorisation should state: the authorising person, the authorised entity, purpose/scenario, channels (website, Douyin, WeChat Channels, live stream, overseas), term, territory, exclusivity, whether AI training/adaptation/synthesis is allowed, whether sub-licensing is allowed, and deletion after withdrawal/termination. A real face/voice is sensitive personal information under the PIPL, requiring a specific purpose, strict necessity, strong protection and separate consent; employees, minors and deceased persons have additional rules.
| Authorisation item | What to write | Evidence | Common gap |
|---|---|---|---|
| Authorising party | The person/guardian/rights holder; matches the material | ID document/authorisation | Treating employee photos as company assets |
| Purpose | Brand promotion/service/live/training; no out-of-scope use | Authorisation, project brief | "Global perpetual" is too vague |
| Channels/territory | Website/social/live/overseas/exhibitions | Channel list | Only "online" |
| Term/withdrawal | Start/end, withdrawal method, deletion after termination | Term, withdrawal record | No term/irrevocable |
| AI training/adaptation | Whether training, synthesis, adaptation and derivative works are allowed | Specific clause | Photo consent is not training consent |
| Sub-licensing/subcontracting | Whether vendors/platforms/distributors may use it | Subcontractor list | Vendor reuses the model |
| Minors/employees | Guardian consent/labour context/voluntariness | Guardian consent, employee communication | Mandatory face scan/recording |
The authorisation checklist is editorial; portrait/voice interests reference the Civil Code personality rights, and personal-information processing references the PIPL; specific consent depends on the facts and legal counsel.
How should deep-synthesis labelling work?
AI-generated/deep-synthesised digital-human video, audio, live streams and virtual scenes need explicit and implicit labels under the Labelling Measures and deep-synthesis rules. Do not remove, alter or conceal labels; platforms also have their own declaration rules.
Digital-human content is usually "AI-generated/synthesised" or "deep-synthesised" content. The Measures for Labelling AI-Generated Synthetic Content, effective 1 September 2025, require explicit labels (text/voice/graphics perceivable by users) on text, audio, images, video and virtual scenes, plus implicit labels in file metadata (generation attribute, service provider name/code, content ID). Users must declare when publishing, and no organisation or individual may maliciously delete, alter, forge or conceal labels. Digital-human live streams/short videos must also follow platform-specific AI declaration and labelling rules.
| Content form | Labelling requirement (editorial) | Evidence | Common failure |
|---|---|---|---|
| Digital-human short video | Prominent notice at the opening/around playback; metadata label | Screenshot/metadata | Only "AI" in the description |
| Digital-human live stream | Prominent notice in the room/frame; platform declaration | Live screenshot, platform record | Pretending to be a real anchor |
| AI voice/audio | Voice or rhythm notice at start/end/middle | Audio sample | No synthetic-voice label |
| Virtual scene | Prominent notice at the opening; appropriate position during service | Screenshot/recording | No virtual-scene notice |
| Download/export | File carries explicit + implicit labels | File inspection | Labels lost after export |
| Re-sharing | User declaration; platform verification/notice | Platform rules/records | Removing/cropping labels |
Labelling references the Measures for Labelling AI-Generated Synthetic Content and the Deep Synthesis Provisions; exact position/style/platform rules follow the original text and platform requirements.
What should you watch in digital-human advertising, endorsement and live selling?
Digital-human ads must be true and verifiable; do not fake experts, customers, medical results or data. A virtual persona can still count as advertising/endorsement. Live commerce must also follow platform and consumer-protection rules.
A digital human is not exempt from advertising law. If it recommends products, endorses a brand or sells live, the content must still be true and lawful: no false claims, no absolutes, no fabricated expert/patient/customer reviews, and no presenting AI-generated results/data as real cases. Using a real person's likeness for endorsement needs their authorisation and must be treated as an advertising endorser; a virtual persona must not imply unauthorised endorsement or make users think it is a real expert/customer. Platforms also impose filing, labelling, category and ad-review requirements on AI live streams/digital-human selling; confirm against the platform rules.
| Scenario | What to do | Evidence | Red flag |
|---|---|---|---|
| Brand endorsement | Real-person authorisation/virtual IP; no implied unauthorised endorsement | Authorisation, contract | "Founder recommends" without consent |
| Product explanation | Specs/price/results true; no absolutes | Product materials, review | "Lowest/best in the world" |
| Expert/doctor image | No fake expert/patient/results; medical ads follow rules | Qualification, ad review | AI posing as a doctor |
| Customer reviews | Real and verifiable; AI-generated reviews labelled/not misleading | Review records | Invented customer reviews |
| Live commerce | Platform filing/labels/category/consumer rights | Platform records | Unattended live stream without label |
| Data/results | Source, sample and definition verifiable; no fake screenshots | Report/raw data | Photoshopped/fake data |
The advertising/endorsement checklist is editorial; ad review, endorsement and special-industry (medical, financial) requirements follow the Advertising Law, industry rules and platform policies.
How should the vendor contract and training data be handled?
The contract must cover material/training-data rights, portrait/voice authorisation, labelling, ad review, data location/deletion, subcontracting, infringement indemnity and exit. "The vendor guarantees compliance" is not enough.
Digital-human projects typically involve likeness/voice capture, model training, video generation, live-streaming tools, platform operations and cloud services. The contract should cover: sources and rights of material/training data, scope of portrait/voice authorisation, whether training and reuse are allowed, labelling of generated content, responsibility for ad review, data location/retention/deletion, subcontracting, infringement/complaint handling, indemnity, suspension/exit, and ownership of material/models. A vendor saying "we handle compliance" is not enough; put obligations, evidence and owners into the contract.
| Clause | What to write | Evidence | Red flag |
|---|---|---|---|
| Material/training data | Source, rights, scope, whether training is allowed | Authorisation, data list | "Found online" |
| Portrait/voice | Person/guardian authorisation, term, channels, withdrawal | Authorisation | Photo consent only, no training consent |
| Labelling | Explicit/implicit labels, platform declaration, retained on export | Samples, platform records | Vendor removes watermark |
| Ad review | Who reviews, who is liable, special-industry qualifications | Review records | "Vendor says it is fine" |
| Data/deletion | Location, retention, deletion, access, subcontracting | DPA, logs | Training data kept forever |
| Infringement/complaints | Notice, takedown, indemnity, defence, insurance (if applicable) | Contract, SOP | All liability pushed to the client |
| Suspension/exit | Export of material/model/account/generated content; deletion | Exit checklist | Vendor disappears on exit |
The contract checklist is editorial and not a substitute for legal review; material/data/labelling/advertising/infringement/exit liability depends on the contract, platform and project.
Launch acceptance checklist
Accept eight items: rights chain, separate consent, labelling, ad review, training data, system security, contract, and monitoring/incident response. If any is missing, launch may become an infringement or false-advertising risk.
This checklist can be shared by brand, e-commerce/live commerce, customer service, legal, IT, procurement and vendors. The pass standard is not "does the digital human look real" but whether rights can be verified, labels are visible, ads are reviewed, data can be deleted and incidents can be handled.
| Check | Pass standard | Evidence | If failed |
|---|---|---|---|
| Rights chain | Portrait/voice/IP/material authorisations complete; scope/term/channels clear | Authorisation, contract | Suspend/obtain consent |
| Separate consent | Real face/voice has separate consent; minors have guardian consent | Consent record | Obtain consent/suspend |
| Labelling | Explicit + implicit labels; platform declaration; retained on export | Samples, metadata | Add labels/takedown |
| Ad review | True/verifiable; special-industry qualifications; no absolutes | Review records | Revise/stop |
| Training data | Lawful/authorised source; no out-of-scope use; deletable | Data list, DPA | Stop training/delete |
| System security | Access/encryption/logs/watermark/tamper protection | Config, logs | Remediate/downgrade |
| Contract | Infringement/complaint/indemnity/subcontracting/exit written | Contract, SOP | Add clauses/change vendor |
| Monitoring/incident | Public-opinion/complaint/infringement notice/takedown/correction process | Monitoring log, drill | Add process/training |
The acceptance checklist is editorial; specific rights, labelling, advertising, data and contract requirements depend on regulations, platform rules, authorisation documents and legal counsel.
How should the 90 days be sequenced?
In 90 days: classify and check the rights chain, then authorisation/consent, labelling/platform, ad review, contract/data, testing/drills and launch. Do not train the model first and fix authorisation later.
The typical failure is "build the avatar first, obtain authorisation later". Start by classifying (real person/virtual/voice/live), checking the rights chain and platform rules, then decide whether to capture/train. Authorisation, consent, labelling, ad review and contract can run in parallel, but testing, training and incident drills must be complete before launch. For real people/employees/minors, communicate early and explain the purpose, scope and withdrawal.
| Phase | Actions | Outputs | Owner |
|---|---|---|---|
| Weeks 0–2 | Classification, scenarios, platforms, material/training data, existing authorisation inventory | Classification table, risk list | Brand/legal |
| Weeks 3–4 | Portrait/voice/IP authorisation and separate consent; minor/employee communication | Authorisations, consent records | Legal/HR |
| Weeks 5–6 | Labelling/platform declaration, ad review, data/contract/vendor | Label samples, contract, DPA | Brand/procurement |
| Weeks 7–8 | Generation/live testing; complaint/takedown/incident drills; training | Test record, drill | Operations/legal |
| Weeks 9–10 | Launch; monitor labels/ads/public opinion/data; handle exceptions | Launch checklist, monitoring | Operations/compliance |
| Weeks 11–12 | Review; authorisation expiry/withdrawal; model/material update/deletion | Review, update plan | Brand/legal |
The 90-day plan is editorial; authorisation, labelling, advertising, platform, contract and launch timing depend on the company, vendor and platform rules.
Who should read this, who should not, and what decision does it support?
It fits companies using digital humans for brand, live commerce, customer service or training. It does not fit teams without portrait/voice authorisation or ad review, or those wanting to impersonate real experts/customers.
Who should read: marketing/brand, e-commerce/live commerce, customer service, legal/compliance, IT, procurement and vendors; teams building a digital human/virtual anchor/digital twin/AI voice-over; anyone deciding "can we use a real face/voice, how to authorise it, how to label it, how to review ads, how to contract". Who should not: projects without clear authorisation/consent that want to impersonate real experts/customers/medical results; teams without ad review/platform rules/incident processes; projects expecting one checklist to replace legal advice. Decisions supported: digital-human classification, portrait/voice authorisation, labelling, advertising/endorsement, training data, vendor contracts, a 90-day plan and launch acceptance.
Lingyi Information is a website and enterprise AI project services company, not a law firm, advertising-review body or platform. This article does not replace portrait/voice authorisation, ad review, platform filing or legal advice. Project scope is confirmed case by case; no platform approval, ad approval, traffic, conversion or result is promised.
FAQ
Do we need authorisation to train a digital human on the founder's/employee's photos?
Yes. Portrait rights are personality rights; being an employee or founder does not make the likeness a free company asset. The authorisation should state purpose, channels, term, whether AI training/synthesis is allowed, and withdrawal/deletion. A real face/voice is also sensitive personal information and usually needs separate consent.
What should we watch with AI voice cloning?
Voice interests are protected by reference to portrait rights. Obtain explicit authorisation from the voice owner/guardian, covering scenarios, term, channels, training/synthesis and sub-licensing. Do not train on audio found online, customer call recordings or a departed employee's voice.
Must digital-human video/live streams be labelled?
When the content is AI-generated/synthesised or deep-synthesised, apply explicit and implicit labels under the Labelling Measures and deep-synthesis rules; platforms also have declaration/labelling rules. Do not delete, alter or conceal labels. Confirm exact position/style against the original text and platform requirements.
Can a digital human be an advertising endorser?
It depends on the scenario and authorisation. A real person's likeness used for endorsement needs their authorisation and must follow advertising/endorsement rules; a virtual persona must not imply unauthorised endorsement or fake an expert/customer/medical result. Special industries (medical, financial) also face industry review and platform rules.
Does digital-human live commerce need platform filing?
Platform rules differ and usually involve AI-live/digital-human labelling, category, ad review, consumer rights and account qualifications. Confirm against the target platform before launch; do not use "unattended live streaming" to evade labelling or review.
Is "the vendor guarantees compliance" enough?
No. The contract should cover material/training-data rights, portrait/voice authorisation, labelling, ad review, data location/deletion, subcontracting, infringement/complaint/indemnity, and suspension/exit. The company still needs to keep authorisation, consent, review and monitoring evidence.
Bottom line
For an enterprise AI digital human or virtual anchor, check the rights chain first: a real face/voice needs authorisation and separate consent, AI-generated/deep-synthesised content needs labels, advertising/endorsement/live selling must be true and verifiable, and training data/vendor contracts must be explicit. Accept with a checklist before launch; looking real is not the first question—lawful use and incident handling are.
Method and limits
This article is public regulation, judicial interpretation and project method, not a compliance certification or legal advice for any company, digital human, platform, advertisement or contract. It uses no unpublished client material, authorisation or contract. The classification, tables and 90-day plan are editorial suggestions; portrait/voice authorisation, labelling, ad review, data and contractual liability depend on the regulations, platform rules, authorisation documents and legal counsel.
Sources
Information is current to 31 October 2026; each official/primary page governs its own scope, version and clauses.
- Measures for Labelling AI-Generated Synthetic Content (effective 1 September 2025): cac.gov.cn. Explicit/implicit labels, download/export, user declaration and the prohibition on deleting/altering labels.
- Provisions on the Administration of Deep Synthesis of Internet Information Services (effective 10 January 2023): cac.gov.cn. Notice/separate consent for face/voice editing, security assessment, labelling and filing.
- Personal Information Protection Law of the PRC: npc.gov.cn. Separate consent, impact assessment and minimisation for sensitive personal information such as faces and voices.
- Civil Code of the PRC, Book on Personality Rights: npc.gov.cn. Portrait rights, voice interests protected by reference, and the prohibition on making/using/publishing without consent.
- Advertising Law of the PRC (2021 amendment): SAMR. Advertising must be true and verifiable; no false/absolute claims or fabricated endorsement/proof.