백서

시청각 저작물의 AI 사용 공개

생성 이미지, 합성 연기자와 합성 음성, 실존 인물의 초상, 사용 도구를 체계적이고 검증 가능한 방식으로 공개하는 방법을 제시합니다.

제1.0판발행일: 2026년 11월 21일읽는 시간 14분Lire en français
이 문서는 프랑스어와 영어로 발행됩니다. 영어판을 기준본으로 하며, 아래 본문은 영어판입니다.

01Executive summary

Generative tools can now contribute to almost every stage of audiovisual production, from concept art to backgrounds, performers, voices and music. Audiences, festivals, buyers, performers and regulators need to know how AI was used in a work. What they usually receive is free text — a line in a press kit, a box on a submission form — that is hard to compare, impossible to verify and easily lost when the work changes hands.

This paper sets out a practical framework for disclosing AI use, as implemented in INDEX.FILM — International AI Audiovisual Asset Registry. Every INDEX record carries a structured AI disclosure: a short summary; whether imagery was generated; whether there are synthetic performers or a synthetic voice; whether a real person's likeness is used and, if so, with consent; and the tools used. An optional AI-use policy states the rights holder's position — allowed, prohibited or negotiable — on training, fine-tuning, style or character reference, and voice or likeness.

The disclosure is attributed to the registrant, versioned and time-stamped. It can be reviewed by an accredited registrar under the AI disclosure review scope, supported by consent evidence referenced by hash, and addressed in a Rights Clearance Schedule at the CLEARED level. It complements provenance embedded in files, such as C2PA Content Credentials, and can help those subject to transparency obligations, such as those in Article 50 of the EU AI Act, without discharging those obligations by itself.

A disclosure states facts about how a work was made. It is not a judgment on the work's merit, authorship or ownership. INDEX records registration facts and verification scope, supporting the recognition of rights worldwide.

02Why disclosure matters

Each party that handles a work asks a slightly different question about AI use. A disclosure model must answer all of them from one statement.

PartyWhat they need to knowWhy it matters to them
AudiencesWhether what they see and hear was generated, and whether real people are depicted syntheticallyTo interpret the work, especially where it resembles documentary or news footage, and to trust those who present it
Festivals and marketsHow AI was used, in a form comparable across submissionsTo apply eligibility and category rules consistently, inform juries and present works transparently
Buyers, distributors and platformsWhich elements are generated, with which tools, and whether consents existChain of title, insurance, platform policies, labelling duties and reputational risk
Performers and their representativesWhether a voice or likeness has been synthesised, and on what consentTheir personality and performers' rights, and the integrity of their professional identity
Regulators and competent authoritiesWhether content is AI-generated or manipulated, and who is responsible for the statementTo supervise transparency obligations where they apply

Disclosure also protects creators. A precise, early, attributed statement of how a work was made is the best answer to later suspicion, and a time-stamped record of what was declared, and when, can help if questions arise about the origin of material.

03Principles

PrincipleMeaningIn the INDEX model
SpecificSay what was generated, not merely that AI was usedSeparate fields for imagery, performers, voice and likeness
StructuredFixed fields and values that can be compared and processedYes/no fields, a three-value likeness field, a tools list, a three-value policy
AttributableName who made the statement and whenDeclared by the registrant through a named registrar; versioned and time-stamped
VerifiableLet a competent party check the statement and say soAI disclosure review scope; evidence referenced by hash; CLEARED schedules
ProportionateAsk for what matters, at a level of detail suited to the workA short required core; an optional policy; verification only where needed
Privacy-respectingAvoid publishing personal data or confidential materialConsents referenced by hash, never published; registrant held as an opaque reference

These principles pull in different directions: specificity calls for more detail, proportionality and privacy for less. The model resolves the tension with a short, categorical core, a free-text summary for nuance, and evidence kept out of the public record as hash references held by the parties.

04The INDEX disclosure model

The AI disclosure is part of every INDEX record; clause 5 of the INDEX Standard 1.0 sets out which elements are required at creation and which from BOUND. The disclosure describes the work as registered — the cut whose fingerprints are bound — and any change creates a new, time-stamped version of the record. The guidance below is informative; the fields themselves are defined by the Standard.

FieldValuesGuidance
SummaryFree text, up to 500 charactersA plain account of how AI was used and to what extent: which stages (for example concept, backgrounds, animation, voice, music) and which parts of the work. No promotional language and no personal data.
Generated imageryYes / NoYes if images or moving images in the work as registered were produced or substantially altered by a generative model. Routine corrections such as noise reduction do not by themselves call for Yes. When in doubt, answer Yes and explain in the summary.
Synthetic performersYes / NoYes if any on-screen performance is carried by a generated or synthetic performer rather than by a human performer recorded for the work. Explain borderline cases, such as a human performance captured and applied to a generated character, in the summary.
Synthetic voiceYes / NoYes if any dialogue, narration or singing is generated or cloned, including conversion of one speaker's voice into another's.
Real-person likenessnone / consented / unspecifiedWhether the work uses a synthetic or AI-altered depiction of the appearance or voice of an identifiable real person. None: it does not. Consented: it does, with the consent of each person concerned or of those entitled to give it. Unspecified: the registrant does not state the position.
ToolsList of up to 30 entriesThe principal generative tools and models used, with versions where known. Naming a tool is not an endorsement and implies no relationship with its maker.

Reading “unspecified”

“Unspecified” should be the exception. It tells readers that the question has not been answered, not that the answer is no, and it invites scrutiny. A registrant who cannot confirm consent for every person concerned should say so rather than choose “consented”.

Writing the summary

  • Useful: “Exterior backgrounds generated from the director's sketches and composited with live-action plates; all performances by human actors on set; no voice synthesis.”
  • Too vague: “AI-assisted film.”
  • Promotional: “A ground-breaking fusion of human artistry and cutting-edge AI.”

Uses that leave no generated material in the work as registered, such as research or early concept exploration, need not appear in the structured fields, but may be mentioned in the summary where they matter to readers.

Borderline cases

SituationSuggested treatment
AI upscaling, noise reduction or colour matching of filmed footageGenerally not generated imagery; mention in the summary if material to the look of the work
Generated storyboards or concept art that do not appear in the workNot in the structured fields; may be mentioned in the summary
Dubbing that clones the original actors' voices into other languagesSynthetic voice: Yes; real-person likeness: consented, where each actor has consented
Digital de-ageing of an actorGenerated imagery: Yes; real-person likeness: consented; synthetic performers: No, as the performance is human

05The AI-use policy

The disclosure looks back: how was AI used to make the work? The AI-use policy looks forward: how may AI be used with the work? It is optional and states the rights holder's position on four uses.

UseCovers
TrainingUsing the work, or material from it, to train AI models
Fine-tuningUsing the work to adapt or specialise an existing model
Style or character referenceUsing the work as a reference to reproduce its visual style or its characters in new output
Voice or likenessUsing voices or likenesses appearing in the work to generate new performances

Each use is marked allowed, prohibited or negotiable, expressed in the Standard as permitted, reserved or by agreement. Negotiable signals that the use may be licensed by agreement with the rights holder.

The policy is a published statement of the rights holder's position, attributed like every other field. It is not a licence: it grants no right by itself, and a use marked allowed remains subject to applicable law and to the rights of others, such as performers whose voices appear in the work. Nor does it replace any reservation of rights that the law of a particular jurisdiction may require in a specific form; rights holders should also make such reservations where and as the applicable law requires.

A rights holder can state a position only on what it controls. Where performers' consents are limited, the voice-or-likeness position should be no more permissive than those consents.

06Verifying disclosure

A disclosure is declared by the registrant, and the record says so. Three mechanisms allow it to be checked and the check to be recorded; a fourth allows it to be challenged.

AI disclosure review

AI disclosure review is one of the five verification scopes. An accredited registrar whose accreditation covers it may review the disclosure against the material made available — production records, generation logs, project files, contracts and consents — and sign a VerificationCredential stating the scope, basis and method, date, reviewer reference and hash references to the evidence. The disclosure is then shown as verified by that registrar; if it changes later, it returns to declared until reviewed again. Review checks consistency with evidence. It is not a forensic guarantee that no undisclosed use exists, and the credential states what was examined so that readers can judge its weight.

Consent evidence referenced by hash

Consent to the use of a person's likeness or voice is personal and often confidential, and INDEX never publishes it. A hash of the consent document enters the record as an evidence reference, while the original remains with the issuing party. The hash reveals nothing of the content, but if the document is later produced — to a buyer under confidentiality, an insurer, a panel or a court — anyone can confirm that it is the document relied on, unaltered. A “consented” statement supported by reviewed consents carries verified status; without review, it remains the registrant's declaration.

CLEARED schedules

At the CLEARED level, an accredited law firm attaches a Rights Clearance Schedule stating territory, media, term, exclusions and conditions, together with the scope of its review. Where counsel's review covers AI-related matters — for example consents for synthetic voices or likenesses, or the terms on which generative tools were used — the scope of review says so, and any resulting exclusions or conditions are stated in the schedule.

Challenge

Anyone who believes a disclosure is materially false or misleading may file a complaint under the INDEX Dispute Resolution Policy (IDRP), which expressly covers record data including AI disclosure and real-person likeness. Where a complaint appears well-founded on its face, the record is publicly flagged as disputed within 2 business days. Remedies include correcting the record in a new version, suspending or revoking an attestation, and revoking a certificate or the record.

Status of a disclosureWhat a reader can rely on
DeclaredThe registrant's statement, attributed and time-stamped
VERIFIED, with AI disclosure reviewAn accredited registrar has checked the disclosure against stated evidence
Addressed in a CLEARED scheduleReviewing counsel's scope of review covers AI-related matters, with conclusions, exclusions and conditions stated

07Relationship with other mechanisms

A registry disclosure is one of several mechanisms. Each has strengths and limits, and they work best together.

MechanismWhere it livesStrengthsLimits
INDEX recordOutside the file, at a permanent addressDescribes the work as a whole; attributed, versioned and verifiable; reachable by ID or fingerprint even when file metadata is lostDoes not see inside a file's editing history
C2PA Content CredentialsEmbedded in or bound to a fileSigned provenance of a specific file, including generation and editing steps, recorded by the tools and organisations involvedMay be lost when files are re-encoded or metadata stripped; describes a file, not the work's overall disclosure or rights position
WatermarksImperceptibly in images or soundCan survive some transformations and signal machine generationRobustness varies by technique; carries little information
On-screen labels and creditsIn the work or its presentationImmediately visible to audiencesNot machine-readable; easily inconsistent across versions and territories
Submission and delivery formsIn each recipient's systemsTailored to the recipient's rulesNeither portable nor verifiable; answered again for every recipient

C2PA and INDEX together

C2PA Content Credentials and a registry record are complementary. A C2PA manifest can show that a particular shot was produced by a particular generative tool; the INDEX disclosure states, for the work as a whole, that generated imagery is present and where. If a manifest is stripped from a distributed copy, the copy's fingerprint — or, for re-encoded copies, an ISCC code where one is bound — can still lead to the record.

Producers who use Content Credentials should keep the work-level disclosure consistent with the manifests in their files, and compute the fingerprints bound to the record on the final files, since embedding a manifest changes a file's bytes.

08Regulatory context

Disclosure of AI use is also a subject of law and regulation. This section gives orientation only; it is not a statement of the law of any jurisdiction.

European Union

Regulation (EU) 2024/1689, the Artificial Intelligence Act, contains in its Article 50 transparency obligations concerning AI-generated or manipulated content, applicable from 2 August 2026. These obligations are addressed to providers and deployers of certain AI systems. How they apply to a particular work, and to the parties involved in making and distributing it, depends on the facts and on the text as interpreted by competent authorities and courts. Readers should consult the current text, together with any amendments, guidelines and codes of practice, and take qualified advice.

Other jurisdictions

Rules on labelling synthetic content exist or are under discussion elsewhere, and they differ in scope and terms. Personality and image rights, performers' rights and data protection law may apply independently of any AI-specific rule.

What the record can and cannot do

An INDEX record does not by itself discharge any legal obligation. Its fields are deliberately neutral: they record facts relevant under many frameworks without presuming the legal characterisation that any one of them applies. What the record offers to those who carry obligations is a structured, attributed and time-stamped source for the information they must convey, and a record of what was disclosed, by whom and when.

Within the Alliance, the Ethics and AI Transparency Committee keeps the disclosure rules under review, and any normative change to the fields follows the Standards Committee's public comment process.

09Recommendations

For creators and producers

  1. Keep production records from the start: which tools and models were used, for which sequences, on what terms. Disclosure written from records is easier and more credible than disclosure written from memory.
  2. Disclose at registration, and update the record when the work changes; a new cut with new generated material calls for a new version.
  3. Be specific in the summary, avoiding both understatement and promotional language.
  4. Obtain written consent before synthesising any real person's appearance or voice, scoped to the intended uses, and keep it so that its hash can be referenced and the original produced when needed.
  5. Use “unspecified” only for a genuine reason, and explain it in the summary.
  6. Consider stating an AI-use policy consistent with your contracts, especially performers' consents.
  7. Ask your registrar for AI disclosure review where festivals, buyers or platforms will rely on the disclosure.

For festivals and markets

  1. Ask for the INDEX ID at submission and read the disclosure from the record instead of asking the same questions again.
  2. Define eligibility and category rules in terms of the disclosure fields, so that decisions are consistent and explainable.
  3. Present disclosure to juries and audiences consistently, drawing on the summary.
  4. Check the record's status and the current certificate at index.film/verify before selection and before screening.
  5. As accredited partners, festivals can issue Certificates of Honour linked to the record.

For platforms, distributors and broadcasters

  1. Ingest the disclosure and AI-use policy through the Registry API or the JSON-LD record, and map them to your own labelling and catalogue fields.
  2. Match deliveries to records by SHA-256 fingerprint, and use ISCC codes, where bound, to support the matching of re-encoded copies.
  3. Distinguish declared from verified disclosure in risk assessment; for sensitive content, ask for AI disclosure review or a CLEARED schedule that addresses AI-related matters.
  4. Preserve Content Credentials where possible, and keep audience-facing labels consistent with the registered disclosure.
  5. Take the AI-use policy into account in your own uses of the work, subject to your contracts.
  6. Report a disclosure you believe to be materially false to legal@index.film under the IDRP.

10Worked examples

The examples are illustrative: they show how the fields might be completed for three typical works, and are not records of actual works.

FieldA. Fully generated short filmB. Live-action film with AI-generated backgroundsC. Animation with the synthetic voice of a consenting actor
SummaryAll imagery and animation generated from the director's prompts and sketches, then edited and graded by the team. No dialogue; narration by a human narrator; music composed and performed by human musicians.All performances filmed with human actors against blue screen. Backgrounds of exterior scenes generated with an image model and composited by the visual effects team.Characters animated by the studio's animators. The lead character's voice is synthesised from recordings of an actor, with the actor's written consent for this work; all other voices performed by human actors.
Generated imageryYesYesNo
Synthetic performersYesNoNo
Synthetic voiceNoNoYes
Real-person likenessnonenoneconsented
ToolsThe image and video generation models used, with versionsThe image generation model and compositing software usedThe voice synthesis tool used
AI-use policy (optional)Training: negotiable; fine-tuning: negotiable; style or character reference: prohibited; voice or likeness: not statedTraining: prohibited; fine-tuning: prohibited; style or character reference: negotiable; voice or likeness: prohibitedTraining: prohibited; fine-tuning: prohibited; style or character reference: negotiable; voice or likeness: prohibited
Verification to considerAI disclosure review against project files and generation logsAI disclosure review against visual effects records; file integrityAI disclosure review including the actor's consent, referenced by hash; for distribution, a CLEARED schedule whose scope of review covers the consent
  • Example A: a fully generated work still has human contributions worth stating — direction, editing, narration, music. The summary is where that balance is described.
  • Example B: generated imagery is Yes although every performance is human. The summary prevents misreading by saying exactly where generated imagery appears.
  • Example C: a synthetic voice and consented likeness go together. Voice or likeness is marked prohibited because the actor's consent covers this work only, and the rights holder cannot permit more than it holds.

11Conclusion

Disclosure of AI use is most useful when it is specific, structured, attributable and verifiable, and when it is made once, at the source, and travels with the work. The INDEX model provides a short required core, a free-text summary for nuance, an optional forward-looking policy, and three levels of confidence — declared, reviewed, and addressed by counsel — each visible to anyone who checks the record.

Disclosure is not a verdict. It does not decide whether a work is art, who its author is or who owns it. It states how the work was made, names who said so and records who has checked. INDEX records registration facts and verification scope, supporting the recognition of rights worldwide.

12About this paper

Disclosing AI use in audiovisual works: a practical framework. Edition 1.0, published 21 November 2026 by the International AI Audiovisual Asset Index Standards Alliance, “the Alliance”, an international non-profit organisation with its seat at Avenue Louise 390, BE-1050 Ixelles, Région de Bruxelles-Capitale, Belgique.

This white paper is published for information only. It describes the AI disclosure and AI-use policy defined in the INDEX Standard 1.0 (index.film/standard), which is normative and prevails in case of difference. English is the reference edition. References to legislation are for orientation and reflect the text cited, not its application to any particular case. Related reading: the white papers on the INDEX model (index.film/knowledge/white-papers/index-model) and on the recognition of rights (index.film/knowledge/white-papers/rights-recognition).

This paper is not legal advice. For advice on a specific matter, consult a qualified lawyer.

신청

글로벌 등록 네트워크에 참여하십시오.

시청각 저작물의 AI 사용 공개 · INDEX.FILM