AI Disclosure Taxonomy

Human Provenance in Film (HPF) is a free, open standard for declaring AI output present in finished film and television productions. It gives audiences and receiving organisations a common record of how AI was used. HPF was originated by The Mise En Scène Company (MSC), which currently maintains the standard and holds its copyright. The standard is published under CC BY 4.0.

This document sets out the three-category classification, the producer declaration and the machine-readable record. HPF is a self-declaration standard. It does not certify productions, audit workflows, or determine consent, rights or regulatory compliance.

The producer answers two questions of fact about the version the declaration names. Is any AI output present in that version as distributed? If not, it is No AI Used. If it is, did the AI process human-created material, which is Assistive AI, or originate new content that appears in the finished work, which is Generative AI? The producer signs the declaration and is responsible for its accuracy. HPF does not verify it.

HPF classifies how AI was used, not the underlying model or product. The same AI system can support an Assistive use in one production and a Generative use in another.

2.1 Key terms

Artificial intelligence, for the purposes of this taxonomy, means software systems that use machine learning, neural networks, or similar techniques trained on data to generate, synthesise, enhance, or analyse content, as distinct from deterministic software that applies fixed rules or algorithms without learned models. Where a tool offers both conventional and AI-powered modes, only the AI-powered mode’s output is in scope. AI tool means any software that uses artificial intelligence, as defined above, to produce, modify or optimise content. Standard digital tools that do not use it are not AI tools. Released version means a separately released editorial edition or substantive localisation of a work, identified by its picture, sound and, where relevant, embedded audience-facing text. Finished production, also called the finished film or the finished work, means the released version a declaration covers, as distributed or exhibited. This taxonomy covers film and television.

2.2 Released versions and declarations

One signed declaration covers one released version of a work and may travel with every delivery of that version. A later released version, such as a later editorial edition or a substantive localisation, has its own signed declaration, even where the classification and factual summary are unchanged; the new declaration may repeat those facts. It does not change the record for the earlier version; the earlier declaration continues to cover its version. HPF does not require the producer to notify recipients of an earlier version, although contracts or distribution arrangements may.

How this rule applies to episodic works, to selectable subtitles and captions, and at the boundary between technical and substantive localisation is open under question 4 in section 9. Classify and declare your production gives informative examples of releases that normally do and do not need a new declaration.

A production receives one classification, at the highest category that applies to any element of the released version. There is no percentage threshold and no de minimis exception. Where in-scope AI output is present, the declaration includes a concise factual summary of what the AI did and what output reached the released version. The table summarises the three categories; full definitions follow.

3.1 No AI Used

3.2 Assistive AI

3.3 Generative AI

3.4 The reconstruction test

Some AI tools work on footage, audio, or images that already exist. For these, classification turns on whether the tool reconstructs content already present or fabricates content that is not.

Reconstruction restores, cleans, or repairs content that was captured or created by a human, such as de-flickering, artefact removal, or format conversion. The output is the same content in better condition. This is Assistive AI.

Fabrication invents content that was never captured or created, such as painting in a background revealed by removing an object, generating frames or detail with no source, or synthesising a performance element the performer did not give. This is Generative AI.

Classification turns on this distinction, not on whether the software technically synthesised pixels, frames or audio, and not on the label a vendor applies. Where a tool both reconstructs and fabricates, the fabrication determines the category.

Two kinds of case remain open. Whether established machine-learning functions such as denoising, upscaling and tracking require Assistive AI disclosure at all is the classification-scope part of consultation question 1. Hybrid restoration, where repair and newly originated content are hard to separate, and the treatment of automated subtitles and captions in a released version, are part of question 4. See section 9.

3.5 Descriptors

Descriptors add factual context about the type of AI output present in the finished work. They never determine or change the classification and carry no threshold. More than one descriptor may apply where each records a distinct fact. Altered performance may be recorded on an Assistive AI or a Generative AI production. Every other descriptor applies only to Generative AI. None applies to No AI Used.

In the second example each descriptor records a different fact: Digital replica, that the generated material represents an identifiable person; Generated voice, that spoken content was generated; Altered performance, that a performance the actor gave on camera was modified. Generated writing spoken by a human actor in the usual way is different: it is Generated writing, not Generated voice or Digital replica.

Descriptors are factual provenance terms. Where a descriptor resembles a term in a collective agreement or contract, it takes its meaning from this standard alone. Descriptors do not state whether consent was obtained or whether rights, contractual or collective-agreement requirements were met, and they do not replace contracts, releases, consent records or chain-of-title documentation. Consultation questions 2 and 3 ask whether the definitions, examples and combination guidance allow consistent use.

This taxonomy covers AI output reflected in the named finished version as distributed or exhibited, from development through to post-production.

AI used only in development, administration, production management or ideation, and unused concept material, are outside the finished-work record where no AI output reached or is embodied in the named finished version. This does not prevent a No AI Used declaration. Development-stage AI is in scope where its output is embodied in the finished version, for example generated dialogue performed in the final film, or pitch visuals reused as a title sequence. Influence alone is not incorporation. Human modification of an AI output does not by itself answer whether that output reached the finished version; the question is what, if anything, was incorporated into the version being declared.

AI used to make marketing or distribution materials is outside scope unless its output is also incorporated into the finished film or programme.

The production company or lead producer records one classification for the finished production, using the highest applicable category. Who produced a given element does not change the classification. Rights, consent, contracts, insurance and legal compliance remain outside HPF; organisations address them through their own processes. HPF does not assess the licensing of AI training data.

Where classification is unclear, use the rows below. Situations still under review are listed in section 8.

The test turns on what is in the finished film, not on whether a human would otherwise have performed the function. AI-generated content that appears in the finished work is Generative AI whether or not a human could have produced it.

Where the answer is unknown or the records are incomplete, the classification is unresolved. Missing, uncertain or absent information is never No AI Used. A No AI Used declaration records that no in-scope AI output was found, not that nothing is known.

No technical integration is required. The declaration is a signed producer document, on paper or digital. The HPF AI Disclosure Form can be used, or any document with the same information. The form below shows the minimum structure.

A declaration identifies the work, the released version it covers, the classification, the party making it, and the date it was made. It is signed by someone authorised to declare on behalf of the production. Where in-scope AI use is present, it carries a short factual summary of what the AI did and what output reached the finished work. Tool names can support the summary but do not replace it.

An authorised party may make a declaration for a production already released, where it can establish the relevant facts. The declaration carries the date it was made, not the release date. No one may infer a classification from a release date, detection output or incomplete catalogue information, or apply one in bulk across a catalogue. Where no HPF record exists, that is not a No AI Used classification. An absent declaration establishes neither that AI was used nor that it was not used.

7.1 Deal documentation

The signed declaration is the complete mechanism. It can be supplied with existing deal, delivery and other production paperwork, such as a sales agency’s delivery package, a platform licensing agreement or a festival submission, and it exists independently of the delivery file. HPF standardises the information in the record. Each receiving organisation decides what to do with it.

7.2 Machine-readable record and C2PA

For a record made under 0.9.3, all three fields are required when an HPF record is present. A record made under 0.9.2 carries the standard version and the classification only and stays valid under the version it names. Where no declaration exists, omit the record: do not write null values and do not default to no_ai.

{
  "hpf_standard_version": "0.9.3",
  "hpf_released_version": "Theatrical cut",
  "hpf_classification": "no_ai" | "assistive_ai" | "generative_ai"
}

The version field was named hpf_taxonomy_version before 0.9.2. A provisional descriptor array, hpf_descriptors, is also proposed; do not build a production integration against it yet. The core fields may be tested in integration during consultation.

C2PA can carry an HPF classification in a Content Credential. It does not decide or validate it, and nothing in HPF depends on C2PA. HPF is developing an optional entity-specific mapping, working label film.humanprovenance.ai-disclosure, which is not part of the C2PA specification and has not been reviewed or endorsed by C2PA. The integration guide and C2PA mapping have the detail.

7.3 Audience disclosure formats

HPF encourages platforms and distributors to display the classification to audiences, but does not require display or prescribe a label, format or placement. Where it is shown, for example in credits, press notes, a platform title page or delivery metadata surfaced at point of play, the category does not vary by format.

These situations need judgement or remain open. All guidance in this table is provisional and may be revised in v1.0. Responses can be submitted through section 9.

HPF is inviting responses before finalising v1.0. Partial responses are welcome: say which question you are addressing and, where possible, include your name and organisation. Technical amendments can go directly to the GitHub repository.

Current v0.9.3 rules

v0.9.3 establishes the current consultation-draft rules below. Like all pre-v1 provisions, they remain subject to evidence received during consultation, but they are not listed as unresolved questions.

• No percentage threshold: the highest applicable category applies to any AI output present, however small (section 3).
• More than one descriptor may apply where each provides distinct factual information (3.5).
• One declaration covers one released version, and a later released version has its own declaration (2.2).

Outside the consultation

• HPF does not determine contractual permission, consent, compensation, rights, or legal compliance, although its terminology may be useful in conversations between producers, performers, representatives and other parties.
• HPF does not currently prescribe an audience-facing label, display format, or placement rule.

Remaining questions for consultation

1. Q1Reasonable enquiry and established machine learning. This question has two parts. Neither is settled in v0.9.3, and concrete workflow examples are welcome for both.
   Declaration basis: Is reasonable enquiry the appropriate basis for a No AI Used declaration where AI functions may be built into software, enabled by default, or difficult to identify? Reasonable enquiry is not a current HPF requirement.
   Classification scope: Which established machine-learning functions, if any, should require Assistive AI disclosure when they process picture or sound retained in the released version?
2. Q2Performance-related descriptors. Are the current performance descriptors sufficiently clear for altered performances, digital replicas, synthetic performances, and combined uses? What examples are needed to support consistent application? This includes when Generated Visual Content should be used alongside Altered performance, Digital replica, Synthetic performance, or other descriptors.
3. Q3Human contribution to synthetic performance. How should HPF describe cases where human performance inputs contribute to a synthetic performance without representing the performer as an identifiable person? This may be resolved through guidance, factual-summary practice or worked examples; a new descriptor is not assumed.
4. Q4Complex finished-work and version cases. What further guidance is needed for transformed output, intermediate VFX control data, restoration and reconstruction, localisation, subtitles and captions, episodic releases, animation workflows, and other cases where the relationship between AI output and the released version is difficult to determine? How HPF should treat AI-generated or AI-segmented mattes, masks, or similar control data that shape a delivered composite is not settled in v0.9.3. Questions about episodic works, selectable subtitles and captions, dubbing and other localisation, and the boundary between a technical delivery variation and a substantively different released version concern applying the released-version rule, not reconsidering it.
5. Q5Implementation. What information can producers reasonably establish when preparing a declaration, and what information is useful to organisations receiving it?

If a real production use does not fit the current categories, we want to hear about it.

Earlier drafts asked some of these questions in other words and numbered them differently. The wording as asked is preserved in each release tag, and the version history records where each now sits.

Every revision of the standard, what it changed and a link to its release tag are recorded in one place: the version history in GOVERNANCE.md. This is version 0.9.3, a draft for consultation. Earlier versions remain as they were released.

Illustrative classifications. Each shows the category and the reason it applies.

The standard is published in human and machine-readable forms.

Definitions of terms used around AI disclosure in film. Terms defined in the standard itself are in section 2 and section 3.

Different kinds of transparency

• Regulatory transparency: disclosure or marking required by law, such as Article 50 of the EU AI Act.
• Technical provenance: records of a digital asset’s origin and history. Content Credentials can carry such records. Watermarking and fingerprinting can support identification or link content to records, but do not necessarily provide a provenance history themselves.
• Consent and compensation: rights mechanisms, such as collective-bargaining provisions, licences and permissions, that govern whether material, performances or data may be used and on what terms.
• Production disclosure: information about how a finished work was made, supplied by the party responsible for it. HPF provides this form of production disclosure. The producer’s signed declaration is the record, and existing deal, delivery and other production paperwork is the route by which it reaches the organisations and audiences that use it.

HPF complements technical provenance and regulatory disclosure and does not replace them. It does not determine consent, compensation or underlying rights.

How information is recorded

How information is communicated

Alignment with existing definitions

Where recognised definitions for these terms exist or emerge in the creative sectors, HPF will seek to align with them rather than create a competing vocabulary. The glossary remains under consultation. Reference to an external standard, technology, organisation or regulatory framework does not imply coordination with or endorsement by that body.