AI Disclosure Taxonomy

Human Provenance in Film (HPF) is an initiative of The Mise En Scène Company (MSC), an international film sales agency with offices in London, New York, and Toronto. This taxonomy is its core framework.

A 2024 Deloitte survey of over 3,500 US consumers found that 70% would prefer to watch a film written by a human rather than by generative AI.¹

A 2025 Baringa survey found that 77% of consumers want to know if content they are watching was created by AI. For film and other creative content specifically, 31% say they would choose not to consume content wholly created by AI, and a further 41% would proceed with reservations.²

These are stated preferences. Audiences currently have no means to act on them. At the point of consumption, AI disclosure is absent from almost all streaming platforms, theatrical release, and acquisition documentation. Without a consistent disclosure standard, the commercial value of human authorship in film cannot be measured or demonstrated.

The absence of disclosure reflects a market failure, not audience demand for AI content. HPF exists to correct it.

The taxonomy gives financiers, producers, sales agents, distributors, broadcasters, platforms, and festivals a consistent way to understand AI use in any production. It works through standard deal documentation, from co-production agreements and chain of title to platform licensing and festival submissions.

The producer answers two questions of fact. First, is any AI output present in the finished film as distributed? If not, the film 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 legal weight sits in the warranty the producer signs. The test is just how they reach an answer they can stand behind.

2.1 Key terms

Two terms in this document have specific meanings. AI tool means any software that uses machine learning, neural networks, or generative models to produce, modify, or optimise content. Standard digital tools without machine learning components don’t count. Finished work means the film as distributed or exhibited, including all versions released after the original declaration was made.

A production receives one classification based on its highest category of AI involvement. The table below summarises all three. Full definitions follow. The terms AI tool and finished work used throughout this section are defined in section 2.1.

3.1 No AI Used

3.2 Assistive AI

3.3 Generative AI

3.4 The reconstruction test

Some AI tools operate on footage, audio, or images that already exist, rather than producing content from a prompt. 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 a human captured or created: noise reduction, de-flickering, artefact removal, upscaling that sharpens detail already in the source, format conversion. The output is the same content in better condition. This is Assistive AI.

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

The label a vendor applies does not decide the category. Where a tool both reconstructs and fabricates, the fabrication determines the category.

This taxonomy covers AI use across the making of the finished film, from development through to post-production, in every version distributed or exhibited. It does not cover what happens after the film is finished: distribution materials, marketing, and promotional content are out of scope. The table below sets out what is and is not in scope, which is open for further consultation.

In a co-production, the classification is the highest category used by any co-producer or contractor. This prevents co-production structures from hiding AI use that would otherwise be disclosed.

Where classification is unclear, apply the following test. Known situations currently under review are listed in section 8.

The line between Assistive and Generative AI is hardest where AI builds on human work but adds something new, particularly in VFX and animation. The test then is whether the AI processed human-created material or originated content that appears in the finished film. Because it turns on what is in the film, not on what a human might otherwise have done, it holds as tools and workflows change. The cases that genuinely sit between the two are logged in section 8 for v1.0.

HPF will publish guidance notes on significant edge cases as they arise. Known situations currently under review are listed in section 8.

No technical knowledge is required from producers or sales agencies. The disclosure is a paper declaration, the same kind producers already make for music clearances, location releases, and archive footage rights.

The producer completes an HPF AI Disclosure Form: three tick-boxes, a line for listing any AI tools used, and a signature. A sample form is available to view and print. You don’t have to use that form: any document with the classification, production details, and a signature works. The form below shows the minimum structure.

The process below applies whether or not a sales agency is involved.

7.1 Deal documentation

AI disclosure under this taxonomy is recorded as a producer declaration in deal documentation. This applies whether the vehicle is chain of title (where a sales agency is involved), a platform licensing agreement, or a festival screening agreement. The receiving party reviews the declaration when the film is acquired or submitted.

Advantages of a declaration-based approach over technical marking alone:

1. Fits existing agreements: A false declaration is covered by existing contracts. No new legal machinery is needed.
2. Survives the file: The document exists independently of the delivery file and survives transcoding, format conversion, and platform ingest.
3. Works for all routes: Sales agency, direct platform, festival, and broadcaster.

7.2 C2PA as a technical complement

HPF is proposing a custom assertion to the C2PA working group that would allow the classification to be carried as a Content Credential directly in the delivery file. This is not yet part of the C2PA specification. The proposal is open for comment via the GitHub repository.

Where technically achievable and once the assertion is adopted, post-production facilities and delivery houses would be able to embed the HPF classification as a C2PA Content Credential directly in the delivery master. Content Credentials embed provenance data in the file itself and survive platform ingest and transcoding.

C2PA credentials are entirely optional. The HPF AI Disclosure Form and the deal documentation declaration are the complete mechanism. Producers and sales agencies never need to use C2PA tools.

The schema below shows the proposed data structure and is stable at v0.9. Both fields are required when an HPF record is present. Where no declaration exists, omit the record entirely: do not write null values and do not default to no_ai. An absent record and a declared no_ai are not the same thing.

{
  "hpf_taxonomy_version": "0.9",
  "hpf_classification": "no_ai" | "assistive_ai" | "generative_ai"
}

A draft C2PA custom assertion (hpf.film.ai_disclosure) is proposed for carrying the classification in a Content Credential alongside c2pa.actions. Full schema, C2PA mapping, and implementation guide: GitHub repository.

7.3 Audience disclosure formats

Disclosure to audiences may take any of the following forms, at the distributor’s or platform’s discretion. The category displayed does not vary by format:

• Start-credits or end-credits statement
• Press notes and production information
• Platform label on the film detail page
• Delivery metadata surfaced at point of play

The following situations require further guidance. All guidance in this table is provisional and may be revised in v1.0. Responses can be submitted via section 9.

HPF is inviting responses before finalising v1.0. Partial responses are welcome: indicate which question or questions you are addressing, and where possible include your name, organisation, and the section your feedback relates to. There is no word limit. Technical amendments can go directly to the GitHub repository.

1. Q1Does the present-and-originated principle, and the three categories it produces, capture the distinction that matters to your organisation? Is anything missing, such as a fourth category?
2. Q2Is the declaration mechanism workable in your part of the industry, and if not, what would need to change? In particular, for a No AI Used declaration, software now switches AI features on by default, so would a "reasonable enquiry" standard, the same one used for other chain-of-title warranties, be a fair basis?
3. Q3Two scope boundaries are open. Should any pre-production AI use that never reaches the screen still be disclosed? And how urgent is it to bring marketing materials, such as trailers, posters, and social cuts, into scope?
4. Q4Are the edge cases in section 8 classified correctly, and are there other situations that need guidance before v1.0? Animation, and content built on human work but newly synthesised, such as in-betweening and style transfer, is where we most want input.
5. Q5For platforms, broadcasters, and distributors: would you use the Assistive AI classification, and would you surface it to audiences?
6. Q6Which ways of showing the disclosure are practical and clear, from end credits to delivery metadata to a platform label, and what would meet your regulatory obligations?
7. Q7Does the chain-of-title mechanism work outside the UK and US, where co-production paperwork differs? Are there regulatory, contractual, or collective-bargaining frameworks HPF needs to account for in v1.0?

All version changes are published. Future changelogs will document substantive amendments by section number.

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

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