August consultation update

The three categories and the organising principle have not changed since July. This month we have published revision 0.9.2, added a provisional descriptor layer for consultation, and opened a short survey to test the standard against real work. We have also completed an international review of published organisational disclosure requirements and begun structured research into how practitioners discuss AI use in production.

What we have heard

• Keeping the top-level classification simple remains important, but people want a way to distinguish materially different kinds of generative use beneath it.
• A percentage threshold still appears difficult to apply: current production tools do not provide a consistent measure of how much of a frame, shot, sequence or finished work an AI system affected.
• The boundary between Assistive and Generative AI remains the hardest part of the standard to apply, particularly where modern tools generate pixels, frames or audio while working from existing human material.
• Different parts of the industry need different levels of detail. A useful disclosure record must travel between them without attempting to replace their own policies, contracts or decisions.
• Published organisational requirements confirm that AI disclosure is already becoming part of screen-industry paperwork, but the questions and definitions vary widely.

These points draw on exploratory conversations across production, workforce interests, industry infrastructure and adjacent creative-sector initiatives. They were exchanges and implementation feedback, not formal reviews of HPF, and they do not imply endorsement, participation in the consultation or any commitment to adopt the standard.

On descriptors: more detail without more categories

Our July position on thresholds stands. HPF does not set a minimum quantity of qualifying AI-generated content before the Generative AI category applies: if that content is present or embodied in the finished production, the classification does not change because the use was brief or small.

Instead, revision 0.9.2 introduces a provisional descriptor layer. The category says whether generated content is present; descriptors say what type of content is involved. They do not determine or change the classification, carry no quantity threshold and are not a fourth category.

The terms under consultation are:

• Altered performance
• Digital replica
• Synthetic performance
• Generated voice
• Generated writing
• Generated music
• Generated visual content

Altered performance is the one descriptor that may also be recorded for an Assistive AI production. Its presence alone does not establish that the AI originated new performance content.

The taxonomy page contains the definitions and worked examples. We particularly want to know whether the descriptors are useful and distinct, whether any overlap, whether Generated visual content is too broad, and what is missing.

These are factual provenance terms. Where a descriptor resembles language used in a contract or collective agreement, it takes its meaning from the HPF standard alone: it does not state whether consent was obtained or whether legal, contractual or workforce requirements were met.

For implementers, 0.9.2 also renames the version field to hpf_standard_version and updates the working label for the proposed C2PA assertion. The mapping remains an HPF proposal: it is not part of C2PA and has not been reviewed or endorsed by it, so the label should not yet be treated as stable.

On corrections and unresolved information

Revision 0.9.2 also corrects or clarifies several statements in the earlier consultation draft:

• An HPF declaration is a producer's statement of fact. It is not itself a contractual warranty, although parties may incorporate it into their own representations and warranties.
• Where the information needed to classify a production is missing or uncertain, the classification is unresolved, not No AI Used: no declaration means that no declaration has been made, and does not establish an absence of AI.
• An authorised party may declare for an already released production where it can establish the facts through reasonable enquiry. A classification must not be inferred from a release date or applied in bulk to a library.
• HPF classifies the finished production, not the companies or contributors involved in making it. Contributors disclose their relevant use and the output that reached the finished work; the producer applies the standard and declares the result.
• Provider marking under Article 50(2) of the EU AI Act and deployer disclosure under Article 50(4) are separate regulatory obligations. HPF is a proposed voluntary industry standard and does not itself establish compliance with either.

We have also added a glossary distinguishing several things now described as provenance: regulatory marking, technical provenance, consent and compensation, and the commercial disclosure HPF provides.

One issue remains open: established machine-learning features such as denoising, upscaling and tracking are increasingly embedded in ordinary production software, and we are consulting on when, if at all, those uses should require disclosure. The presence of machine learning does not by itself make a production Generative AI.

On the disclosure questions organisations already ask

We reviewed current published rules to understand how widely screen organisations already ask about AI use. As of August 2026, at least 50 film funds, festivals, awards bodies and broadcasters across 17 countries required AI disclosure, asked a formal question about AI use or expressly reserved the right to request that information.

This is a documented lower bound, not a complete international census: it includes only organisations whose current public rules we opened and read. Commissioner, distributor and platform requirements are likely to be undercounted because many appear in acquisition questionnaires, delivery documents and private contracts rather than published rules.

The more useful finding is the variation: the same production may encounter a yes-or-no question in one place, a free-text request in another and a list of tools in a third, without a shared definition of what counts as AI use. Producers may have to describe the same work differently each time, while recipients receive answers that are difficult to compare across organisations.

This is the disclosure interoperability problem HPF is designed to address. HPF standardises the disclosure input without determining what an organisation must permit, fund, acquire, exhibit or display. We are preparing worked examples showing how an HPF declaration could sit alongside the questions organisations already ask.

On testing the standard against production practice

The organisational review examines formal requirements; a separate qualitative study is testing how those requirements and the HPF vocabulary relate to the way practitioners discuss AI use in production.

We are conducting structured research across selected public online communities concerned with established screen-production practice and AI-native filmmaking. The study uses a documented research protocol and a frozen analytical corpus, and is intended to identify difficult classification cases, differences in production language and questions that the consultation should test across contrasting production contexts.

Analysis is still underway, so we are not reporting findings in this update. Any eventual publication will set out the methodology and limitations, including the self-selecting nature of public forum discussion and the fact that users' accounts cannot be independently verified. The study cannot estimate how prevalent particular practices are across the screen industries.

On audiences and the evidence gap

We also reviewed Luminate's consumer research on specific uses of generative AI in film and television, including scripts, music, voiceover, dubbing, animation, visual effects, synthetic actors and digital replicas. The relevant point for HPF is not that one use is good and another bad; it is that audiences distinguish between uses, which provides an evidence base for testing which distinctions are useful enough to carry as descriptors. Audience attitudes do not determine classification; the taxonomy does.

Current research can measure stated attitudes, but it cannot yet show how consistent provenance disclosure affects real viewing or purchasing behaviour: works are not labelled consistently at the point of choice, so there is no established market evidence showing whether disclosure changes demand.

We are looking for academic partners to design that research independently of HPF and of any commercial interest in the outcome, including the possibility that disclosure makes no difference. If you work in audience research, cultural economics or media consumption, email contact@humanprovenance.film.

On governance and the path to v1.0

In July we committed to finding a neutral home for HPF, preferably through an existing standards body and, failing that, through an independent coalition or membership structure. We are assessing what HPF would need in order to be governed independently of The Mise En Scène Company after consultation. No permanent host, legal form or funding model has been selected.

We are also identifying a small pre-v1 advisory working group to test the taxonomy, descriptors and implementation model against real production and institutional use cases. The group would not be HPF's future governing body, and participation would not imply endorsement.

Conversations are continuing about how HPF could sit alongside existing industry standards, policies and workflows. We will identify an organisation's involvement only where there is an agreed public basis for doing so. A conversation with HPF does not imply support for or a commitment to adopt the standard.

On international policy and research

HPF submitted feedback to the European Commission's call for evidence on an AI strategy for the cultural and creative sectors. The submission argued that assessments of AI's impact should distinguish AI adoption, consumption of AI-assisted or AI-generated works, stated audience attitudes, and behaviour when a work's production provenance is disclosed. It called for European research into the last of these without presuming what audiences will prefer. The submission is separate from the consultation and does not imply Commission support for HPF.

On 31 July, HPF was included as a case study in CoSTAR Foresight Lab's Creative Technologies International Scan #7, prepared by Olsberg•SPI, in its section on AI governance and disclosure. The report's wider discussion identifies clear, interoperable disclosure standards as potentially reducing consumer confusion and friction in sales and licensing. Its inclusion of HPF does not imply endorsement by CoSTAR or Olsberg•SPI.

On 7 October, HPF will take part in the Society for Computers and Law's Annual AI Conference in London. The AI and Creativity panel will examine the boundary between human creativity and AI assistance, ownership of AI-generated content, transparency and audience expectations.

How to respond

We are asking practitioners and organisations to test whether the Assistive and Generative definitions can be applied consistently, whether the provisional descriptors are useful and distinct, what is missing, and which edge cases still need guidance.

A short consultation survey covers those questions. It closes on 30 September 2026, ahead of the wider consultation, which remains open until 31 October. Partial responses are welcome.

The survey also allows respondents to register interest in the pre-v1 working group. You can also email contact@humanprovenance.film or respond through the GitHub repository.

Research update, 9 September 2026

This update originally reported that at least 50 organisations across 17 countries required AI disclosure, asked a formal question about AI use or expressly reserved the right to request that information. That was the documented lower bound when it was published. A subsequent wider review examined publicly available material relating to 456 screen organisations across 32 national or territorial contexts, together with regional and multinational bodies, and documented 136 organisations with an institutional AI-information requirement.

The text above is unchanged as the record of what HPF published on 1 September 2026. The current findings and method are in The fragmented state of AI disclosure in film.