The fragmented state of AI disclosure in film

HPF examined publicly available material relating to 456 screen organisations across 32 national or territorial contexts, together with regional and multinational bodies. The review documented 136 organisations with a published requirement to disclose, describe, notify, seek approval for or otherwise provide information about AI use. The figure is a documented lower bound drawn from public sources, not a global census.

Across the requirements documented, there is no shared definition of AI, common scope, common point in the workflow or common consequence. Collecting AI information is already part of screen-industry practice; describing it in a common way is not.

Some organisations ask only whether AI was used. Others seek details about the material affected, what the AI did, the tools involved or the stage of production. Requirements may also address rights, permissions, human authorship or prior approval. The examples below illustrate how those differences reflect the organisation’s role and the decision the disclosure is intended to support.

Each of these concerns AI use by filmmakers, but each serves a different institutional purpose: a fund assesses a proposal, a festival or awards body considers eligibility, a commissioner decides whether a planned use is acceptable, and a platform may communicate information to viewers. Those purposes shape both the questions asked and the level of detail requested.

None of the organisations named here has adopted or endorsed HPF.

Variation is not limited to the information requested. Published requirements can also leave important terms, thresholds or consequences unresolved.

Organisations variously refer to “AI”, “generative AI”, “AI-generated content”, “synthetic media” and particular forms of alteration. Some require disclosure of any use; others use thresholds such as “substantial”, “significant”, “realistic” or “misleading”. A question about “AI use” may not distinguish an AI function that enabled production work, temporary material generated along the way and content that remains in the finished work.

Scope can also be uncertain. A question may apply to the finished film, the wider production process, material created during development, or the preparation of an application. Requirements do not always distinguish between an intended use and what was ultimately done, or identify which version of a production the answer covers.

The consequences may be equally unclear. Disclosure can support information gathering, further review, prior approval, eligibility decisions or audience communication. Where a policy does not say what follows from disclosure, filmmakers may be unable to determine whether a use is permitted, subject to review or capable of making the entire work ineligible.

These are two distinct problems: fragmentation between organisations, which creates repeated work, and ambiguity within an individual requirement, which can prevent filmmakers from knowing how a policy applies at all.

Individual requirements can contain several questions, so the figures overlap.

The pattern is weighted towards describing the affected material and the purpose of the use. Far fewer requirements ask how AI was used or what human role or oversight remained.

Disclosure can be requested before a project exists in finished form, when a film is submitted, or after production is complete. Its timing generally reflects the organisation’s role.

Even where the underlying facts overlap, the required response may differ because disclosure occurs at different points and supports different decisions. Some organisations also ask about AI used to prepare an application, develop a pitch, administer production or create material that does not appear in the finished work, which sits outside the provenance of a finished version.

Most of the requirements reviewed are designed for institutional use. Of the 136 organisations with an institutional AI-information requirement, 103 had no confirmed audience-facing component and 33 also had an audience-facing requirement. Across the full review, 34 organisations had a public requirement, expectation or formal commitment to label, credit or explain AI use to viewers, so one audience-facing measure was recorded without an accompanying institutional requirement. Audience-facing mechanisms include labels, credits, captions, editorial notes and other viewer-facing notices.

Published practice is considerably more developed for assessment, eligibility, permission and delivery than for informing audiences. This does not imply that organisations without a public requirement oppose audience disclosure, and institutional collection does not by itself establish audience demand. Collecting information within a workflow and communicating provenance to viewers are separate tasks.

The review did not identify a shared way of describing the provenance of a defined finished-work version. Organisations ask about AI in their own terms, for their own decisions, and a disclosure prepared for one recipient may not satisfy another.

That points to a narrower opportunity for standardisation than a shared AI policy or a universal questionnaire: a common provenance record for the finished version that recipients can use within their own policies, while continuing to request whatever information is specific to their decisions. Standardising the record does not make an ambiguous policy clear. Each organisation still defines the scope of its rules and explains how disclosed uses affect its decisions.

HPF reviewed publicly accessible rules, guidance and other policy materials relating to 456 screen organisations across 32 national or territorial contexts, together with regional and multinational bodies. The review covered festivals and awards, funds and public agencies, broadcasters and commissioners, development programmes, and distribution and exhibition platforms.

For the institutional count, an organisation was included where public evidence showed a requirement to disclose, describe, notify or seek approval for AI use. Audience-facing measures to label, credit or explain AI use were coded separately. Each organisation was counted once within each measure, and requirements limited to particular schemes, uses or thresholds were included as qualified requirements.

The findings are a documented public lower bound, not a global census. They describe published requirements, not how consistently they are applied. Private, unpublished, unindexed, portal-only and contractual requirements may therefore be underrepresented.

• Add HPF to your workflow
• When an organisation asks about AI
• Read the HPF taxonomy
• How HPF works