How to Write a Useful Synthetic Media Disclosure

How to Design a Useful Synthetic Media Disclosure

A useful synthetic media disclosure tells the audience what they are seeing or hearing, what AI materially generated or changed, which human and source contributions remain, and what evidence supports any real-world claim. It should appear where the audience encounters the work, not only in a distant policy page.

"Made with AI" is often too vague. A complete software log is usually too much. The goal is informed interpretation.

This guide provides a conservative editorial method. It is not legal advice or a universal platform-compliance template. The rules that apply depend on jurisdiction, platform, content, audience, people depicted, and distribution date.

Separate four different questions

A disclosure becomes clearer when it stops asking one label to do four jobs.

flowchart TB
    A["Synthetic media disclosure"] --> B["Production: what AI generated or changed"]
    A --> C["People and rights: whose likeness, voice, or work appears"]
    A --> D["Provenance: what process record travels with the asset"]
    A --> E["Truth: what real-world claims were independently verified"]

Production disclosure explains how the asset was made.

People and rights disclosure explains material human performances, licensed sources, consent, and authorization where appropriate.

Provenance records signed assertions and history.

Factual verification explains why a publisher believes a claim about the real world.

One layer cannot silently substitute for another.

Start with the audience-facing fact

The first sentence should resolve the most likely misunderstanding.

If a realistic scene did not occur, say that. If a real person's voice was synthesized with permission, say that. If only the background was generated, do not imply that the entire performance was synthetic.

Use ordinary words before tool names. "This is a fictional scene generated from text prompts" tells a viewer more than "Created using Model X."

The tool and version can follow when they help reproduce or interpret the work.

Describe the material contribution

Not every AI-assisted action matters equally to an audience.

Spelling correction, caption timing, noise reduction, an invented person, a cloned voice, a rewritten statement, and a generated news scene create different expectations.

A concise record can use these fields:

The public disclosure does not need every internal field. The production record should preserve them.

Use a short statement and a deeper record

The immediate label should be clear enough for the viewing context. A longer note can explain the production.

For a fictional generated scene:

This fictional scene was generated with AI from original prompts and did not record a real event. Dalton Anderson selected the takes, edited the sequence, and reviewed the final audio. No real person's likeness or cloned voice was intentionally used.

For a mixed live-action and generated piece:

The on-camera performance was recorded with the named actors. AI tools generated the city background and some environmental audio. The actors' voices and dialogue were not synthesized. The final composite was edited and reviewed by the production team.

For a documentary reconstruction:

This is an AI-assisted reconstruction, not footage of the event. The sequence is based on the cited records linked below. Generated imagery and voice performance illustrate the timeline; they are not independent evidence that the event occurred.

For a product demonstration:

This demonstration includes selected AI-generated outputs. The prompt, model version, settings, attempt count, selected and failed examples, edits, and capture date are documented below. Results may differ by configuration and date.

These examples are editorial starting points. They do not determine whether the content is lawful, licensed, nondefamatory, accessible, or compliant with a platform rule.

Put the disclosure where misunderstanding happens

A disclosure hidden after the viewer has acted arrives too late.

Use a visible on-screen or adjacent statement when realism could materially mislead. Repeat it in the description and metadata where the platform supports those layers. For audio, use an audible disclosure when listeners may not see a caption. For an installation or live experience, disclose at the first exposure.

The label should survive common sharing paths. A description can disappear when a clip is downloaded and reposted. An on-screen label can be cropped. Embedded provenance can be removed. Redundancy is a feature.

Accessibility matters. Use readable contrast, sufficient duration, plain language, captions, transcripts, and alternative text appropriate to the medium.

Use platform controls as well as prose

Platform disclosure settings can trigger labels, preserve structured information, and satisfy platform-specific processes. They should not replace the creator's own accurate description when the audience needs more context.

YouTube's current AI disclosure help requires creators to disclose photorealistic content materially generated or altered with AI, including realistic scenes that did not occur and depictions of real people doing things they did not do. Its examples and interface can change.

YouTube also explains how it may surface C2PA-based production information. A platform label and a Content Credential are different. One is a user-facing platform presentation. The other is a signed provenance record interpreted through compatible systems.

Check the actual upload surface on the release date.

Carry provenance when available

The current C2PA 2.4 specification can record signed assertions about origin, actions, ingredients, and asset history. That can make a creator's disclosure harder to separate from the file and easier to inspect.

It still does not prove that the public description is complete or that a depicted event happened. C2PA's guiding principles explicitly avoid judging whether provenance is good or bad.

Use machine-readable provenance and human-readable disclosure together. Preserve the project files, source records, approvals, model and tool versions, and final exported asset as internal evidence.

Treat real people and real claims as a higher-risk class

A label does not create permission to use a person's face or voice. It does not cure impersonation, defamation, privacy invasion, deceptive advertising, missing license rights, or unsafe distribution.

If a work depicts a real person saying or doing something they did not say or do, the production needs a documented rights and risk decision before publication. Satire, commentary, art, news, advertising, and commercial entertainment may receive different treatment under different laws and policies.

If a work makes a factual claim, preserve the underlying sources and verification. "AI-generated" does not excuse a false statement. "Human-created" does not make one true.

Account for the August 2026 EU change

Article 50 of the EU AI Act contains transparency duties for certain providers and deployers of AI systems. The European Commission's July 20, 2026 guidance says the relevant obligations apply from August 2, 2026.

The regulation distinguishes roles and includes scope conditions, exceptions, machine-readable marking duties, and disclosure duties for certain deepfakes and public-interest text. It cannot be converted safely into one global label.

This page was reviewed five days before that application date. Any release on or after August 2 requires a current EU scope review, as well as review of every other applicable jurisdiction and platform.

Run a disclosure review before export

Read the statement while viewing the final asset, not the storyboard.

Ask whether a reasonable audience could still misunderstand who performed, whether the scene occurred, what was generated, whose work was used, or what evidence supports the claim. Compare the disclosure with the platform setting, Content Credential, credits, rights record, and actual final edit.

If those layers conflict, stop. A machine-readable claim saying "digital capture" should not accompany a public caption implying the whole scene is synthetic. A written disclosure should not name a model that was not used. A credit should not imply consent that the contract does not provide.

Keep the disclosure with the asset through revisions. Material changes require a new review.

The standard is useful context

A good disclosure does not apologize for using a tool. It gives the audience enough information to understand the work.

Say what did not happen. Say what the system generated or changed. Preserve the human and source contributions. Identify real people and rights decisions. Separate production history from factual verification. Carry structured provenance where possible. Refresh the statement for the actual platform and law at release.

This guide was freshly written from E070, C2PA 2.4, current YouTube help, the EU AI Act, and July 2026 European Commission guidance reviewed on July 28, 2026. It remains in editorial and qualified legal review because the EU application date and release context are material. AI assistance was used for research organization, drafting, and validation. Publication remains unauthorized.