Skip to main content

Content Generation and Synthetic Media

This page covers features that generate or manipulate text, images, audio, or video.

Why this is usually transparency-triggered (and sometimes prohibited)​

Synthetic and manipulated content can mislead users, so labeling duties commonly apply. If the feature intent crosses into harmful manipulation or deception patterns prohibited by the regulation, the path changes from transparency to prohibition.

Split the duty by role (Art 50)
  • Provider — mark synthetic audio/image/video/text in a machine-readable format (watermark/metadata) as artificially generated (Art 50(2)). Exempt for assistive or standard-editing functions that do not substantially alter the input.
  • Deployer — disclose deep fakes and AI-generated text published to inform the public on matters of public interest (Art 50(4)). Artistic-work and human-editorial-review exceptions apply.

Required obligations​

WHEN​

  • AI creates or edits user-visible content.
  • Content may be consumed as factual, human-origin, or unmodified without clear indication.

MUST​

  • Apply required labels/disclosures for generated or manipulated content.
  • As provider, embed machine-readable AI-generated marking (Art 50(2)); as deployer, disclose deep fakes and public-interest AI text (Art 50(4)).
  • Keep labels attached in all user-visible delivery paths (UI, exported media, downstream surfaces where applicable).
  • Escalate immediately if the use case could match prohibited manipulative/deceptive behavior (Art 5(1)(a)).

OUTPUT artifacts​

  • Implementation evidence for labeling (UI/API examples).
  • Traceability from asset to model/provider/version.
  • Escalation record if prohibited-risk indicators are found.
Meeting the machine-readable marking duty (Art 50(2))

Two widely adopted, technology-neutral approaches:

  • C2PA / Content Credentials — cryptographically signed provenance metadata attached to media. (c2pa.org)
  • Google SynthID — imperceptible watermarks embedded in AI-generated image, audio, text, and video. (SynthID)

The Act is technology-neutral: it does not mandate a specific standard, only that AI-generated or manipulated output be detectable in a machine-readable format.