Synthetic content marking and provenance

What it does

It makes generated content identifiable as generated. Images, audio, video and published text produced by the organisation's AI systems carry a machine-readable signal of their origin, by watermark, signed metadata or both, and where the law or the organisation's policy requires it, a visible label for the audience. The organisation also decides how it treats the same signals in content it receives. The system owner is accountable for marking; engineering implements it; the owner of each publication applies the visible disclosure. It operates at generation and at publication.

How to implement it

• Establish which outputs must be marked and which need a visible disclosure, from the law and from policy.
• Choose tools that apply marking by default, and confirm that the provider's marking is present.
• Attach provenance metadata to an open standard where the format supports it.
• Preserve marking through editing, export and publishing, and check that it survives.
• Test robustness against common changes such as cropping, compression and re-recording.
• Keep a record of what was generated, so origin can be shown when marking has been lost.

Evidence it operates

Design: a marking and disclosure policy; tool configuration. Operation: sampled published content with marking intact; robustness test results; records of generated assets; examples of disclosures on published material.

Where it fits

At the point of generation and through the publishing workflow. It is the primary control for unmarked synthetic content and supports generative AI used for disinformation. It is checked by review of AI-generated communications before publication, and sits beside AI transparency and disclosure labelling, which tells people when they are dealing with a machine.

Limits

Marking can be removed, deliberately or by ordinary processing, and metadata is stripped by many platforms. Watermarking of text is weak. The absence of a mark proves nothing, since content from tools that do not mark, or from which marks were removed, looks the same as authentic material. Standards and detection tools are still developing, so the control reduces doubt about the organisation's own output more than it helps identify other people's.