AI-generated-image disclosure is machine-readable and self-consistent
What it checks
Audits AI-generated-content disclosure at the only layer that is machine-readable and interoperable: the IPTC Iptc4xmpExt:DigitalSourceType XMP property. Catches the two failure modes that make disclosure worthless — malformed values outside the controlled vocabulary, and XMP that contradicts the C2PA manifest on the same asset.
| State | Result |
|---|---|
| no sampled image declares a digital source type | na |
| every declaration is a full vocabulary URI, agreeing with C2PA | pass |
a bare concept, an https spelling or a trailing slash |
warn |
| free text, or a concept outside the vocabulary | fail |
| the XMP declaration contradicts the asset’s C2PA manifest | fail |
Why it matters
IPTC types DigitalSourceType as [URI <External>], meaning the value must be a full NewsCodes URI from the ratified controlled vocabulary (base cv.iptc.org/…/digitalsourcetype, note the http scheme), not a bare token. A consumer matching against the vocabulary therefore silently ignores free text such as ‘AI-generated’ — the publisher believes it disclosed and every machine reader sees nothing. A bare ‘trainedAlgorithmicMedia’ or an https-scheme variant is a near miss: it is not the form IPTC defines, but Google’s documentation shows exactly those spellings without saying which it reads.
FALSIFIABLE: parse the XMP packet and test membership in the fetched vocabulary; separately, compare against the digital source type asserted in the asset’s C2PA manifest, where a disagreement is a hard contradiction one of the two pipelines produced.
Evidence
- IPTC Photo Metadata Standard 2025.1 — Digital Source Type — IPTC (spec, verified 2026-10-17)
- Property 11.13. Data type “CV-code”, with the CV note “Required: Digital Source Type NewsCodes” at
http://cv.iptc.org/newscodes/digitalsourcetype/. The XMP form is “Iptc4xmpExt:DigitalSourceType [URI <External>]”. The value is a vocabulary URI, not free text. - IPTC publishes metadata guidance for AI-generated “synthetic media” — IPTC (spec, verified 2026-10-17)
- Software that generates images with trained AI models should write the Digital Source Type
trainedAlgorithmicMediainto the file’s XMP packet, or into a C2PA manifest. “The complete URI for the recommended Trained Algorithmic Media term is”http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia. - Google Images SEO: Image Metadata — Google Search Central (vendor-doc, verified 2026-10-17)
- A named consumer. Google reads the IPTC Digital Source Type field and “supports the following IPTC NewsCodes”:
trainedAlgorithmicMedia,compositeSynthetic,algorithmicMediaandcompositeWithTrainedAlgorithmicMedia. - Counter-evidence for the scheme rule. The page shows the codes as bare tokens linked to
https://cv.iptc.org/…URIs. It does not say which spellings Google’s parser accepts. A bare token or anhttpsURI may still be read by Google, though neither is the form the IPTC standard defines.
How it scores
Tier per evidence policy: scored — grade B meets the A/B bar required for scored audits.
Grade: B — IPTC defines the property and its vocabulary, and Google documents reading it. The grade stays at B because no consumer documents rejecting the malformed values this audit fails, and no AI agent is documented to read the field.
Example finding
A publisher configures its CMS to stamp every AI-illustrated article header with DigitalSourceType = ‘AI-generated’. Because the value is free text rather than the cv.iptc.org/…/trainedAlgorithmicMedia URI, every conforming reader treats the field as unrecognized. The organization reports full AI-disclosure compliance internally while shipping zero machine-readable disclosure.
Sources
- IPTC Photo Metadata Standard 2025.1 — Digital Source Type — IPTC, spec (verified 2026-10-17)
- IPTC publishes metadata guidance for AI-generated “synthetic media” — IPTC, spec (verified 2026-08-20)
- Google Images SEO: Image Metadata — Google Search Central, vendor-doc (verified 2026-10-17)