Provifai
All posts
Technology8 min read

C2PA, Metadata, Watermarks: How Machine-Readable AI Labelling Really Works

The EU Code of Conduct refers to machine-readable labelling. What lies behind C2PA, IPTC DigitalSourceType, and watermarks — and why visible labels alone are not enough.

When people talk about AI labelling, most think of a visible label on the image. That is only half the story. The other half takes place where no visitor looks: inside the file itself. Machine-readable labelling is the aspect targeted by auditors, platforms, and the EU Code of Conduct — and it is refreshingly concrete.

C2PA: Proof of Origin Within the File

The Coalition for Content Provenance and Authenticity (C2PA) — backed by Adobe, Microsoft, Google, Sony, and others — has established a standard that cryptographically signs and records a file’s provenance: Which tool created the content? Was it edited? Who issued the signature? Tools such as Adobe Firefly already write this evidence automatically. The crucial point: the signature can be verified, but not forged without detection. A C2PA record is therefore the strongest form of labelling — proof, not mere assertion.

IPTC DigitalSourceType: The Industry’s Vocabulary

Much simpler, but widely used: the IPTC field DigitalSourceType describes in the metadata how an image was created. The value trainedAlgorithmicMedia means: generated by a trained AI. digitalCapture means: genuine camera shot. This is precisely the vocabulary referenced by the EU Code of Conduct — it is the lowest common denominator for machine-readable labelling.

Watermarks: Invisible Yet Detectable

A third approach: invisible watermarks such as Google’s SynthID, which are embedded directly into the image pixels and are designed to withstand compression or cropping. Their strength is robustness; their weakness is exclusivity — so far, they can mainly be read by the provider themselves.

Why Visible Labels Alone Are Not Enough

  • Anyone can remove a visible label — a signed proof of origin cannot be tampered with unnoticed.
  • Platforms and search engines read metadata, not image corners.
  • Evidential value comes from verifiability: in case of doubt, what is in the file counts.

The Flip Side for Operators

The same standards that enable labelling also make missing labelling visible: an image with a C2PA signature from an AI tool, published without disclosure, documents the gap at the same time. An inventory of your own content reads precisely these traces — and reveals where action is needed.

Free audit

What is actually on your website?

Enter your domain, wait less than a minute — you will see which content shows traces of AI generation. No registration required.