Skip to content
Disclosemark

Docs

C2PA, metadata, and a visible label

Three different things, doing three different jobs. Most of the confusion comes from treating them as one decision.

Embedded metadata

Your image file has a header, and the header has room for text fields. The relevant one is IPTC DigitalSourceType, carried inside an XMP packet, whose value is an IRI from a controlled vocabulary — for a generated image, http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia.

It is plain text. Anybody with exiftool can read it, change it or delete it. That is the weakness and also the reason it is the practical answer: every image tool on earth already understands XMP, so the marking is readable everywhere without anyone adopting anything new.

This is what the EU AI Act's “machine-readable format” language is being read to mean in practice, and it is what the labeler writes on every plan including the free one.

C2PA Content Credentials

C2PA wraps the same kind of assertion in a cryptographically signed manifest. The manifest records what was done to the image, hashes the pixel data, and is signed by a certificate. Change one pixel and the claim no longer validates.

That gives you something metadata cannot: proof that a specific party made a specific declaration at a specific time, which nobody can forge after the fact. In a dispute — a takedown appeal, a client arguing about what you delivered — that is a materially stronger record than a text field anyone could have typed.

The part that gets oversold: a manifest signed with a certificate that is not on the C2PA trust list validates as present but untrusted. It still proves the file has not changed since signing. It does not prove who you are. Getting a certificate from a listed CA is a separate, paid, identity-checked process, and no tool can do it for you.

C2PA also breaks easily in ordinary use. Any resize, re-encode or edit by software that does not understand C2PA invalidates the claim, and most of the software your images pass through does not understand C2PA yet.

A visible label

Text burned into the pixels. It is not machine-readable, so it satisfies nothing in Article 50(2), and it does not appear in your listing description, so it does not satisfy 50(4) either.

It is still worth using, for a reason that has nothing to do with compliance: on a marketplace, buyers open the image before they read the description. A corner label is the only one of the three that a human sees at the moment they form an impression of the product. It is the mechanism that prevents the complaint, rather than the one that wins the appeal.

What survives what

 MetadataC2PAVisible label
Uploading to Etsy or ShopifyUsually survivesUsually survivesAlways survives
A platform resizing your imageOften strippedStripped, and the claim breaksSurvives
Screenshotting itGoneGoneSurvives
Cropping in Canva or PhotoshopUsually keptKept if the editor supports C2PA, else goneSurvives unless cropped out
Saving from a browser as WebPGoneGoneSurvives
Someone deliberately removing itTrivialTrivial to remove, impossible to forgeNeeds an editor and leaves a mark

So which do you need?

Metadata plus a written disclosure line. That combination covers both halves of Article 50 and costs nothing. Do that first, and do it on every image, before you think about anything else.

Add a visible label on marketplaces where images carry the listing — prints, wall art, digital downloads, anything where the image is the product rather than a photograph of the product.

Add C2PA if you are handling work for clients who may later dispute what you delivered, if you publish to LinkedIn or Meta surfaces where the credential is read and displayed, or if you want a signed record for your own protection. It is a paid feature here because signing costs us certificate infrastructure, not because the disclosure it makes is any better.

None of the three stops someone else stripping the disclosure from a copy of your image and reposting it. That is not what any of them are for. They establish what you published, which is the thing you are accountable for.

Checking your own files

Do not take our word for it. Install exiftool and run it over a file the labeler gave you:

exiftool -G1 -a your-image.jpg | grep -i "digital source"

You should get back [XMP-iptcExt] and the IRI. Then upload the same file to your platform, download it back from the live listing, and run the same command. That second result is the one that tells you whether the marking survived, and it is the only test that matters. What we found when we did that.