Skip to main content

How generated images declare they were made by AI

Every image Rational AI generates carries a machine-readable marker saying a model made it. The marker is embedded in the file itself, so it keeps saying so once the file has left the platform — in a download, an export, or a response from the API.

What is embedded​

Generated images carry an XMP packet containing the IPTC DigitalSourceType property, set to trainedAlgorithmicMedia.

That property is the interoperable way for a file to declare it was produced by a generative model. It's what search engines and platform ingest pipelines read when they label content as AI-generated, which is what makes it worth more than a note in a database:

  • A database record doesn't travel with the file. Download the image and every trace of where it came from is left behind.
  • A visible watermark doesn't survive a crop.
  • An embedded marker goes wherever the bytes go.

When it is applied​

The marker is written before the image is stored, not added on the way out. Every copy that subsequently leaves — a download, the gallery, an API response — carries it, because they're all serving an already-marked file.

Marking on the way out would leave the stored object unmarked and make correctness depend on every future download path remembering to add it.

Two practical consequences:

  • Re-marking is harmless. An image that already carries the marker isn't given a second one.
  • Unrecognised formats pass through untouched. PNG and JPEG are marked. Anything else is stored exactly as generated rather than risking a corrupted file — an unlabelled image is a smaller failure than a broken one, and you can't tell the difference after the fact.
⚠️THIS IS A DECLARATION, NOT A SIGNED PROVENANCE MANIFEST — IT IS NOT C2PA

The distinction matters if you're answering a question about provenance, because the two are often spoken of as though they were the same thing.

The marker establishes what the file is: something a model generated. It does not establish that nobody has altered the file since, and anyone who wants to strip it can — it's metadata, not a seal.

C2PA additionally binds the claim to a signing identity, so the assertion can be shown not to have been tampered with. That requires a certificate and a signing service, and Rational AI does not do it today. If a tender or a policy asks specifically for C2PA, this does not satisfy it.

Result​

You can state precisely what a generated image discloses about itself: an embedded, interoperable IPTC declaration that a model made it, written before storage so every copy that leaves carries it — and you know the honest limit of that claim, which is that it is an unsigned declaration rather than a tamper-evident C2PA manifest.