How we mark files
AI-generated assets delivered from the Caisting.ai catalogue carry two independent layers of marking. This page explains what exactly is in the file, what a third party can check on their own, and what we do not promise.
We will submit our watermark identifiers to the public C2PA algorithm registry; the entry will list this page's address as the place to learn more.
Layer one: the watermark
We embed an invisible watermark in images and video, carrying an identity identifier and — in assets from a licence package — a licence identifier. The watermark is keyed — we are the ones who read it, on the Caisting.ai verification page. Without our key nobody outside can read it or detect it. It survives ordinary processing of the file: rescaling, re-compression, cropping. Our library also handles audio, although today's catalogue assets are silent.
The algorithm identifiers describing this layer:
- ai.caisting.wm.image — images
- ai.caisting.wm.video — video
- ai.caisting.wm.audio — audio
Layer two: signed C2PA metadata
We attach signed metadata in the C2PA standard, also known as Content Credentials. It carries the AI-generated content label and a soft binding identifier, which is the same watermark payload written out in the metadata.
This layer anyone can check without us — with any public Content Credentials tool, for as long as the file keeps its metadata.
The exception: real photographs of the person
A digital twin's licence package may also include real photographs of the person who lent their likeness. Such files are not AI content, so they carry no AI-generated label, no watermark and no soft binding identifier. We sign them with provenance metadata only — so that it is clear they came from us.
If you are checking such a file and see no AI label, it does not mean the file has been tampered with. It means it is a photograph, not generated material.
Two caveats we state plainly
Metadata is fragile. Most image editors and social platforms strip it on save. After such a save only the watermark remains in an AI file — and that one we read. In a real photograph of the person, described above, nothing remains: there is no watermark there, because it is not generated content.
Our signing certificate is not yet on the C2PA trust list. Checking tools will confirm the signature is present, but may flag the issuer as untrusted. We have applied to the C2PA conformance programme, which leads to a trust-list certificate — a different path from the algorithm registry mentioned above. Until it is complete, we describe this exactly as it stands.
How to check a file
Open it in any tool that reads Content Credentials — for AI material you will see the AI-generated content label and the soft binding identifier. If you hold a licence code in the CAI-XXXXX-XXXX format, you can check its status at caisting.ai/verify.
Why two layers
Each has a different weakness. Metadata is readable by anyone but disappears on an ordinary save. The watermark survives processing but is unreadable without our key. Together they give a marking that outlives both a trip through an editor and the absence of any contact with us.
Contact
For questions about file marking: hi@caisting.ai