AI labelling
Last updated: 19 November 2026
Some of my images and voice recordings are made with AI systems that I run myself. I mark such content in a machine-readable way, as required by Article 50(2) of the EU AI Act (Regulation (EU) 2024/1689). This page explains how the marking works and how you can check it for free.
What is marked
Every published or delivered file that contains AI-generated parts from my own systems: images and AI speech. The marking sits in the final file, not in intermediate versions.
Two layers
1. Invisible watermark in the content
- Images: Adobe TrustMark, variant P, error-correction schema 0 (40 bit).
- Speech: Meta AudioSeal 0.2.1, 16-bit model. A recording counts as marked when at least 50 % of its frames score above 0.5, either across the whole file or in its strongest 4-second window.
The watermark can be neither seen nor heard. It is designed to survive common processing such as compression and resizing.
2. Signed C2PA manifest (Content Credentials)
- Source declared with the IPTC digital source type:
trainedAlgorithmicMedia(fully AI-generated),compositeSynthetic(composite with AI elements) orcompositeWithTrainedAlgorithmicMedia(AI edit on top of real material). - For images, the manifest points to the watermark (soft binding
com.adobe.trustmark.P). - Signed with the certificate “kz Content Credentials”, issued by Tom Langer kz Root CA, timestamped by DigiCert. This CA is not on the C2PA trust list, so general tools show “issuer not recognized”. Root certificate fingerprint for comparison:
SHA-256: 16ea94869c99d8bf284e216a310d21392e16d1a8297d4ab2dfcd43dbd4762ed0
Both markings carry a random identifier only and no personal data.
Social networks often remove the manifest on upload. The watermark is there so the content stays recognisable even then.
How to check a file
- kz-verify (free, open source, runs on your computer, requires Python): checks both layers and matches the signature against my root certificate. Instructions and code: github.com/muclanger/kz-verify
- Manifest only: verify.contentauthenticity.org, Adobe Inspect or the Content Authenticity Initiative’s
c2patool. - Watermark only: the reference decoders of the open projects TrustMark and AudioSeal with the parameters above.
Please do not remove the markings
The markings exist so that anyone can recognise AI-generated content. Please do not remove, alter or imitate them.
Questions: see the imprint (German) for contact details.