Why Content Credentials disappear after upload, and how to keep them

Published 12 October 2026

You sign an AI-generated image with a C2PA manifest, publish it, download it again, and the Content Credentials are gone. This is the most common surprise in marking AI images. The manifest is metadata, and a great deal of software throws metadata away.

Where the manifest gets lost

• Social platforms and messaging apps. Digimarc lists uploading to a social network, sending through a messaging app and editing in a phone app as everyday ways manifests are lost (Digimarc, "How Digital Watermarks Strengthen C2PA Content Credentials").
• Your own delivery pipeline. Image resizers, format converters, content delivery networks and content management systems often produce a new file instead of passing the original through. A new file usually has no manifest unless the tool is built to carry it.
• Re-encoding. Saving the image again as a new JPEG or WebP, for example after compression, writes a new file. The old metadata is only kept if the encoder copies it.
• Screenshots. A screenshot is a new image with no metadata at all.
• Deliberate removal. Anyone can strip metadata with ordinary tools.

Even when a tool keeps the metadata, changing the pixels breaks the manifest's hash of the image, and a validator then reports the manifest as no longer matching the file.

Test your own pipeline first

You control one part of this: what happens between your model and your users. Check it before you rely on it.

1. Mark one image at the point of generation.
2. Publish it through your normal path: storage, resizing, the content delivery network and the page that shows it.
3. Download the image as a user receives it, from the page itself, not from your storage bucket.
4. Drop that file on a C2PA reader, such as the free checker. It should report the manifest, and report it as valid.
5. Repeat for each image size or format your pipeline produces, such as thumbnails and WebP versions.

If the manifest is missing, find the step that rewrote the file, and either configure it to keep metadata or move marking after it.

Recover what platforms strip: durable Content Credentials

You cannot stop other platforms stripping metadata. You can make the mark survive anyway. An invisible watermark in the pixels carries an identifier. A stored copy of the manifest is kept under that identifier. When a file arrives with its manifest stripped, a checker decodes the watermark and looks the manifest up. C2PA calls this link a soft binding, and Adobe's open-source TrustMark watermark was built for it (Adobe TrustMark: using TrustMark with C2PA).

This only works while the watermark survives. Heavy processing, cropping or a low-quality screenshot can destroy it. Tellmark measured its TrustMark watermark surviving JPEG quality 75, a 50% resize and a WebP round trip on 20 of 20 test images, and nothing else (Tellmark, what marking cannot do).

How Tellmark handles this

Every image Tellmark marks gets both a signed C2PA manifest and a TrustMark watermark carrying its mark ID, and Tellmark stores the manifest. The free checker and each customer's verify page read the manifest if it is there. If it is not, they decode the watermark and find the original record, so a stripped copy can still be traced to its marking.

Sources

• Digimarc, "How Digital Watermarks Strengthen C2PA Content Credentials"
• Adobe TrustMark: using TrustMark with C2PA
• C2PA, the Coalition for Content Provenance and Authenticity
• Tellmark, what marking cannot do

Related guides

• C2PA and invisible watermarks: how they differ, and why AI images use both
• How to add C2PA Content Credentials to AI-generated images in Python
• The 2 December 2026 deadline for marking AI-generated content

Describes technical measures. Not a legal assessment.