See and remove C2PA Content Credentials

What are C2PA Content Credentials?

Content Credentials is the public name of a C2PA manifest: a signed record, stored inside the image file, of which tool made the image and what was done to it. For an AI image, it usually states that generative AI was used.

The C2PA (Coalition for Content Provenance and Authenticity) publishes the standard. A file holds a manifest store with one or more manifests. Each manifest contains a claim: the software that wrote it (the claim generator), a list of assertions such as the actions performed (created, edited, generated by AI) and the images it was made from, and a cryptographic signature from the certificate of whoever signed it. A hash ties the manifest to the exact bytes of the file, so any later edit shows.

Where the manifest lives depends on the format (C2PA 2.2, annex A):

What does PromptWipe show about them?

The report lists how many manifests the file holds and their size, the software the file says made it (“Made with”), the signer’s name as written in its certificate, and the title, if there is one. These values are declared, not checked: PromptWipe reads them but does not verify the signature.

To check a signature, use a verifier such as Verify, from the Content Authenticity Initiative. It is a separate website: read how it handles your file before you give it a private image.

Two more cases appear in the report. Some files carry only a link, in their XMP metadata, to Content Credentials stored elsewhere: PromptWipe shows it as “Link to Content Credentials” and removes it with the rest of the XMP. And when a manifest cannot be read, the report says so and the block is removed all the same.

When should you keep Content Credentials?

Keep them whenever a rule or a promise depends on them: images made with Adobe’s generative AI, platforms or clients that require AI content to be labelled, and work where provenance is the point. In those cases, use PromptWipe to read them, not to remove them.

• Adobe generative AI. Adobe’s Generative AI User Guidelines (last updated 15 May 2026, section 6) say: “You must not remove, alter, or disable any Content Credentials.” If your image comes from Firefly or another Adobe generative AI feature, keep them.

• Labelling rules. If a platform, a stock agency, a publisher or a client asks you to disclose AI-generated content, removing the credentials does not change that duty. Label the image yourself.

• Deepfakes. Article 50(4) of the EU AI Act requires deployers of an AI system that generates deepfakes to disclose that the content was artificially generated or manipulated, with a lighter duty for works that are evidently artistic, creative, satirical or fictional. If that duty applies to you, it stays, whatever the file contains.

• Your own provenance. A photographer or illustrator who signs their work can use the credentials to show where it comes from. Removing them throws that proof away.

Why would you remove them?

For privacy and control over what you share. Depending on the tool that wrote it, a manifest can name the organisation that signed it, the software and its version, the list of edits, and the images it was made from, sometimes with their thumbnails.

That is more than many people expect when they send a draft to a client or post a picture. A manifest also takes space: in a JPEG it can fill several segments of 64 KB. If none of the cases above applies to your image, you may prefer to share it without that record.

Note that PromptWipe removes all metadata, not only C2PA: EXIF, XMP, prompts and other text blocks go too. It keeps only what the picture needs to look the same: the colour profile and the rotation. It has no option to remove C2PA alone: if you need to keep the credentials, keep the original file.

What happens after they are removed?

The file then has no Content Credentials, and a verifier reports that none were found. Anyone who needs the provenance record will no longer find it in this copy, so keep the original if you may need it. The pixels are copied byte for byte and the picture does not change.

Three things stay:

• Invisible watermarks written into the pixels, such as SynthID. They are not metadata, and no metadata tool removes them.
• A link through a watermark. The C2PA standard defines a soft binding: a fingerprint or an invisible watermark that lets a verifier find a manifest stored elsewhere, for example in an online repository. If the image has one, it can still be matched with its credentials.
• Labels that platforms apply from other signals: watermarks, detection models or what the uploader declares.

What do AI services put in their Content Credentials?

Mostly a statement that the image is AI-generated, which service made it and when, rather than your prompt. Two vendors document their fields:

• Microsoft says images from Azure OpenAI carry a manifest signed by Microsoft, with the description “AI Generated Image”, a software agent such as “Azure OpenAI DALL-E” or “Azure OpenAI ImageGen”, and the time of creation.
• Black Forest Labs says images from its FLUX API carry an AI-generated flag, the claim generator, the UTC time and the FLUX model name and version, and no prompt, negative prompt, API key, account ID or IP address.

We have not yet inspected files from ChatGPT, Gemini, Midjourney or Firefly, so this page makes no claim about them. A page for a generator is published only after a real file from it has been checked. The guide to AI image metadata lists what the open-source tools write.

Which files does it work on?

PNG, JPEG and WebP, up to 50 MB, one image at a time. AVIF, HEIF and GIF files are refused for now, with a link to request the format. Everything runs in your browser; how it works explains how to check that nothing is sent.

Sources: C2PA 2.2 specification (annex A.3, A.5, soft binding); Adobe Generative AI User Guidelines; Microsoft Learn, Content Credentials; Black Forest Labs, C2PA in FLUX images; EU AI Act, Article 50.