Guide
How to check Content Credentials (C2PA)
Content Credentials are a signed history attached to a file. Reading the words is not the same as checking the signature. This site does the first and does not do the second.
17 October 2026 · Check AI Free
The Coalition for Content Provenance and Authenticity publishes the C2PA technical standard. The user-facing name for that history is Content Credentials. A manifest can name a tool, a signer and a sequence of actions, including a generative step or an edit. The C2PA site describes the standard; ContentCredentials.org is the public explainer the coalition links to. Neither page checks your file for you.
Check a file in this order
Step 1
Get the file the signer would have produced
A screenshot, a recompressed download and a “save image” copy are new files. Our metadata experiment did not verify any C2PA signature, and it found that an XMP container in all eight photographic originals was absent after a specified JPEG export and after a browser screenshot. An XMP container is not itself a Content Credential. The practical point is narrower: a transform that rewrites the file can leave you with nothing to validate. See what is and is not known about Instagram, WhatsApp and Facebook before you blame a platform you have not tested.
Step 2
Open a validator, not a string search
Adobe’s Content Authenticity Verify tool at verify.contentauthenticity.org asks you to select a file, or to drag it in, and it lists the formats it inspects, including JPEG, PNG, WebP and several video and audio types. The Content Authenticity Initiative’s Inspect documentation describes what a conforming validator shows: no manifest, a manifest whose certificate is not on the trust list (“Unrecognized”), or a signer and time taken from the claim when the certificate is on the list. Use that class of tool for the signature. The open specification is at spec.c2pa.org.
Step 3
Read the validation state before the story
A manifest can be present and still fail. The CAI’s validation notes say errors are reported in a validation status list, including an untrusted, invalid, revoked or expired signing credential, and a claim that no longer matches bytes that were edited after signing. Success is the absence of those errors, which you have to look for. A thumbnail and a friendly producer name do not replace that list.
Step 4
Read the actions as claims, including AI actions
A valid signature binds a signer’s claim to the asset the validator checked. It can say an image was generated, opened in an editor, or captured. A valid “generated” action is evidence of generation and evidence that the manifest is intact. It is not evidence that the depicted event happened. A valid capture claim is not a background check on the person in the frame.
Step 5
Keep an unrecognized signer visible
Inspect’s documentation says a certificate that is not on the C2PA trust list, including test certificates shipped with the SDK, is shown as unrecognized. Record that state. An unrecognized signature is a different result from a trusted one and a different result from a broken hash. Do not upgrade it in your notes.
Step 6
If there is no manifest, say that and move on
Many cameras and many generators still produce files with no Content Credentials. Absence is not a finding of AI, and it is not a finding of a camera. Continue with embedded PNG text, reverse search and, if you want it, a model score. The checklist for AI-generated images puts those steps in order. The checker on this site can still report ordinary source clues while leaving the signature unchecked.
What a string search will miss
Manifests are structured data, often inside a JPEG, PNG or similar container, and they are signed. A search for the letters “C2PA” can miss a valid manifest and can hit an unrelated comment. The PNG reader on the Stable Diffusion PNG info page shows tEXt, iTXt and zTXt fields. Those fields are how Automatic1111-style parameters and ComfyUI workflows are commonly stored. They are not a C2PA validator, and a workflow JSON is not a signed credential.
If you also want a pixel-model estimate, run it on the same original and label it separately. Detectors disagree for reasons that have nothing to do with whether a signature checked out. A high model score does not repair a failed manifest, and a valid manifest does not need the model score to be “more true.”
Sources
- C2PA — The coalition and the public description of the Content Credentials standard.
- ContentCredentials.org — Explainer linked from the C2PA site.
- C2PA technical specifications — The specification a validator implements.
- Content Authenticity Verify — Adobe’s inspector for dragging in a file and reading its Content Credentials.
- Inspect tool documentation — How a conforming validator distinguishes a missing manifest, an unrecognized certificate and a trusted signer.
- Validating manifests — Error codes for untrusted, invalid, revoked or expired credentials, and for assets edited after signing.
- Method and limitations — This site observes source declarations and does not verify signatures.
After the manifest
- Check an image — Source clues and an optional model score. Signatures stay unverified.
- Metadata after sharing — Why the download may no longer be the signed file.
- DALL·E context — A suspected generator still needs the original file, not a style guess.
- Reverse image search vs AI detector — A manifest does not tell you whether the picture was posted last year under another caption.