What are Content Credentials?

Content Credentials are signed records stored inside an image file. They follow the C2PA standard, and each record says which software made a statement about the file, what the statement is, and who signed it. They are the strongest provenance signal an image can carry, because a signature can be checked. They are also easy to lose, and a valid signature shows who made a claim, not that the claim is true.

Our image checker reads these records in your browser, tests the signature and the file match, and checks the signer against two dated trust lists. This page explains each step.

What a record contains

A C2PA record, also called a manifest, is a bundle of signed statements. It names the claim generator, the software that made the claim. It lists actions, such as "created" or "edited", and an action can carry a digital source type, for example trainedAlgorithmicMedia, which is the standard way to say that generative AI created the media. It can list ingredients, which are earlier files used to make this one. It contains a hash of the image data, which ties the record to that exact file content, and it ends with a signature made with the signer's certificate.

Which tools attach them

These are the cases we can source. "Reported" means a secondary source, not the company's own page.

• OpenAI. Images from ChatGPT and the API carry a record signed by OpenAI (reported). OpenAI's API documentation says a record counts as detected only when it is valid and trusted, names OpenAI and includes an AI-generation action.
• Azure OpenAI. Microsoft's documentation says Content Credentials are added automatically to output from its DALL-E and GPT-image models, issued by Microsoft.
• Google. Google's documentation lists Content Credentials, signed by Google LLC, for gemini-3-pro-image and gemini-2.5-flash-image. The Gemini app is reported to have added them in November 2025.
• Adobe Firefly. Reported to add them automatically to Firefly and API output that is entirely generated.
• Black Forest Labs. Its help page says images from its API, playground and hosted service carry a record with the claim generator, the model name and version, and a SHA-256 hash binding. Open-weight Flux models add none.
• Samsung. Reported to add a record to images edited with Galaxy AI.
• Midjourney. Reported to add none. We found no documented records for Ideogram, Leonardo or Canva.

What a check can tell you

1. Is there a record? Many images have none.
2. Is the signature valid? This shows the record was not altered after it was signed, and that it was signed with the key that belongs to the certificate inside it.
3. Does it match this file? The record's hash must match the image bytes. If it does not, the image has changed since signing and we say so.
4. Is the signer trusted? The signer's certificate must chain back to an anchor on a trust list. A record can pass the first three tests and still fail this one, which means the signature is sound and the signer is unknown to the list.

Only then do we read what the record declares. A change to the image data of an ingredient, an earlier version the file was made from, counts against the whole file, so a tampered ancestor never reads as trusted. Other problems in an earlier version's record, such as a certificate on neither list or an expired one, are reported about that earlier version and are not blamed on the file's own signer. Anyone can make their own certificate, name it after anyone, and sign a record with it, which is why a name on a certificate that no list vouches for is shown as what the certificate says, never as who signed. A sample image from the c2pa-rs project, signed with the C2PA test certificate, reads as valid but not trusted on both lists, which is the correct result.

The two trust lists

A trust list is a published set of certificates that signers chain to. We check two, and we always tell you which one matched.

• C2PA Trust List. The list C2PA publishes in its conformance repository. Our copy comes from a commit dated October 17, 2026. It has no certificate with Microsoft, OpenAI, Truepic or Samsung as the subject. A "trusted" result on this site means a match here.
• Interim Trust List. A list named for its freeze date, January 1, 2026, published in the verify-site repository. It includes Microsoft, Truepic, Samsung, Leica, Sony, Canon, Nikon and others. We use it as a second, legacy check. A match is labelled "legacy trust-list match" and is never counted as a match on the C2PA Trust List.

Here is what happened when we ran real sample files through both lists on October 27, 2026. It is a small test, not a survey.

In all nine files the record's own signature validated. Where the signer was not trusted, the cause was the certificate, not a broken record, so the result is "signer trust not established" and not "failed validation". We do not know whether newer Adobe files behave differently from our three samples. Privacy and method explains why we show OpenAI files this way and never add trust anchors of our own.

What a record cannot tell you

• That its statements are true. A trusted signature shows who made a claim, such as "this tool created this image". It does not audit the claim.
• Anything once it is gone. Most images lose their records when they are shared, screenshotted or re-saved. No record is not evidence of anything.
• Revocation. We do not ask online services whether a certificate has since been revoked, because that would send a request from your browser to a third party.
• Records stored elsewhere. A file can point to a record held online instead of containing it. We never fetch it, and we say the credentials are not available locally. Adobe says it backs up credentials in its public Content Credentials cloud so they can be recovered with Adobe Inspect.

How to check one

Drop the file into our checker, which runs entirely in your browser. For a second look, Adobe Inspect and Verify also read Content Credentials. They are other companies' sites, and anything you upload goes to them. How to tell if an image is AI covers the other kinds of signal, including SynthID.

Frequently asked questions

Are Content Credentials and C2PA the same thing?

Content Credentials are records written to the C2PA standard. People use the two names for the same thing, with C2PA being the technical name.

Can a Content Credentials record be faked?

Anyone can sign a record with their own certificate. A checker should show that the signer is not on a trust list, which is why we test trust separately from the signature.

If a record names an AI tool, is the image AI-generated?

The record says what its signer states, and a trusted record shows that the signer made that statement. We report it as what the signed record says, never as our own conclusion about the image.

Why does a file from ChatGPT say "signer trust not established"?

In our test, OpenAI's certificates were on neither list, and two older samples had expired certificates. The signature itself was valid. For the same reason we cannot tell a genuine OpenAI file from one signed with a certificate that someone made and named to look like OpenAI's.

Do Content Credentials survive a screenshot?

No. A screenshot makes a new file with no record. Re-saving or uploading can drop the record too.