What Are Content Credentials (C2PA)?

You have probably seen the phrase on an image caption or a little "CR" pin in a corner: Content Credentials. It sounds like a stamp of authenticity. It is closer to a receipt. This is what one is, what it can tell you, and why its absence tells you almost nothing.

What a Content Credential actually is

A Content Credential is a block of provenance metadata tucked inside a file. It records things like which tool made the file, which tool edited it, and when. The block is signed with the issuer's certificate, so a reader can check that it has not been altered since it was written and can see who issued it.

The underlying format is C2PA, and inside a file the signed block is called a manifest. "Content Credentials" is the consumer name for the same thing. When ChatGPT, Firefly or a Claude-generated file carries provenance, that is a C2PA manifest under a friendlier label.

A manifest can hold more than an origin line. It can list a chain of edits, name the software used at each step, and optionally carry information the creator chose to attach. What it holds is up to whoever signed it, so two files can both have Content Credentials and say very different amounts.

Where the standard comes from

C2PA stands for the Coalition for Content Provenance and Authenticity, an open standards group whose members include Adobe, Microsoft, Google, OpenAI, Intel and the BBC. It publishes a free technical specification, which is why a manifest written by one company can be read by another company's tool.

The consumer-facing side is the Content Authenticity Initiative, which Adobe started in 2019. It runs the branding, the little "CR" marker, and a public Verify page. So C2PA is the standard, and Content Credentials is the name most people see.

A record, not a watermark

This is the distinction that decides what a result means, so it is worth stating plainly.

A watermark is designed to survive the things people assume will defeat it: screenshots, re-encoding, resizing, light edits. A Content Credential is not. It is a normal part of the file, so it comes off the moment the file is transformed in a way that drops metadata. That is a feature of how metadata works, not a flaw in C2PA.

Because of that, the two signals have different jobs. A pixel watermark answers "did this model make it," within the provider's own detector. A Content Credential answers "what does this file say about itself," which is only as trustworthy as the fact that the manifest is present and signed by someone you recognise.

What strips them

Plenty of ordinary handling removes a manifest without anyone meaning to. A screenshot rebuilds the image from scratch, so the credential is gone. Many social platforms re-encode uploads and drop metadata. Saving a copy through an editor that does not support C2PA writes a clean file. And the image metadata remover on this site deletes it on purpose, in your browser, when you ask.

So a file with no Content Credentials is not a suspicious file. It is the default state of most images on the internet. The absence of a manifest is the absence of a record, nothing more.

How to read one in a browser

DetectMark reads Content Credentials locally, with nothing uploaded. Drop a PNG, JPEG or WebP into the image tool and the local scan runs first, entirely in your browser: the C2PA manifest, EXIF, XMP, GPS and PNG text chunks are parsed and listed. If a manifest is present, you see what it says and who issued it. If it is missing or broken, the app reports that rather than turning it into a verdict.

Reading a manifest is a provenance check, not a watermark check. That is why the pixel-watermark row on the same scan reads Unknown: no browser can read SynthID without the provider's key. For OpenAI images you can go one step further and send the original file to OpenAI's verifier, but only after an explicit consent click. That flow lives on the OpenAI image verifier page and the AI image watermark detector, and it never runs automatically.

Adobe also runs a public reader at contentcredentials.org/verify, and Anthropic and OpenAI have their own checkers for files they generate. They all read the same kind of manifest.

What a manifest proves, and what it does not

A present, intact manifest signed by a name you recognise proves that the named issuer wrote that provenance into the file, and that it has not changed since. Combined with a pixel watermark, that is close to the strongest statement a file can make about where it came from.

A manifest issued by someone you have never heard of proves only that someone signed it. A missing manifest proves nothing at all. None of these is evidence that a human made the file, and none of them is a percentage. DetectMark does not convert an absent or unfamiliar credential into a "this is real" or "this is AI" score, because that would be a guess wearing the clothes of a verdict.

If you strip a credential from your own file, DetectMark confirms the removal by re-scanning the export. It will never claim to have removed a pixel watermark, because it cannot, and what DetectMark cannot remove explains why the pixel mark stays put no matter what a remover promises.

Who signs them today

The list of tools that attach Content Credentials keeps growing. OpenAI signs a manifest on ChatGPT images alongside a SynthID pixel mark. Adobe Firefly applies one to its output. Gemini images carry one next to SynthID, and Claude-generated files ship a signed manifest with the text watermark. The AI watermark status matrix tracks who does what, dated, so the gaps stay visible instead of hidden.