What is C2PA? Content Credentials explained
C2PA is an open standard for attaching a signed record of how an image was made. Content Credentials is the name you'll see on the label. Very useful when it's there, and silent when it isn't.
Updated
The short version
C2PA stands for the Coalition for Content Provenance and Authenticity. It was founded in 2021 by Adobe, Microsoft, the BBC, Intel and others to agree on one way of recording where digital media comes from.
The standard describes a manifest: a small, cryptographically signed record stored with the file. It can say which device or app made the image, what was done to it afterwards, and whether AI was involved. If someone changes the image or the record without re-signing it, verification shows that the record no longer matches.
"Content Credentials" is the friendlier name for the same thing, often shown as a small "CR" icon.
What's inside a credential
- Who signed it: the camera, app or service that created the record, identified by a certificate.
- What happened: actions such as created, opened, edited, cropped, or generated by AI.
- Ingredients: earlier versions or other images that went into this one, each with its own record where available.
- A binding to the pixels, so changes made after signing can be detected.
- Optional extras, such as a creator's name, if the person or tool chose to include them.
Who adds them
A growing set of tools attach credentials. Some AI image generators from large companies mark their output as AI-generated. Some editing apps, including Adobe's, can record edits and AI features used. A number of newer cameras and phones can sign photos at the moment they are taken. Some platforms read credentials and show a label.
Coverage is patchy. Many tools add nothing at all, including most open-source image generators that people run on their own computers.
For photographers and newsrooms, signing at capture is the interesting part: it lets a photo carry a signed record of which camera took it, and of every edit after that, all the way to publication. For everyone else, the most common credential you're likely to meet is an AI tool declaring that it made the image.
Why most images have none
Even when a credential is added, it rarely survives the trip to your screen:
- Many social platforms and messaging apps strip metadata when you upload, and re-compress the image.
- Screenshots create a brand-new file with no history.
- Some editing and export tools don't carry credentials over.
- Anyone who wants to hide an image's origin can remove the record.
How to check credentials yourself
The simplest way is the free Verify tool at contentcredentials.org: upload the image and it shows any credential, who signed it and what it records. Some apps and browser extensions can show the same information.
When you read one, look at the signer first. A record signed by a known camera maker or AI service means more than one from an unknown issuer. Then look at the actions: "created with a generative AI tool" is about as direct as image evidence gets. Always check the original file if you can; a copy saved from a social feed has probably lost the record.
What Content Credentials can't do
- Prove anything by their absence. Most genuine photos online have no credential, and so do most AI images.
- Tell you whether the scene is true. A signed photo of a staged scene, or of a screen, is still signed.
- Survive every platform. Once stripped, the record is gone from that copy.
- Replace judgement. A credential is only as trustworthy as whoever signed it.
Watermarks are something else
Some AI companies also hide invisible watermarks in the pixels of images their tools make. Google DeepMind's SynthID is the best-known example. Unlike a C2PA record, a watermark lives in the image itself, so it can survive some screenshots and edits. But it can usually only be read with the maker's own tools.
Checkobot does not read SynthID or other invisible watermarks.
What Checkobot does with credentials
When you upload an image or paste a link, Checkobot reads any Content Credentials it carries. If the image still has signed credentials from a trusted AI tool that record AI generation or editing, it is flagged as AI, and the result says the credentials were the reason.
If there is no credential, nothing changes: the two detectors (whole-image analysis, and face analysis on the largest face) decide on their own. That's also why the checker takes uploads and links but not clipboard paste: browsers re-encode pasted images and strip their credentials on the way.
Questions
Does a missing credential mean an image is AI?
No. Most images online have no credential, because platforms strip them on upload. A missing credential tells you nothing either way.
Does a credential mean the image is true?
It tells you where the file came from and what was done to it, according to whoever signed it. It doesn't tell you whether the scene in front of the camera was genuine or staged.
Does Checkobot detect SynthID watermarks?
No. Checko reads C2PA Content Credentials when they're present, and runs its own two detectors. It doesn't read invisible watermarks.
Can I add Content Credentials to my own photos?
Yes, if your camera, phone or editing app supports it. Look for a Content Credentials or provenance option in its settings or export dialog. Remember that many platforms will strip the record when you upload.
Why should I upload the file instead of a screenshot?
A screenshot is a new file: it has no credentials and more compression. The original file keeps whatever record it had, and more of the detail the detectors need.