C2PA is an open technical standard for recording where a piece of digital content came from and how it was changed. It’s published by the Coalition for Content Provenance and Authenticity, whose steering committee includes Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Sony, TikTok and Truepic (C2PA). The record it creates is called a Content Credential: a signed, tamper-evident label inside the file that can say which tool made the image, what edits were made, and whether generative AI was involved.
C2PA describes this as a “nutrition label” for digital content. Like a nutrition label, it tells you what went in, not whether you should believe what the image shows.
How Content Credentials work #
A Content Credential is technically a C2PA manifest. It has three parts.
- Assertions. Statements about the file: when and where it was created, which tools edited it, and whether AI was used (C2PA explainer).
- A claim that bundles those assertions together.
- A digital signature from whoever created the credential, such as a camera maker, an editing app or an AI generator.
The manifest is tied to the image with cryptographic hashes. If anyone changes the pixels or the provenance record without a C2PA-aware tool, the hashes no longer match, and a validator can tell the credential has been broken. That’s why it’s called tamper-evident rather than tamper-proof.
When a C2PA-aware editor changes the image, it adds a new manifest that references the old one. Over several edits, you get a chain showing the file’s history.
Who uses C2PA today? #
A few examples from official announcements:
| Company | What it does with C2PA |
|---|---|
| Adds C2PA metadata to images from Nano Banana Pro in the Gemini app, Vertex AI and Google Ads, announced November 2025 (Google) | |
| Google Search | “About this image” shows whether an image with C2PA metadata was created or edited with AI tools, announced September 2024 (Google) |
| Meta | Reads C2PA and IPTC “AI generated” signals to apply its “AI info” label (Meta) |
| TikTok | Reads Content Credentials to auto-label AI content from elsewhere, and attaches credentials to its own AI content that stay on when downloaded, announced May 2024 (TikTok) |
Adobe’s Firefly and OpenAI’s image tools are also associated with Content Credentials, and both companies sit on the C2PA steering committee.
How to check an image for Content Credentials #
- Content Credentials Verify. The free page at contentcredentials.org/verify, run by the Content Authenticity Initiative, reads a file’s credentials and shows the signer, the tools and any AI assertion.
- The CR icon. Some sites display a small “CR” pin on images that carry credentials. Clicking it shows the history.
- On a phone. Expose AI reads embedded Content Credentials as part of its metadata pass, along with XMP and generator tags, and recognizes signed C2PA credentials. If it finds a credible AI claim, that decides the verdict, and the “How we decided” card says so. If there are no credentials, which is the usual case for images from social media, its on-device model judges the pixels instead. Nothing is uploaded.
For a wider look at file metadata, see how to check image metadata for AI.
What C2PA can’t do #
This part matters as much as the rest.
It doesn’t say whether an image is true #
The C2PA explainer is explicit: Content Credentials don’t make value judgments about whether provenance data is “true,” and they can’t determine whether an asset depicts the truth. A real camera photo with valid credentials can still be staged, or captioned misleadingly.
Credentials can be removed #
Manifests usually live inside the file, and they can be separated from it. Most social platforms strip metadata on upload, and a screenshot creates a new file with none of the original’s credentials. The C2PA FAQ notes that “soft bindings” (invisible watermarks or fingerprinting) can help rediscover a credential stored in the cloud after it’s been removed (C2PA FAQ), but that depends on the creator having set that up.
Missing credentials prove nothing #
Most images, real and AI, have no credentials. An image without them isn’t suspicious by default. It’s just unlabeled.
Non-aware edits leave gaps #
If someone crops or edits a file in software that doesn’t support C2PA, the record may not reflect that change, or the credential may simply be dropped.
C2PA vs AI detectors #
C2PA and AI detectors answer different questions. C2PA tells you what a file says about its own history, backed by a signature. A detector estimates, from the pixels, whether an image was generated. C2PA is much stronger when it’s present, and useless when it’s absent. A detector always gives you something to go on, but only a probability. The sensible approach is to check for credentials first and fall back on a detector, which is how Expose AI orders its two passes. For how C2PA compares with invisible watermarks like Google’s SynthID, see SynthID vs C2PA.
Privacy #
According to the C2PA FAQ, the core Content Credentials specification doesn’t attribute content to individuals or organizations, so that it can stay as privacy-preserving as possible. Credentials can include location and time data, though, depending on what the creating device or app records.
Frequently asked questions #
Is C2PA the same as a watermark? #
No. C2PA is signed metadata stored with the file, which is easy to read but can be stripped. A watermark like SynthID is hidden in the pixels and is designed to survive edits, but you need a specific detector to read it. The two are complementary, and C2PA even allows watermarks as a way to recover stripped credentials.
Can Content Credentials be faked? #
They’re hard to forge because they’re cryptographically signed, and tampering breaks the signature. But someone could strip credentials from an AI image, or sign content with a certificate that isn’t trustworthy, so validators check who signed a credential as well as whether it’s intact.
Do phones add Content Credentials to photos? #
Some cameras and phones can sign photos with Content Credentials at capture, and support is growing. Check your device maker’s documentation to see whether yours does and whether it’s switched on.
Why don’t most images have Content Credentials? #
Many tools don’t add them yet, and most platforms strip metadata on upload. Credentials only help if they’re created at the start and preserved all the way to the viewer, and both halves are still being rolled out.