C2PA, the Coalition for Content Provenance and Authenticity, is the industry body and the open technical standard it publishes for attaching cryptographically signed provenance records to images, video, audio and documents. In AI safety and across artificial intelligence more broadly, C2PA is the main interoperable mechanism for declaring that content was generated or edited by AI and for letting anyone verify who signed that declaration and whether the file changed afterward.
The consumer-facing name for C2PA data is Content Credentials. The coalition was formed in 2021 by technology and media organizations, and its specification defines how a manifest of assertions is bound to an asset, signed with an X.509 certificate and validated by any conforming reader. C2PA records provenance; it does not judge truth, which is the single most important point for interpreting it correctly.
How a C2PA manifest is built and verified
A C2PA manifest is a signed bundle of statements about an asset, embedded in the file or stored remotely and linked to it. Each manifest contains assertions, which are individual facts such as capture device, editing actions or AI generation, and a claim that lists hashes of those assertions. The claim is then signed, producing a claim signature in the COSE format that ties the statements to a certificate holder.
Binding to the content happens through a hard binding, a cryptographic hash over the asset bytes, so that any pixel or sample change breaks validation. Newer versions of the specification also allow soft bindings, such as invisible watermarks or perceptual fingerprints, which help recover a manifest after metadata has been stripped. When a file is edited, the editing tool adds a new manifest and references the earlier one as an ingredient, forming a verifiable provenance chain.
The table below lists the core components a validator inspects.
| Component | What it contains | Why it matters |
|---|---|---|
| Assertion | One fact, such as actions or creation tool | Carries the actual provenance information |
| Claim | Hashed references to all assertions | Prevents silent changes to individual statements |
| Claim signature | COSE signature with X.509 certificate chain | Identifies the signer and proves integrity |
| Hard binding | Hash of the asset bytes or byte ranges | Detects any modification of the content |
| Ingredient | Reference to a parent asset and its manifest | Links edits into a history |
| Manifest store | All manifests of an asset, in a JUMBF container | Holds the full chain inside the file |
How C2PA labels AI-generated content
C2PA labels AI involvement through the actions assertion and a digital source type value drawn from the IPTC vocabulary. A generator that creates an image from a prompt records an action such as created with the source type trained algorithmic media, while a tool that applies generative fill to a photograph records an edit with composite AI involvement. Readers can then display a disclosure such as an AI-generated badge.
The distinction between fully synthetic and partially edited content is essential for honest disclosure, because a lightly retouched news photograph and a fully synthetic scene carry very different risks. The table shows typical source type values and their meaning.
| Source type | Meaning | Typical producer |
|---|---|---|
| Digital capture | Recorded directly by a camera or microphone | Signing camera or phone |
| Trained algorithmic media | Fully created by a generative model | Image, video or audio generator |
| Composite with trained algorithmic media | Real content with AI-generated parts | Editor using generative fill |
| Algorithmic media | Created by non-learning algorithms | Procedural or rendering software |
| Composite synthetic | Mix of several synthetic elements | Compositing pipelines |
Regulators increasingly expect machine-readable marking of synthetic content. The transparency obligations of the EU AI Act, for example, require providers of generative systems to mark outputs in a detectable format, and C2PA is one of the leading candidate techniques alongside watermarking.
C2PA compared with watermarking and detection
C2PA, AI watermarking and AI content detection solve different parts of the provenance problem and work best together. C2PA provides rich, signed, human-readable history but depends on metadata surviving; watermarking embeds a signal inside the content itself; detection classifiers guess after the fact without any cooperation from the producer.
| Approach | Signal location | Strength | Main weakness |
|---|---|---|---|
| C2PA manifest | Metadata or remote store | Signed, detailed, verifiable history | Lost when metadata is stripped |
| Invisible watermark | Inside pixels or samples | Survives many re-encodes | Carries little information, can be removed |
| Detection classifier | None, inferred from content | Works on uncooperative content | Probabilistic, error-prone, easily outdated |
| Fingerprint lookup | External database | Recovers lost manifests | Requires a registry and near-duplicate matching |
The practical design pattern is layered provenance: sign a C2PA manifest, embed a watermark as a soft binding, and register a fingerprint, so that the manifest can be recovered even after a social platform removes metadata.
Trust, certificates and validation states
A C2PA signature is only as trustworthy as the certificate that produced it. Validators check the certificate chain against a trust list, and the coalition operates a conformance program that defines which signers and products may appear on the official list. A valid signature from an unknown self-issued certificate proves integrity but not the identity of the signer.
Readers therefore report several distinct outcomes, and interfaces should present them carefully rather than as a single pass or fail.
| Outcome | What it means | How to treat it |
|---|---|---|
| Valid and trusted | Intact content, signer on trust list | Provenance claims can be relied on as declared |
| Valid but untrusted | Intact content, unknown signer | Integrity holds, identity unverified |
| Invalid | Hash or signature mismatch | Content changed after signing |
| No manifest | No Content Credentials present | No conclusion about authenticity |
Limitations and common misconceptions
C2PA is frequently overestimated as a deepfake detector, which it is not. Its guarantees are narrow and precise, and most misuse comes from reading more into a manifest than it states.
- Absence proves nothing: most authentic and most synthetic media still carry no manifest, so missing credentials are not evidence of manipulation.
- Signed does not mean true: a manifest proves who made a statement, not that a staged scene or a misleading caption is accurate.
- Metadata stripping by upload pipelines removes embedded manifests unless soft bindings or remote manifests are used.
- The analog hole remains open: photographing a screen produces a new, unsigned capture of synthetic content.
- Key compromise or careless certificate issuance lets bad actors sign false claims, so revocation and trust list hygiene are critical.
- Privacy exposure can occur when manifests reveal location, device or author details that the creator did not intend to publish.
The short version
C2PA is like a tamper-evident label stuck to a photo or video. The label says who made the file, which tools touched it and whether AI created any part of it. The label is sealed with a digital signature, so if someone changes the picture or the label, the seal breaks and software can tell.
The label does not say whether the picture shows the truth. It only tells you where the file came from and how it was changed. If a picture has no label at all, that does not make it fake; many genuine photos simply do not have one yet.
Where it makes a difference
A regional news agency receives a dramatic flood photograph from a freelance photographer during a storm. The image arrives with Content Credentials created by the photographer’s camera, showing digital capture, a timestamp and a certificate issued to the camera manufacturer. The photo desk opens it in a verification tool and sees a valid and trusted signature, plus one later manifest recording a crop and exposure adjustment in an editing application.
On the same afternoon, a second flood image circulates on social media without credentials. A copy found through reverse search carries a manifest declaring trained algorithmic media from an image generator. The desk publishes the first photograph with its provenance panel visible and declines the second image, noting in its verification log that the AI origin was confirmed by a signed manifest rather than by guesswork. The result is a faster, documented decision and a public record readers can inspect themselves.