What Is C2PA? The Content Provenance Standard Explained

What Is C2PA and Who Created It?

C2PA stands for the Coalition for Content Provenance and Authenticity. It was announced on 22 February 2021 as a Joint Development Foundation project, with Adobe, Arm, BBC, Intel, Microsoft and Truepic as founding members. It merged two earlier efforts: the Adobe-led Content Authenticity Initiative and Project Origin, which Microsoft and the BBC ran to trace news media.

The coalition publishes the C2PA technical specification, now at version 2.4 (April 2026). The specification defines how provenance information is structured, signed, embedded in or linked to a file, and validated. Content Credentials is the consumer-facing name for the same records. The C2PA explainer says plainly that the credentials do not make value judgements about whether provenance data is true, and that provenance complements fact-checking and forensics rather than replacing them.

Adoption now reaches capture devices. Google announced on 10 September 2025 that Pixel 10 phones attach Content Credentials to JPEG photos taken in Pixel Camera, and that the app reached Assurance Level 2, the highest rating the C2PA Conformance Program defines today. Google Photos validates and displays credentials it finds.

How Does C2PA Work?

Every C2PA record starts with assertions: labelled statements about the asset, such as the capture device, the edit actions performed, the ingredients a composite was built from, or a digitalSourceType value saying the content was produced by a generative model. A claim then lists hashed references to those assertions together with a binding to the content, and the signer's private key produces the claim signature. Together these form a manifest, and the manifests for one asset sit in a manifest store, usually embedded in the file as JUMBF boxes.

When the file is edited in a conforming tool, the tool writes a new manifest that names the previous version as an ingredient. The last manifest in the store is the active manifest, and following ingredient links backward gives the provenance chain. The C2PA manifest guide on this site walks through that structure field by field.

What Are C2PA Hard Bindings and Soft Bindings?

A hard binding is a cryptographic hash over the asset bytes, or a defined part of them, recorded inside the claim. The specification offers several forms, including a plain data hash and a BMFF hash for MP4-style video. If even one protected byte changes outside a conforming tool, the hash no longer matches and the validator reports a mismatch. That makes tampering visible, but it also means routine re-encoding by a social platform breaks the binding.

A soft binding is a non-unique identifier such as an invisible watermark or a perceptual fingerprint. The specification says soft bindings must never stand in for hard bindings. Their job is recovery: when metadata has been stripped, a watermark decoder or fingerprint lookup can find the original manifest in a repository, which then has to be validated against the asset like any other. Durable Content Credentials combine both, so one channel can survive when the other is lost.

How the C2PA Trust Model Works

C2PA signatures use X.509 certificates, the same family of credentials that secures websites. During validation the verifier builds a chain from the signing certificate to a trust anchor. The C2PA Trust List names certificate authorities accepted for signing, and a separate TSA Trust List covers time-stamping authorities, whose RFC 3161 time-stamps let a signature stay checkable after the signing certificate expires. If no chain reaches an accepted anchor, the specification reports the status signingCredential.untrusted.

Verifiers are free to use other lists, including a private trust store for partners they know. That flexibility is deliberate, and it is also why a green check needs reading carefully. NIST AI 100-4 (November 2024) makes the key point: the most signed metadata can establish is that the signer claims the metadata is accurate. A camera signature does not reveal on-device AI processing the camera did not record, and a generator that never writes a credential leaves no record to check.

Does C2PA Detect Deepfakes?

No. C2PA records provenance that honest tools choose to write; it does not analyse pixels or audio. A deepfake made with an open-source face-swap model, a screenshot of a credentialed image, or a video re-recorded off a screen will usually arrive with no manifest. NIST AI 100-4 notes that if provenance becomes widespread, a missing record may become a signal in some contexts, but today absence mostly reflects non-adoption or ordinary platform processing.

That gap is where content-based detection works. Neural Defend's NeuroDetect analyses an authorised image, video, audio file or document inside a case and records model and version context next to whatever provenance the file carries, so an investigator can report both. NeuroVerify runs six detection signals in one API call for onboarding and account recovery selfies, returning a unified risk score with signal-level evidence in under 200 ms p95 on eligible image requests. The comparison of C2PA and forensic detection on this site covers when each answers the question you have.

How Should Teams Use C2PA Today?

Treat a valid credential as strong evidence about history and an absent credential as no evidence either way. A practical review order works for newsrooms, platforms and fraud teams alike: check whether a manifest exists; run validation and keep the exact status codes; read the signer and decide whether that signer is relevant to your question; follow ingredients if the file is a composite; and run content-based detection whenever the record is missing, invalid, or does not cover the claim being made.

Publishers who want their own output to carry credentials should sign at capture or export, keep a copy of the original manifest in a repository, and consider a soft binding so the record can be recovered after social platforms strip metadata.