C2pa
C2PA, the Coalition for Content Provenance and Authenticity, is an industry standards body and the technical specification it publishes for attaching tamper-evident provenance information to images, video, audio and documents. It was formed in 2021 by Adobe, Arm, the BBC, Intel, Microsoft and Truepic, joining two earlier efforts: Adobe's Content Authenticity Initiative, begun in 2019, and Project Origin, led by the BBC and Microsoft to authenticate news content. The user-facing name for the resulting metadata and its small "cr" icon is Content Credentials, while C2PA is the open specification underneath, hosted under the Linux Foundation's Joint Development Foundation.
The specification defines how a signed record, called a manifest, is built and embedded. A manifest holds a set of assertions, such as the creating device or tool, a list of editing actions, whether generative AI was used, and ingredients taken from other files. A claim collects those assertions and their hashes, and a claim signature, made with an X.509 certificate, seals the claim. The manifest is stored inside the file in a container format based on JUMBF, a box structure related to JPEG metadata, or kept externally and linked from the file. A hard binding ties the manifest to the exact bytes of the asset by hashing; if the asset is changed without re-signing, validation reports a mismatch. Version 1.0 of the specification appeared in January 2022, and later versions added features such as richer actions, support for more formats, and the possibility of retrieving a manifest by a watermark or fingerprint.
Adoption has spread across the capture, edit, generate and publish stages, though unevenly. Adobe built Content Credentials into Photoshop and its Firefly generator, camera makers including Leica and later Sony and Nikon added signing in selected models, and Truepic and Qualcomm worked on capture-time signing in hardware and phones. Google and OpenAI joined the steering committee in 2024, and OpenAI began attaching C2PA metadata to images from its DALL-E 3 model in early 2024. News organizations such as the BBC have piloted displaying credentials with their journalism. Counting logos is not the same as measuring effect, however, and the number of ordinary viewers who ever see or understand a credential remains small.
Trust in C2PA is not automatic; it is anchored in who is allowed to sign. A validator checks the certificate chain against a list of recognized certificate authorities and a conformance program, so a credential signed by an unknown party appears as "unverified" rather than as proof of anything. This design resists casual forgery but creates governance questions: who decides which signers count, how revoked or stolen keys are handled, and whether open-source tools and independent creators can participate without costly certification. Self-signed credentials are cryptographically valid yet carry no external trust, which is a frequent source of confusion for newcomers.
The limits are well documented by the project itself and by outside critics. Metadata is easily removed, so an attacker can strip credentials from a synthetic image and the viewer sees only the absence of a label, which is also what most genuine photographs look like today. C2PA records claims, not facts: a real camera can capture a staged scene and a human can sign a deceptive edit. Capturing device identity or location raises privacy and safety concerns, so the specification makes most assertions optional and allows redaction. Critics also warn that a credential regime might disadvantage people without access to compliant hardware, or lead to platforms treating unsigned content with undue suspicion. Supporters answer that C2PA never claims to detect fakes and is meant to give honest creators a way to prove authenticity.
A common misconception is that C2PA is a watermark or an AI detector, or that it labels content as AI-generated automatically. It is none of these. It is an interoperable envelope for signed statements, and whether a given image says it was generated by a model depends on the generating tool choosing to say so. Open-weight generators run locally have no obligation to sign anything. For that reason it complements watermarking and detection research, and it sits alongside regulation, such as transparency provisions in the European Union's AI Act that require marking of synthetic content and have encouraged machine-readable disclosure.
Read C2PA as a shared grammar for signed provenance claims rather than as a verdict machine: it lets cooperating cameras, editors, AI tools and publishers show their work in a form anyone can verify, but only where signers participate, only while the metadata survives, and only as far as the signer is trusted.