Content Credentials (C2PA)

Content Credentials are cryptographically signed records attached to an image, video, audio file, or document that state who or what created it and how it was edited. They implement the open technical standard of the Coalition for Content Provenance and Authenticity (C2PA), which Adobe, Arm, the BBC, Intel, Microsoft, and Truepic founded in February 2021 to unify the technical work of the Adobe-led Content Authenticity Initiative and Project Origin, led by Microsoft and the BBC. "Content Credentials" is the public name for C2PA data and its "cr" icon. In generative AI, a credential can declare that an image or clip came from a text-to-image or video generation model, so viewers do not have to guess from the pixels.

How a C2PA Manifest Works#

The C2PA specification packages provenance into a manifest with three parts:

• Assertions: Statements about the asset, such as the capture device, editing actions like cropping or color correction, the ingredients it was made from, and a digital source type.
• Claim and signature: The claim lists hashes of every assertion and is signed with the X.509 certificate of the camera, app, or service issuing it, optionally with a trusted timestamp.
• Bindings: A hard binding is a cryptographic hash of the asset's bytes, so changing a single pixel breaks it. A soft binding, such as an invisible watermark or a perceptual fingerprint, lets a verifier look up credentials that were stripped from the file.

The digital source type uses the IPTC vocabulary, in which trainedAlgorithmicMedia marks content created entirely by a generative model. Manifests are usually embedded in the file and can also be stored in the cloud. Each edit in a C2PA-aware tool adds a new manifest that lists the previous version as an ingredient, building a chain of data provenance from capture or generation to publication. A verifier checks that chain step by step:

What Credentials Show and Their Limits#

A valid credential, which anyone can inspect by uploading a file to the Content Credentials verifier, proves that a known signer made specific statements and that the asset has not changed since. It does not prove that a photo depicts real events, since a camera can photograph a staged scene or a screen. The absence of credentials proves nothing either: they are optional, and screenshots, re-encoding, and many upload pipelines strip embedded metadata, which is why soft bindings exist. Trust also depends on the signer, so the C2PA runs a conformance program and trust list for products that issue credentials. Content Credentials therefore work alongside AI watermarking, which survives metadata stripping because the signal lives in the pixels, and detection of deepfakes, which looks for the visual and statistical cues of AI-generated images in content that carries no provenance at all.

AI regulation is pushing adoption. Article 50 of the EU AI Act, its main transparency provision, requires providers of generative AI systems to mark synthetic audio such as cloned voices, images, video, and text in a machine-readable, detectable format from 2 August 2026, and C2PA manifests combined with watermarks are among the techniques providers use.

Adoption Across Devices and Platforms#

• Cameras and phones: The Leica M11-P, launched in October 2023, was the first camera to sign photos with Content Credentials at capture. Google's Pixel 10 phones attach them to every photo taken with the Pixel Camera app, which reached Assurance Level 2, the highest level the C2PA conformance program currently defines.
• AI generators: Adobe Firefly and OpenAI's image models attach credentials stating that outputs were AI-generated.
• Publishing platforms: LinkedIn shows a "cr" label on supported images and videos that opens their credential details.

Credentials also help computer vision teams separate captured photos from synthetic data when assembling training data, since training repeatedly on untracked generated images can lead to model collapse. On Ultralytics Platform, each image can carry custom metadata that the dataset search matches, so a team that records a credential's digital source type there can find every trainedAlgorithmicMedia image before training.