Category leaders

Guide

What Is C2PA? Content Credentials Explained

How C2PA and Content Credentials work: what a signed credential contains, how it is tied to a file, what a valid credential shows and does not show, how to check a file, where the standard is used, and its known limits.

September 30, 2026 · Updated November 13, 2026 · The AI Rankings

Quick answer: C2PA is an open technical standard for recording where a piece of digital content came from and what has been done to it. Content Credentials is the name for the signed record itself. A Content Credential is attached to a file. It lists the device or software that made the file and the edits applied to it, and it is cryptographically signed, so a change to the file or to the record can be detected. The standard is published by the Coalition for Content Provenance and Authenticity (C2PA), whose steering committee is Adobe, Amazon, the BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony, TikTok and Truepic (C2PA). A valid credential shows which device or software produced the record and that the content has not changed since it was signed. It does not show that the content is true. A file with no credential may still be AI-generated, because a credential is metadata and can be lost when a file is converted, re-saved or captured in a screenshot.

This guide covers what a credential contains, how it is checked, where the standard is used and its known limits. For invisible watermarks such as SynthID, see what is AI watermarking. For tools that estimate from the image alone whether it is AI-generated, see best AI image detectors and how AI detectors work.


Four names are used around the standard.

NameWhat it is
C2PAThe Coalition for Content Provenance and Authenticity, which publishes the standard. The name is also used for the standard itself. C2PA describes itself as “a project of the Joint Development Foundation, a Washington-based 501c6 non-profit” (C2PA)
Content CredentialsThe name of the signed record, and of the technology as a whole. The specification calls a Content Credential “the preferred non-technical term for a C2PA Manifest” (C2PA specification)
Content Authenticity Initiative (CAI)A group announced by Adobe, The New York Times Company and Twitter at Adobe MAX 2019 (Adobe). C2PA says it joined the efforts of the CAI and Project Origin to develop open technical standards (C2PA)
Creator Assertions Working Group (CAWG)The group describes itself as “a working group of the Decentralized Identity Foundation”. It publishes the identity assertion, an optional addition to a credential (CAWG)

C2PA’s founding press release is dated 22 February 2021, and it released its specifications in draft for review on 1 September 2021 (C2PA).

What a Content Credential contains

The specification defines a nested structure.

ComponentDefinition in the specification
Manifest store”A collection of C2PA Manifests that can either be embedded into an asset or be external to its asset”
Manifest”The set of information about the provenance of an asset based on the combination of one or more assertions (including content bindings), a single claim, and a claim signature”
Assertion”A data structure which represents a statement either made (or ‘created’) by the signer or simply gathered at claim generation-time, concerning the asset”
Claim”A digitally signed and tamper-evident data structure that references a set of assertions, concerning an asset, and the information necessary to represent the content binding”
Claim signature”The digital signature on the claim created using the private key owned by a signer”
IngredientContent from another asset that was used to make this one. The specification says the parts of a composed asset are “referred to as ingredients”

Definitions are from version 2.5 of the C2PA technical specification.

In short, assertions are the statements, the claim references them, and the signature makes a later change to either detectable.

The specification groups its standard assertions by function. They include:

An action can carry a digitalSourceType field that says how the content was made. The specification’s example is a video generated by an AI model from a text prompt. Its manifest has a c2pa.created action whose digitalSourceType is the IPTC value trainedAlgorithmicMedia.

Hard bindings and soft bindings

A binding ties a manifest to the content it describes. The specification defines two kinds:

A hash matches only the exact content it was computed from. A soft binding is a way to find a manifest again. The specification’s example is an asset that “can become separated from its C2PA Manifest due to removal or corruption of asset metadata”, where a fingerprint “could be used to search a database to recover the asset with an intact C2PA Manifest”.

How it works

  1. Capture or generation. A camera, phone, application or AI model creates the file and writes a manifest. The manifest records what made the file and holds a hash that binds it to the content.
  2. Editing. An editing tool that supports the standard records the original as an ingredient, adds actions describing what it did, and signs a new manifest. The manifest store then holds a chain.
  3. Signing. Each manifest is signed with a certificate. C2PA’s FAQ says signing certificates “are issued by Certification Authorities (CAs) listed on the C2PA Trust List” (C2PA).
  4. Distribution. The file is shared. The manifest stays with it only where each tool and platform that handles the file keeps the metadata.
  5. Validation. A validator checks the signature, the certificate and whether the hash still matches the content, and reports the result.

What a validator reports

The specification defines three states for a manifest and one for the file.

StateWhat it means in the specification
Well-formed manifestThe manifest follows the specification’s rules for its contents, its assertions and its ingredients
Valid manifestThe manifest is well-formed, it “has not been modified since the manifest was signed”, its signature validates, and the signer’s credential is not rejected as revoked
Trusted manifestThe manifest is valid and its signing credential is trusted. A validator checks this against the trust anchors it keeps
Valid assetThe parts of the file “that are covered by content bindings have not been modified since the active manifest was produced”, and that manifest is valid or trusted

The specification leaves the conclusion to the reader. It says the person consuming the file “has to interpret those statements to arrive at a set of conclusions of their own about the provenance of the asset”. A file with no manifest gives a validator nothing to report.

What a Content Credential shows and does not show

It shows which claim generator produced the record. The specification defines a claim generator as the hardware or software that generates the claim, and says that if a manifest is valid, “the manifest’s claim can be attributed to the claim generator”. If the asset is also valid, the content covered by the bindings has not changed since.

It does not show:

Content Credentials, watermarks, labels and detectors

Four approaches are used to identify AI-generated or edited content.

ApproachWhat it isCovered in
Content CredentialsSigned metadata attached to a fileThis guide
Invisible watermarkA signal embedded in the content itself, read with the provider’s detectorWhat is AI watermarking
Visible labelA mark shown to the person viewing the contentWhat is AI watermarking
AI detectorSoftware that estimates from the content alone whether a model produced itBest AI image detectors, best AI detectors, best deepfake detectors

The first two are used together. The specification treats a watermark or a fingerprint as a soft binding that points back to a manifest. The AI glossary defines the terms used here.

C2PA’s FAQ says Content Credentials “are not a form of DRM” and “record and preserve the provenance of digital media without limiting its use” (C2PA). On 22 January 2026 C2PA said its technical specification “does not contain a standard assertion related to TDM (Text/Training and Data Mining)” (C2PA).

Timeline

DateEvent
2019Adobe, The New York Times Company and Twitter announce the Content Authenticity Initiative at Adobe MAX (Adobe)
22 February 2021C2PA’s founding press release
1 September 2021C2PA releases its specifications in draft for review
December 2021Version 1.0 of the specification
January 2024Version 2.0 of the specification
8 February 2024Google joins C2PA as a steering committee member
7 May 2024OpenAI joins the steering committee
5 and 12 September 2024Meta and Amazon join the steering committee
17 September 2024Google says it will bring Content Credentials to About this image in Search (Google)
June 2025C2PA opens its conformance programme (C2PA)
27 August 2025Nikon releases Z6III firmware version 2.00 with a C2PA function (Nikon)
December 2025Version 2.3 of the specification
22 January 2026C2PA says the specification has no standard assertion for text and data mining (C2PA)
9 February 2026C2PA announces Content Credentials 2.3 and marks five years since its launch (C2PA)
April 2026Version 2.4 of the specification. A security analysis of versions 2.2 to 2.4 is dated 23 April (arXiv)
27 July 2026C2PA announces that TikTok has joined the steering committee (C2PA)
31 July 2026C2PA publishes a guide to labelling AI-generated content with Content Credentials (C2PA)
17 August 2026Version 1.3 of the CAWG identity assertion, a DIF Ratified Specification (CAWG)
October 2026Version 2.5 of the specification

Specification dates are from the version history in the C2PA specification. The 2021 dates and the dates members joined are from C2PA’s news index.

Versions, membership and conformance

ItemWhat the source says
SpecificationThe version history dates version 2.5 to October 2026, version 2.4 to April 2026 and version 2.3 to December 2025. Version 2.5 added support for more file formats, extended live media beyond video and allowed ML-DSA post-quantum signatures (C2PA specification)
MembershipC2PA said on 9 February 2026 that “More than 6,000 members and affiliates have live applications of Content Credentials” (C2PA)
Steering committeeEleven members, named at the top of this guide. An announcement datelined 27 July 2026 said TikTok had moved from General Member to Steering Committee Member (C2PA)
Conformance programmeOpened in June 2025. C2PA describes it as a process “intended to hold generator products, validator products and certification authorities accountable” to the specification, the Certificate Policy and the Security Requirements (C2PA)
Trust ListC2PA’s FAQ calls it “a curated list of Certification Authorities (CAs)”, and says “Only products evaluated and approved by the Conformance Program are eligible to receive certificates from these trusted CAs” (C2PA)
Earlier trust listThe FAQ calls the earlier list the ITL. It says content signed with ITL certificates “will continue to validate in tools that use the ITL”, and that validators “may choose to respect the ITL” but “are not required to do so” (C2PA)
GuidanceOn 31 July 2026 C2PA published a guide to using Content Credentials to label AI-generated and AI-modified content, with a companion Deployment Guidance document (C2PA)
IdentityVersion 1.3 of the CAWG identity assertion is a DIF Ratified Specification dated 17 August 2026. It lets a credential holder “prove control over a digital identity” and record the role that identity played (CAWG)
ISOA draft international standard, ISO/DIS 22144, was in ISO’s fast-track process in March 2025, as set out below

ISO standardisation

A March 2025 page from the Association for Information Science and Technology (ASIS&T) covers a draft international standard, ISO/DIS 22144, titled “Authenticity of Information – Content Credentials”. It says the draft “is in the ISO Fast-Track process which means with a positive vote, it goes to publication” (ASIS&T). ISO’s catalogue entry for ISO 22144 records the stage the standard has reached.

Where Content Credentials are used

WhereWhat the source says
AI image generatorsChatGPT, Codex and the OpenAI API images carry C2PA metadata plus a SynthID watermark (OpenAI). OpenAI states coverage “can vary by product, model, export path, file type”. The Next Web reported on 19 May 2026 that OpenAI was adopting C2PA and embedding Google’s SynthID watermark “across OpenAI’s image outputs” (The Next Web)
Files made by AI assistantsAnthropic says that when Claude produces a file of a supported type, “such as a .png, .jpg, or .svg”, it attaches a content credential “saying that the file was made or processed with Claude” (Anthropic)
PhonesThe CAI’s senior director, Andy Parsons, wrote on 18 January 2026 that the Google Pixel 10 phone launched “supporting C2PA credentials” (CAI)
CamerasThe same post cites Sony’s PXW-Z300 video camera, “which brings Content Credentials directly into high-end video capture”. Nikon added a C2PA function to the Z6III on 27 August 2025 (Nikon) and suspended the service the following month, as set out under Limits
Creative softwareThe same post says Adobe’s Content Authenticity for Enterprise “addresses the realities faced by brands, publishers, and institutions working with large-scale production workflows”
SearchGoogle said on 17 September 2024 that it would bring Content Credentials to its About this image feature in Google Images, Lens and Circle to Search in “the coming months” (Google)
AssistantsThe Gemini app checks an uploaded image, video or audio file for SynthID and Content Credentials (Google)
Social platformsTikTok said on 19 November 2025 that it would continue “reading C2PA Content Credentials and adding them into AI-generated content made on TikTok” (TikTok)

How to check a file’s Content Credentials

There are three ways.

  1. Open the file in a C2PA reader. Anthropic says of a Content Credential that “Any C2PA-aware tool can read it” (Anthropic). A reader shows the manifest chain if one is still attached to the file.
  2. Ask the Gemini app. Upload the file and ask whether it was created or edited with AI. Google says Gemini uses two technologies for this check, SynthID and Content Credentials. Its help page sets the limits: a file of 100 MB or less, and “approximately 10 image checks, 10 video checks (up to 5 minutes total) and 10 audio checks (up to 3 hours total) within any 24-hour rolling window”. Google advises cropping tightly around an image and not uploading a collage. It says verification with Content Credentials “is available at gemini.google.com and in the Gemini app for Android”, and is coming to the Gemini app for iOS (Google).
  3. Use the checker of the company whose model may have made the file. For suspected OpenAI output, openai.com/verify checks for both the C2PA metadata and the SynthID watermark, and OpenAI offers a verification API for volume. The Next Web reported on 19 May 2026 that OpenAI was previewing a public tool “that checks for both C2PA credentials and the SynthID watermark” (The Next Web). For a file that may have come from Claude, Anthropic’s free Claude Content Checker “checks a file for a Content Credential from Claude”. It runs in the browser, accepts files of up to 100 MB and does not check text (Anthropic).

A file can also declare how it was made in ordinary metadata. The IPTC’s guidance of 9 May 2023 recommends that generated images and video carry the Digital Source Type value trainedAlgorithmicMedia in their XMP data, and says “the work has been picked up by the Coalition for Content Provenance and Authenticity (C2PA) via the use of digitalSourceType in Actions” (IPTC).

For the steps to take when a file has no credential, see how to tell if an image is AI-generated.

Limits

A credential can be removed

A credential is metadata, and handling that does not keep metadata removes it. Anthropic lists “format conversion, re-saving, screenshots, or other means” (Anthropic). OpenAI says of its own files that Content Credentials “can sometimes be removed by platforms, editing tools, or file conversions”.

The CAI’s response is durable Content Credentials. Andy Parsons described them on 8 April 2024 as a combination of three things (CAI):

  1. Secure metadata, the signed record, stored in a way that “cannot be altered without leaving evidence of alteration”.
  2. Watermarking, hidden information “not detectable by humans”. Parsons wrote that watermarks “can be impervious to alterations such as the cropping or rotating of images”.
  3. Fingerprinting, “a unique code based on pixels, frames, or audio waveforms” that “can be re-computed on the fly and matched against a database of Content Credentials”.

Parsons wrote that, combined, “the three form a unified solution that is robust and secure enough to ensure that reliable provenance information is available no matter where a piece of content goes”. The specification defines a Durable Content Credential as one “for which there exists one or more soft bindings that enable its discovery in a manifest repository”.

A security analysis of versions 2.2 to 2.4

A paper dated 23 April 2026, “Verifying Provenance of Digital Media: Why the C2PA Specifications Fall Short”, examined versions 2.2 to 2.4 of the specification. Its eleven authors are from the University of Maryland, Baltimore County (UMBC), Hacker Factor and the US National Security Agency (arXiv). It sets out six key findings.

Finding, as the paper titles itWhat the paper reports
Timestamps Can Be Forged or Altered”Nothing in the signed data references the timestamp, allowing removal and replacement without detection.”
Revoked or Compromised Credentials Are Still Accepted”Conforming validators are not required to check for revoked certificates.” The paper says many fail to check revocation properly and some “continue to accept known compromised credentials”
Different Tools Produce Contradictory ResultsFor an AI image signed with a Nikon Z6 III, “Adobe Inspect (pictured here) reports the signature as valid, while Verifieddit reports it as invalid.” The paper adds: “Neither conforming validator reports the revocation.”
Parts of Files Can Be Modified Without DetectionThe specifications “allow certain parts of a file (such as metadata) to be excluded from protection”. After the authors changed the GPS location in a Pixel 10 Pro image, the Proofmode Verify validator “does not detect our alteration and displays the false GPS location”
Credentials Expire and Become UnverifiableAn Arizona Secretary of State image “validated in January 2025, but fails to validate a year later”
Certification Does Not Ensure Compliance with the SpecificationsThe paper says “the conformance program does not define security requirements for conforming validators”, and that certification rests largely on self-reported compliance, with “no examination of the product’s functionality or source code”

The paper concludes that C2PA “should not yet be relied upon for high-stakes uses”, and gives financial disclosures, journalism and legal evidence as examples. Its recommendations include “strict certificate revocation checking”, consistency across validation tools and “independent security audits for certified products”.

The paper says the Pixel 10 Pro and version 2.3 of the specification “incorporated some of our suggestions”, and that version 2.4 “does not address any of our concerns”. Version 2.5 of the specification is dated October 2026, after the paper.

The Nikon example comes from 2025. Nikon added a C2PA function to the Z6III camera with firmware version 2.00 on 27 August 2025, as part of its Nikon Authenticity Service (Nikon). heise online reported on 23 September 2025 that Adam Horshack, a user of the DPReview forum, had shown how the camera’s multiple exposure function could be used to make it sign an image it had not captured, and that Nikon had temporarily suspended the service. Nikon said it was taking the matter very seriously and wanted to restore confidence in its services, heise reported (heise online). The paper says Nikon revoked the camera’s certificate in November 2025.

What the EU AI Act requires

Article 50 of the EU AI Act applies from 2 August 2026. It requires providers of AI systems that generate synthetic audio, image, video or text to ensure the output is “marked in a machine-readable format and detectable as artificially generated or manipulated”. Their technical solutions must be “effective, interoperable, robust and reliable as far as this is technically feasible”.

The article names no technology. It does not require C2PA, a watermark or any other named method. It also requires anyone who uses an AI system to produce a deepfake to disclose that the content was artificially generated or manipulated.

What is AI watermarking sets out how providers mark their output.


Frequently asked questions

What is C2PA?

C2PA is the Coalition for Content Provenance and Authenticity, and the open technical standard it publishes for recording where digital content came from and how it has been edited. The signed record the standard defines is called a Content Credential. C2PA describes itself as a project of the Joint Development Foundation. Its steering committee is Adobe, Amazon, the BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony, TikTok and Truepic. Version 2.5 of its specification is dated October 2026.

What are Content Credentials?

Content Credentials are the signed provenance records that the C2PA standard defines. The specification calls a Content Credential “the preferred non-technical term for a C2PA Manifest”. A manifest holds one or more assertions, which are statements about the file, a claim that references them, and a claim signature made with the signer’s private key. Assertions can record what made the file, what was done to it, which other files it was made from, and a hash that binds the manifest to the content. A later change to the content or to the manifest can then be detected.

Is C2PA an ISO standard?

In March 2025 a draft international standard titled “Authenticity of Information – Content Credentials”, numbered ISO/DIS 22144, was in ISO’s fast-track process, according to a page from the Association for Information Science and Technology. The page says a positive vote in that process sends a draft to publication. ISO’s catalogue entry for ISO 22144 records the stage the standard has reached.

Does a Content Credential prove an image is real?

No. A valid credential shows which device or software produced the record and that the content covered by its bindings has not changed since. It does not show whether what the image depicts happened. It also does not show that the image was not made by AI: an AI product can attach a valid credential, and the credential then records that a model made the file. A paper dated 23 April 2026 by researchers at UMBC, Hacker Factor and the US National Security Agency, which examined versions 2.2 to 2.4 of the specification, concluded that C2PA “should not yet be relied upon for high-stakes uses”.

Why do Content Credentials disappear from images?

A credential is metadata attached to the file, and handling that does not keep metadata removes it. Anthropic says a file’s metadata can be “stripped through format conversion, re-saving, screenshots, or other means”. OpenAI says of its own output that Content Credentials “can sometimes be removed by platforms, editing tools, or file conversions”. A file with no credential may therefore still be AI-generated. The Content Authenticity Initiative describes durable Content Credentials, which pair the metadata with a watermark and a fingerprint so that a record removed from a file can be found again.

What is the difference between C2PA and a watermark?

A C2PA Content Credential is signed metadata attached to a file. Any C2PA-aware tool can read it, and it can be removed. A watermark is a hidden signal embedded in the content itself, and reading it needs the provider’s detector. The specification calls a cryptographic hash of the content a hard binding, and a watermark or fingerprint a soft binding. What is AI watermarking covers watermarks.

How do I check if an image has Content Credentials?

Open the file in a C2PA reader, which shows the manifest chain if one is still attached. Or upload the file to the Gemini app, which checks for Content Credentials and Google’s SynthID watermark. Google’s help page gives a quota of approximately 10 image checks in any 24-hour rolling window. For suspected OpenAI output, use openai.com/verify, which checks for the C2PA metadata and the SynthID watermark that OpenAI embeds in images from ChatGPT, Codex and its API. For a file that may have come from Claude, Anthropic’s free Claude Content Checker checks for a Content Credential from Claude. Google said on 17 September 2024 that its About this image feature would show Content Credentials in Google Images, Lens and Circle to Search.

Which cameras and phones support Content Credentials?

The Content Authenticity Initiative’s senior director wrote on 18 January 2026 that the Google Pixel 10 phone launched “supporting C2PA credentials”, and that Sony’s PXW-Z300 video camera “brings Content Credentials directly into high-end video capture”. Nikon added a C2PA function to the Z6III with firmware version 2.00 on 27 August 2025, and heise online reported on 23 September 2025 that Nikon had temporarily suspended the service. A manufacturer’s own documentation says what a given model supports.

Does C2PA let me opt out of AI training?

The specification has no standard way to record that. C2PA said on 22 January 2026 that its technical specification “does not contain a standard assertion related to TDM (Text/Training and Data Mining)”. Its FAQ says Content Credentials “are not a form of DRM” and “record and preserve the provenance of digital media without limiting its use”.

Is C2PA required by law?

No law names it. Article 50 of the EU AI Act applies from 2 August 2026 and requires providers of AI systems that generate synthetic audio, image, video or text to mark the output in a machine-readable format so that it is detectable as artificially generated or manipulated. The article names no technology: it does not require C2PA, a watermark or any other named method.

Does C2PA identify who took a photo?

Not by default. C2PA’s FAQ says the core specification “does not support attribution of content to individuals or organizations, so that it can remain maximally privacy-preserving”. An optional addition does: the identity assertion published by the Creator Assertions Working Group, a working group of the Decentralized Identity Foundation. Version 1.3 is a DIF Ratified Specification dated 17 August 2026. It lets a credential holder “prove control over a digital identity” and record the role that identity played.

What is the C2PA Trust List?

C2PA’s FAQ describes the Trust List as “a curated list of Certification Authorities (CAs)”. Signing certificates are issued by the authorities on the list, and the FAQ says “Only products evaluated and approved by the Conformance Program are eligible to receive certificates from these trusted CAs”. A validator uses a trust list to decide whether a valid signature comes from a trusted signer. C2PA’s conformance programme opened in June 2025. The security paper dated 23 April 2026 says the programme “does not define security requirements for conforming validators”.

← All guides