C2PA is the Coalition for Content Provenance and Authenticity, and by extension the open technical standard it publishes for attaching a signed record of origin and edit history to a digital file. The record, called a Content Credential, travels inside an image, video, audio file or document. It states which camera, application or artificial intelligence (AI) model produced the file, what changed afterwards and which organisation vouches for those statements; a cryptographic signature makes later tampering detectable. The standard exists because ordinary metadata can be rewritten by anyone with a text editor, and because generative AI has made convincing fakes cheap enough that proving where a file came from is worth building infrastructure for. It does not judge whether content is true. It records who said what about it.

How a Content Credential is built

The record is layered. At the bottom sit assertions, labelled statements about the asset: an actions assertion (c2pa.actions) listing steps such as creation, cropping or colour correction; a thumbnail; ingredient assertions pointing to source files; and a hard binding, a cryptographic hash of the file's bytes, SHA-256 in the specification's examples. Since version 2.1, every standard manifest must open its history with either a c2pa.created or a c2pa.opened action.

Assertions are gathered into a claim, which the signing software, the claim generator, signs with a private key. The signature uses the CBOR Object Signing and Encryption (COSE) format and, since version 2.0, must rest on an X.509 certificate, the same public-key system that secures websites. A time-stamp from an RFC 3161 time-stamping authority is recommended; without one, a manifest stops validating once its certificate expires or is revoked. Assertions, claim and signature together form a manifest. Manifests accumulate in a manifest store, serialised in the JPEG universal metadata box format (JUMBF, ISO 19566-5) and embedded in the file: APP11 segments for a JPEG, a dedicated box for an MP4.

A worked example shows the history building up. A phone camera that supports the standard writes a manifest recording a digital capture, a thumbnail and a hash, then signs it. An editing application that opens the photo adds a second manifest, cites the first as an ingredient and lists its own actions. The last manifest in the store is the active one: a validator checks its signature, recomputes the hash and walks back through the ingredients. An image from a generative model instead carries a c2pa.created action tagged with the IPTC digital source type trainedAlgorithmicMedia, machine-readable shorthand for model-made.

Since version 2.1, validators report three states, well-formed, valid and trusted, the last meaning the certificate chains to an accepted list, typically the official C2PA Trust List. A soft binding, either a content fingerprint or an invisible watermark, lets a validator find a stripped credential again in an online manifest repository. Hard bindings detect alteration; soft bindings are built to survive it.

From Project Origin to version 2.4

Two efforts came first. Microsoft, the BBC, CBC/Radio-Canada and The New York Times convened Project Origin in 2019, and Adobe launched the Content Authenticity Initiative (CAI) with The New York Times and Twitter in November that year. On February 22, 2021, Adobe, Arm, the BBC, Intel, Microsoft and Truepic merged the work into the C2PA, a project of the Joint Development Foundation, an affiliate of the Linux Foundation.

Version 1.0 is dated December 2021 and was announced on January 26, 2022. Version 1.2, in October 2022, added the digitalSourceType field on which AI labelling now depends. Version 2.0, in January 2024, was the break: X.509 certificates only, a default C2PA Trust List and the removal of assertions naming human authors, now handled by the separate Creator Assertions Working Group (CAWG). Version 2.1 (September 2024) closed every publicly disclosed security flaw, according to its change log. Versions 2.3 (December 2025) and 2.4 (April 2026) added live video, plain text, HTML embedding and a dedicated AI disclosure assertion.

Membership tracked the AI boom. Publicis Groupe became the first advertising holding company on the steering committee in June 2023. Google joined on February 8, 2024, OpenAI on May 7, 2024, Meta on September 5, 2024 and Amazon a week later. The C2PA Conformance Program opened on June 4, 2025, and 54 products from 31 companies had passed it by May 14, 2026, according to the coalition's deployment guidance. The Interim Trust List used by early implementations was frozen on January 1, 2026. Version 2 is on the International Organization for Standardization (ISO) fast track as draft standard ISO/DIS 22144.

Why it reaches media plans

Regulation turned a voluntary format into compliance plumbing. Article 50 of the EU AI Act became applicable on August 2, 2026. The accompanying Code of Practice, published in June and finalised with Commission guidelines on July 20, obliges signatory providers to apply at least two machine-readable marking layers, one of them signed and time-stamped metadata. Google signed on July 24 and Meta four days later, with fines under the Act reaching EUR 15 million or 3% of worldwide turnover.

Ad platforms already write and read the credentials. Google said in September 2024 that C2PA signals would inform its ad policy enforcement. It now embeds SynthID watermarks and C2PA markup in every image and video generated inside its advertising tools, independently of the AI label setting advertisers control, and cites C2PA beneath its How this ad was made panel. YouTube uses the metadata to apply AI labels creators cannot adjust. TikTok became the first video platform to read Content Credentials on May 9, 2024, and says it has labelled more than 3 billion videos using them alongside creator labels and watermarks; its Symphony suite attaches credentials to generated ads. Cloudflare's image service can preserve credentials through transformations.

The IAB built its disclosure rules on the same layer. Its January 2026 framework paired consumer labels with C2PA metadata, and version 2, issued on August 18, 2026, defines two custom assertions: com.iab.threshold records whether AI use met the labelling threshold, com.iab.disclosure whether a label was shown. The stakes are commercial, since a Raptive study found AI content cut reader trust by half and the IAB cites an NYU Stern finding of a 31.5% click-through loss on labelled ads.

Where it breaks

Stripping is the chronic weakness: a credential lives in the file, and upload pipelines rewrite files. When The Washington Post uploaded an AI-generated video to eight platforms in October 2025, only YouTube disclosed anything, inside the expanded description. "No platform that The Post tested kept the Content Credentials data on the video or let users access it," wrote Kevin Schaul. The IAB framework concedes that platforms render advertiser-encoded C2PA mainly for on-platform content, not third-party creative.

Signing hardware can fail. Nikon added credentials to the Z6 III with firmware 2.00 on August 27, 2025. Within days, Adam Horshack, a DPReview forum contributor, wrote: "I've found a way to circumvent the C2PA protections on the Z6 III using the camera's multiple exposure feature." He later had the camera sign a Gemini-generated picture of a pug flying a jet. Nikon suspended the service on September 5 and revoked every certificate issued.

Absence of a credential proves nothing, a caveat OpenAI and Anthropic both document. Provenance catches output from cooperating tools and misses everything else. A March 2026 arXiv preprint described "authenticated fakes" made by omitting one assertion field the specification permits, with no cryptography broken.

Privacy cuts the other way. The World Privacy Forum argued in a September 2025 review that C2PA detects no deepfakes but lays down shareable data about creators that can link to commercial, government or biometric identity systems. Google says Pixel 10 certificates are designed so that no image can be linked to another or to its owner.

Even the regulatory fit is disputed. The IPTC, the news-media metadata body, says only C2PA meets the EU code's signed-metadata criteria. Law firm Bird & Bird notes the code names no standard, referring only to established ones.

Not the same as

Content Credentials and the CAI. Content Credentials is the specification's non-technical name for a C2PA manifest and for the technology generally. The CAI is an Adobe-run advocacy community that builds open-source tools; it does not write the standard.

SynthID and other watermarks. SynthID hides a signal in pixels that survives screenshots but carries little information; a manifest carries detailed history but is lost on re-encoding. The specification can cite a watermark as a soft binding, making the two complements.

IPTC and EXIF metadata. Google Merchant Center required the IPTC trainedAlgorithmicMedia tag on AI-generated product images from February 2024. Those fields are unsigned and editable; C2PA reuses the IPTC vocabulary but signs it.

Platform attestations. The syntheticContentAttestationStatus field in Display & Video 360 and Campaign Manager 360 is an advertiser declaration stored by Google, not a signature travelling with the file. Nor does C2PA authenticate who operates a website, the gap Japan's Originator Profile targets.

Recent developments

TikTok said on July 10, 2026 that it was joining the steering committee, and the coalition confirmed the upgrade on July 27. Anthropic attaches signed C2PA metadata to .svg, .png and .jpg files from Claude models launched since August 2, 2026, while OpenAI has paired C2PA with SynthID on images since May 2026 and exposes an API that checks for both. YouTube Shorts avatar videos carry both, though whether buyers filter on the signal is unclear. Under an IAB Austria guide on AI labelling covered today, neither C2PA metadata nor SynthID amounts on its own to the visible label a qualifying ad needs. Generative systems already on the market must bring their marking into conformity by December 2, 2026, and code signatories owe a watermark-detection interoperability solution by February 2, 2027.

Timeline

  • 2019: Microsoft, the BBC, CBC/Radio-Canada and The New York Times convene Project Origin
  • November 2019: Adobe, The New York Times and Twitter launch the Content Authenticity Initiative
  • February 22, 2021: Adobe, Arm, the BBC, Intel, Microsoft and Truepic found the C2PA
  • December 2021: Version 1.0 of the C2PA specification is completed
  • January 26, 2022: The C2PA publicly announces its specification
  • March 16, 2022: Sony joins the steering committee
  • October 2022: Version 1.2 adds the digitalSourceType field to the actions assertion
  • June 5, 2023: Publicis Groupe becomes the first advertising holding company on the steering committee
  • November 2023: Version 1.4 adds embedding for ZIP-based formats and fonts
  • January 2024: Version 2.0 restricts signing to X.509 certificates and introduces the C2PA Trust List
  • February 8, 2024: Google joins the steering committee
  • May 7, 2024: OpenAI joins the steering committee
  • May 9, 2024: TikTok becomes the first video platform to implement Content Credentials
  • September 2024: Version 2.1 defines well-formed, valid and trusted manifest states
  • September 5, 2024: Meta joins the steering committee
  • September 12, 2024: Amazon joins the steering committee
  • September 17, 2024: Google announces C2PA use in Search and its ad systems
  • October 15, 2024: YouTube shows its first Captured with a camera label, on a Truepic video
  • Early 2025: Cloudflare Images adds an option to preserve Content Credentials
  • May 2025: Version 2.2 adds the Soft Binding Resolution API
  • June 4, 2025: The C2PA Conformance Program launches
  • August 27, 2025: Nikon adds C2PA signing to the Z6 III with firmware 2.00
  • September 5, 2025: Nikon suspends its authenticity service after a signing flaw is disclosed; certificates are later revoked
  • September 10, 2025: Google details Content Credentials on Pixel 10 at Assurance Level 2
  • September 2025: The World Privacy Forum publishes its technical review of C2PA
  • October 22, 2025: The Washington Post publishes tests of AI video disclosure across eight platforms
  • December 2025: Version 2.3 adds live video and unstructured text
  • January 1, 2026: The Interim Trust List is frozen
  • January 2026: The IAB publishes its AI Transparency and Disclosure Framework
  • April 2026: Version 2.4 adds an AI disclosure assertion and HTML embedding
  • May 14, 2026: 54 products from 31 companies have passed conformance
  • July 8, 2026: The C2PA publishes Content Credentials Deployment Guidance 1.0
  • July 20, 2026: The Commission publishes Article 50 guidelines and the finalised Code of Practice
  • July 27, 2026: The C2PA confirms TikTok's move to the steering committee
  • August 2, 2026: Article 50 transparency obligations apply across the EU
  • August 18, 2026: IAB framework version 2 defines two custom C2PA assertions
  • December 2, 2026: Deadline for generative systems already on the market to conform on marking
  • February 2, 2027: Deadline for code signatories to deliver watermark-detection interoperability

Summary

Who. The Coalition for Content Provenance and Authenticity, founded by Adobe, Arm, the BBC, Intel, Microsoft and Truepic and now steered by companies including Google, OpenAI, Meta, Amazon, Sony, Publicis Groupe and TikTok. Signers include camera makers, editing tools and AI model providers; readers include YouTube, TikTok, Meta and Google's ad systems.

What. An open specification for Content Credentials: signed, tamper-evident manifests embedded in media files that record origin, edits and AI involvement, validated against certificate trust lists and paired with watermarks or fingerprints for recovery.

When. The coalition formed on February 22, 2021 and released version 1.0 in December 2021. Version 2.4, from April 2026, is current as of September 2026, and EU rules that lean on signed metadata have applied since August 2, 2026.

Where. Inside JPEG, PNG, MP4, PDF and other files, across creation tools, social platforms, ad platforms and content delivery networks, with regulatory weight concentrated in the European Union.

Why. Advertisers and publishers now carry disclosure duties for synthetic creative, and C2PA is the main machine-readable way to record AI involvement. Its value depends on credentials surviving upload, trustworthy signing hardware and platforms choosing to display them, and all three remain uneven.