What exactly is C2PA and how does Content Credentials work?

C2PA (Coalition for Content Provenance and Authenticity) is an open technical standard that binds cryptographic provenance data to digital media files. It stores a 'manifest store' in a JUMBF container inside the file, containing assertions (actions, author, ingredients) sealed by an X.509 digital signature.

Does a valid C2PA signature prove that an image represents a real event?

No. A valid C2PA signature only proves that the file has not been altered since the signer signed it, and that the claims within the manifest belong to that signer. It does not prevent someone from photographing a computer screen displaying an AI image (the 'analog hole'), nor does it verify the objective truth of what was captured.

How do OpenAI, Adobe, and Leica implement C2PA differently?

OpenAI embeds C2PA manifests declaring the image as synthetic/generated by DALL-E 3 or ChatGPT. Adobe records editing actions (such as Generative Fill vs. standard brush adjustments) and ingredient histories. Leica uses hardware-level cryptographic chips (Secure Element) inside cameras like the M11-P to sign RAW sensor captures directly at the moment of shutter release.

Why do most images on Twitter/X, Instagram, and WhatsApp lack C2PA credentials?

Social media and messaging platforms routinely re-encode and compress uploaded images to optimize bandwidth and protect user privacy, stripping all metadata and JUMBF containers in the process. The absence of C2PA records on social media is completely normal and never proves that an image is synthetic or manipulated.

Can someone forge or fake a C2PA Content Credentials manifest?

An attacker cannot forge the cryptographic signature of a trusted entity like OpenAI, Leica, or Adobe without stealing their private keys. However, anyone can create a valid-looking C2PA manifest using a self-signed or unverified certificate. Validators must check both the mathematical signature integrity and the trustworthiness of the root certificate authority.

Technical provenance & forensics

How to Read C2PA Content Credentials — Manifests, Signatures & Provenance

Short answer: C2PA Content Credentials act as a tamper-evident digital nutrition label for media files. By inspecting the manifest store, cryptographic claims, and signing certificates, you can verify who created or modified an image (such as OpenAI for DALL-E 3, Adobe for Photoshop edits, or Leica for hardware camera captures). However, a valid signature is not a truth machine—it verifies provenance continuity, not physical reality.

Published 4 October 2026 · What Made It? research team

The Architecture of C2PA: Inside the Manifest Store

The Coalition for Content Provenance and Authenticity (C2PA) standard does not store plain text metadata that anyone can modify with a basic EXIF editor. Instead, it embeds an encrypted, cryptographically verifiable Manifest Store directly inside the media file:

Component	Technical Role	What You Can Inspect
JUMBF Container	JPEG Universal Metadata Box Format embedding the C2PA binary block within file headers (e.g., APP11 in JPEG, caPI in PNG).	Presence of C2PA markers without needing to decrypt or contact an external server.
Active Manifest	The most recent provenance record describing the current state of the media asset.	The latest author, software application, creation timestamp, and primary action taken.
Assertions & Actions	Individual data claims describing how the file was produced, edited, or combined.	`c2pa.created` (capture/synthesis), `c2pa.edited` (modifications), and `c2pa.placed` (composited elements).
Claim & Hash Binding	A cryptographic SHA-256 hash calculated across all assertions and image byte offsets.	Guarantees that no pixels or metadata values were modified after the claim was formed.
Digital Signature	An X.509 public key cryptographic certificate signed by an authorized Certificate Authority (CA) with a timestamp authority (TSA).	Identity of the signing organization (e.g., Adobe Systems, OpenAI, Leica Camera AG) and validity period.

How Major Implementations Differ: Adobe vs. OpenAI vs. Leica

Not all C2PA manifests mean the same thing. Depending on where the credential was generated, the assertions tell fundamentally different stories:

1. Adobe (Photoshop, Lightroom & Firefly)

When an artist edits an image in Adobe Photoshop with Content Credentials enabled, the active manifest records an audit trail of editing operations:

Tool Distinction: The manifest differentiates between standard adjustments (levels, curves, cropping) and generative AI features (Firefly Generative Fill).
Ingredient History: If multiple source photos were composited into one canvas, the manifest lists the parent ingredients, showing where each visual element originated.
Signer Identity: Signed directly by Adobe's corporate certificate.

2. OpenAI (DALL-E 3 & ChatGPT)

OpenAI embeds C2PA manifests into images generated through ChatGPT and the OpenAI Images API:

Explicit Synthetic Declaration: The assertion explicitly declares the digital asset as synthetic media created by generative algorithms.
No Ingredient Chain: Because the image was synthesized from pure algorithmic diffusion rather than a camera sensor, there is no upstream photographic ingredient.
API vs. Web: Both ChatGPT web outputs and direct API responses carry OpenAI's signing keys, allowing instant detection if the file remains unaltered.

3. Hardware Cameras (Leica M11-P, Nikon Z6 III, Sony α9 III)

Hardware-based provenance represents the cutting edge of physical photo authentication:

Secure Element Hardware: Leica M11-P includes a dedicated cryptographic secure element chip inside the camera body.
Sensor-Level Signing: The camera computes the cryptographic hash and signs the RAW and JPEG files the exact millisecond the sensor captures light, embedding the camera serial number and hardware public certificate.
Tamper Proof: Even if someone alters a single pixel on the memory card, the signature check immediately fails.

The 3 Critical Limitations: Where C2PA Stops

To evaluate media responsibly, analysts and users must understand the engineering boundaries of C2PA. A verified signature does not solve authenticity on its own:

Limitation	How It Occurs	Forensic Takeaway
The "Analog Hole" (Re-photography)	An attacker displays a photorealistic Midjourney AI image on an 8K monitor and photographs the screen using a Leica M11-P camera.	The resulting file has a 100% valid, authentic hardware C2PA signature from Leica. The signature proves a camera captured light; it does not prove the subject existed in the real world.
Social Media Metadata Stripping	Platforms like X, Instagram, WhatsApp, and Reddit re-compress images on upload, stripping all EXIF and JUMBF boxes.	Over 98% of images on social networks have zero C2PA data. The absence of a manifest is never evidence that an image is AI-generated or manipulated.
Self-Signed / Untrusted Signatures	Anyone with basic developer tools can create a private cryptographic key and embed a valid C2PA structure claiming anything.	Mathematical validity is useless without verifying the Certificate Authority (CA) root of trust. Always verify who issued the certificate.

Step-by-Step: How to Inspect C2PA Credentials

When evaluating an image you suspect contains Content Credentials, follow this verification sequence:

Check for Local Container Markers: Open our free browser tool, C2PA Viewer. It inspects the file on your device to confirm whether JUMBF boxes and C2PA markers are present without uploading your image to external servers.
Inspect the Active Manifest: Use a full C2PA validator (such as Content Credentials Verify or the C2PA CLI) to unpack the manifest and inspect the signing certificate and declared actions.
Cross-Reference Hardware EXIF: Use our Free Image Metadata Checker to inspect camera shutter speed, aperture, and ISO settings. Genuine hardware captures will show consistent optical physics and matching EXIF timestamps.
Analyze Visual Artifacts: If credentials are missing or suspicious, scan the file in our multi-signal authenticity scanner to look for synthetic texture patterns, edge consistency, and generator fingerprints.