Provenance standard

C2PA content credentials, explained without the marketing language.

C2PA (Coalition for Content Provenance and Authenticity) is an open technical standard for attaching a signed, tamper-evident history to a media file. It's the most durable answer to 'was this AI?' that currently exists — when it's actually present.

How it works

A manifest, not a watermark.

A C2PA-compliant file carries a "manifest": a structured, digitally signed record of assertions — who or what created the asset, what tools touched it, and in what order. Assertions are hashed and chained, so any later edit either extends the manifest (if made with a compliant tool) or breaks the signature (if the file is edited with a tool that strips or ignores it, which is most tools today). The manifest is typically embedded in the file itself, in JUMBF format inside JPEG/PNG/HEIC containers, or as XMP metadata pointing to it.

Field

digitalSourceType

An IPTC-defined vocabulary term the manifest can carry, with values including trainedAlgorithmicMedia (fully AI-generated), compositeWithTrainedAlgorithmicMedia (AI-assisted composite) and digitalCapture (camera original). This is the single most direct machine-readable "this was AI" signal currently standardized.

Signature

Cryptographic, not cosmetic

The manifest is signed with a certificate chain back to an issuing authority (a camera maker, a generator vendor, a publisher). A valid signature means the recorded history hasn't been altered since signing — it does not mean the history is complete or that no earlier, unsigned edit happened before the chain started.

Adoption

Growing, but not universal

Adobe (Firefly, Photoshop), several camera makers, and some newsrooms support C2PA to varying degrees. Most social platforms strip it on upload today, and most images in general circulation carry no manifest at all — its absence tells you nothing about origin.

In practice

What this means for a normal verification workflow.

Treat C2PA the way you'd treat any signed document: a real credential when present is strong evidence, and its absence is simply "no credential," not "suspicious." Check for it, but build your workflow — reverse image search, metadata inspection, visual review — around the assumption that most files you check won't have one yet.

FAQ

C2PA questions.

Does C2PA prove an image is real or AI?

A valid, present manifest is strong evidence for whatever it asserts. A missing manifest proves nothing — most images online were never issued one.

Can C2PA credentials be faked?

The signature chain resists tampering with an existing manifest, but a bad actor can simply use a tool that doesn't attach one, or strip it before distribution. It's a provenance record, not an unbreakable seal on every image.

Do social platforms preserve C2PA data?

Inconsistently. Many platforms recompress or resave uploads in ways that strip embedded manifests, though some are adding support to preserve or display credential badges.

Worked example

What a manifest assertion roughly looks like.

Simplified, a C2PA manifest for an AI-generated image carries a chain of assertions similar in spirit to: creator tool = "Adobe Firefly 3", action = "created", digitalSourceType = "trainedAlgorithmicMedia", timestamp, and a signing certificate identifying the issuer. Each assertion is hashed, the hashes are chained together, and the chain is signed. Software that respects C2PA (rather than stripping it) reads that chain and can display a "Content Credentials" badge; software that doesn't simply passes the file through, credential intact but unread, or discards it during recompression.

Comparing provenance signals

C2PA vs. EXIF vs. PNG text chunks.

SignalTamper resistanceSurvives social re-upload?Tells you
C2PA manifestHigh — cryptographically signed chainRarely — most platforms still strip itStructured, standardized origin claim (digitalSourceType)
EXIF (JPEG)Low — plain fields, easily editedRarely — stripped on upload by most platformsCamera Make/Model, software, timestamps
PNG text chunksLow — plain text, easily edited or strippedRarely — most re-encoding drops ancillary chunksFull generator prompt/parameters, when present

Takeaway

None of the three survives typical sharing well.

All three are genuinely useful when you have the original, unshared file. All three are usually gone by the time an image has been through Instagram, WhatsApp or a screenshot. Build a verification workflow — reverse image search, visual review, source context — that doesn't depend on metadata surviving.