Field notes / AI provenance
AI provenance · C2PA

How to check C2PA Content Credentials in an image

Content Credentials are closer to a signed history card than an “AI detector.” They can tell you useful things about origin and editing history when the credentials are present and trusted. They cannot prove that every image without credentials is human-made.

Published September 15, 2026 · CreatorPrivacyKit Research

Verified

C2PA is an open provenance standard used to attach signed claims about how media was created or changed.

Checkable

A file can be inspected for a C2PA manifest, signer information, assertions, and whether the credential validates.

Not implied

No credential does not mean “not AI.” Metadata can be absent, stripped, unsupported, or replaced by another provenance system.

What C2PA actually is

C2PA stands for the Coalition for Content Provenance and Authenticity. The standard lets a creator, camera, editor, or generative system attach cryptographically signed provenance information to a file. The visible picture is still the picture; the credential is a structured record associated with it.

That record can describe the tool that created or exported a file, editing actions, ingredients used in a composition, timestamps, and other assertions. The exact fields depend on the software and workflow. C2PA is not limited to AI. A camera can use provenance credentials, a newsroom can use them, and an AI image generator can use them.

How to check a file

  1. Keep the original file. Do not begin by taking a screenshot or re-saving it through another app.
  2. Use a C2PA-aware verifier. Look for a manifest and whether the signer and assertions validate.
  3. Read the claims, not just the badge. “Credential present” is less useful than who signed it and what the history says.
  4. Check for other provenance signals. Pixel-domain watermarks such as SynthID are a separate mechanism from C2PA metadata.
  5. Interpret a negative result narrowly. “No supported signal detected” is not the same conclusion as “this image was definitely not AI-generated.”

C2PA and SynthID are different layers

SignalWhere it livesWhat it can help establishMain limitation
C2PA / Content CredentialsStructured signed provenance associated with the media fileOrigin and history claims from an identified signerCredentials can be absent or lost during some workflows
SynthIDImperceptible watermark embedded in generated contentWhether supported content carries a watermark associated with a participating generatorDetection is system-specific and not universal across every AI model
Ordinary EXIF/XMPFile metadataCamera, software, timestamps, GPS, authoring fieldsUsually not cryptographically trustworthy by itself
Visual classifierModel inference over pixelsProbability-like evidence of synthetic visual patternsCan produce false positives and false negatives

OpenAI and Google now expose provenance verification

OpenAI’s current verification flow checks supported files for provenance signals associated with OpenAI tools, including trusted C2PA manifests and supported SynthID watermarks. A positive result can tell you that the file contains an OpenAI-associated provenance signal. It does not tell you that the image is factually correct, unedited, or owned by a particular person.

Google’s Gemini verification flow similarly combines SynthID with Content Credentials. For SynthID, the important boundary is vendor scope: a missing Google watermark can rule out some Google-generated paths without ruling out other generators.

Why a missing credential proves less than people expect

Provenance is strongest when you have a positive, verifiable chain. The absence of that chain is ambiguous. The image may have come from software that never added C2PA, from a generator that uses a different watermark, from a social platform that transformed the file, or from an editor that exported a fresh container without carrying the original metadata forward.

This is why “AI detector” pages should not turn one missing field into a certainty. A useful checker should report evidence by layer: metadata, C2PA, watermark support, and visual signals. It should also make clear which systems each check can and cannot identify.

Good verification language: “A supported provenance signal was found” is strong. “No supported provenance signal was found” is also useful. “Therefore this image is definitely real” is usually not justified.

What to save if provenance matters

If you are a creator, publisher, investigator, or newsroom, keep the original file rather than only a social-media derivative. Preserve the file hash, the credential, and the verification result. If you need a privacy-safe public copy, create that copy separately rather than destroying the only provenance-rich original.

Inspect before you trust or strip

CreatorPrivacyKit is building local-first tools around metadata, provenance, and file privacy. Use the toolbox to inspect what is in a file before deciding what should leave your device.

Open the toolbox →