Glossary

Provenance: the recorded origin and history of a file

Provenance is the documented origin and edit history of a piece of media, recorded in signed manifests, declared metadata fields and saved tool records, as opposed to detection, which guesses origin from the content itself.

Controlled example PNG file that carries this record: Content Credentials found

test-signed.pngPNG file · controlled example

  1. Signature + IHDR
  2. iTXt XML:com.adobe.xmp
  3. tEXt parameters
  4. tEXt prompt / workflow
  5. tEXt invokeai_metadata
  6. eXIf
  7. caBX C2PA manifestrecord
  8. IDAT pixels
  9. IEND

Content Credentials found

A signed manifest describes how the file was made.

Signer, recording tool, actions and the signature check result are reported separately, so you can weigh each one.

Open the Content Credentials checker →
Content Credentials found: where this record sits inside a PNG file and what the isGenAI report says about it. The picture is a code-drawn controlled example that carries this record; download it and check it yourself.

Kinds of provenance evidence

Provenance comes in layers of strength. Cryptographic provenance is a C2PA manifest signed by a certificate. Declarative provenance is an unsigned field such as the IPTC DigitalSourceType in XMP or an “AI generated” statement. Tool records are saved prompts and workflows written by generators. Matching provenance ties a file to an earlier copy by hash or perceptual fingerprint. Each layer answers a different question and can be edited or stripped to a different degree.

The EU AI Act’s Article 50 uses provenance mechanisms as the machine-readable marking providers must apply.

How isGenAI uses provenance

Every isGenAI finding is a provenance record with a named source: signed record, declared source type, AI statement, saved prompt or workflow, or matching copy. The report labels each with its class and strength and refuses to blend them into one score. The provenance checker tool and the Article 50 provenance page explain which transformations strip which layer.

isGenAI does not run a pixel-based guess in place of provenance.

Provenance versus detection

Detection estimates origin from pixels and can be wrong in both directions. Provenance reads what was recorded and can be absent. isGenAI leads with provenance because an absent record is an honest “unknown”, whereas a wrong estimate is a confident error.

In practice, start with the strongest layer the file still carries and work down. If a signed record exists, read it first; if only a declaration remains, weigh it as editable; if nothing remains, try the matching-copy check before concluding that the origin is unknown.

How to use this term when checking a file

  1. Upload the original file to the AI image detector or the tool named on the page.
  2. Find the finding that uses this term under “How we checked” and read the source field it names.
  3. Weigh it by kind: a signed record, a declared field, a saved prompt and a matching copy are different strengths of evidence.
  4. Remember that absence of a record is not evidence of a human author.

Open the provenance checker

Frequently asked questions

What is provenance in simple terms?

The paper trail of a file: who made it, with what, and what changed it, as recorded in the file or in signed records.

Is provenance the same as AI detection?

No. Provenance reads recorded history; detection guesses from the picture. isGenAI reports provenance.

What provenance survives a screenshot?

None inside the file. Only matching by fingerprint against a known copy can connect a screenshot to its origin.

Which provenance is strongest?

A valid, trusted C2PA manifest, because it is signed and tamper-evident. Declared fields and saved prompts are editable.

Sources

Definitions follow the published standards linked above; the isGenAI behaviour described is the deployed checker as of October 2026.