ProvenanceLensSEE THE ORIGIN.
GitHub
EN
Scan a file

Provenance Lens

Why AI Detectors Fail

AI detector scores are estimates, not provenance evidence. Learn why false positives happen and what technical verification can establish instead.

Updated: Published by Provenance Lens

What Provenance Lens checks4
  1. C2PA Content CredentialsC2PA 2.2Check available
  2. Embedded metadataXMP / EXIF / IPTCCheck available
  3. Unicode anomaliesUnicodeCheck available
  4. Anthropic text watermarkProvider-specificNot checkable

Support status

C2PA Content Credentials

Check available

Source

Embedded metadata

Check available

Source

Unicode anomalies

Check available

Source

Anthropic text watermark

UNSUPPORTED

Source

Anthropic image C2PA

Check available

Source

Evidence, not guessing

Provenance Lens reports only what can be checked: a cryptographically signed manifest, a metadata field, a specific codepoint, a documented provider mechanism. When a signal cannot be checked, it says so. It never converts silence into a verdict.

False positives have consequences

A score based on style cannot name a signer, reproduce a signature check or show a concrete codepoint. It should not be treated as provenance evidence.

What a report cannot prove

  • A valid signature proves who signed a manifest, not that the manifest's claims are true.
  • Metadata is trivially editable. Treat generator fields as an indication, never as proof.
  • Hidden Unicode characters have many innocent causes, including ordinary copy-paste.

Primary sources