Report examples
Read img2det results as evidence, not a one-word verdict.
These anonymous examples show how to translate common image provenance reports into practical review language. Each case separates the final assessment, the main driver, and the action a reviewer should take next.
Marker-only AI provenance clue
Situation: A PNG downloaded from a generation workflow includes OpenAI-style and C2PA-like byte strings, but no trusted C2PA manifest can be verified from the submitted copy.
Final assessment: AI marker found, not verified
Main driver: Provider-related marker strings exist without a validated signature, trusted signer, or confirmed asset binding.
Review action: Quote it as marker-only evidence, ask for the original file, and avoid saying verified OpenAI provenance.
Trusted C2PA credentials
Situation: A JPEG arrives with a C2PA manifest, a valid signature, trusted signer status, and asset binding that matches the analyzed file.
Final assessment: Verified provenance evidence
Main driver: The credential chain is stronger than visual clues because the manifest is signed and bound to the file bytes.
Review action: Use the report to support file provenance, then still check caption accuracy, source context, and consent before publishing.
Screenshot with missing provenance
Situation: A screenshot of a suspicious image has no C2PA, no camera identity, and no useful byte-marker context.
Final assessment: Inconclusive
Main driver: The screenshot is a new capture that usually removes the original image provenance and metadata.
Review action: Treat the report as weak triage evidence and request the original download or upload before making a claim.
Social-media copy after metadata loss
Situation: A social-platform download has no Content Credentials and only compression-heavy forensic context, while the creator says the original file still has metadata.
Final assessment: No verified provenance in this copy
Main driver: The checked copy is platform-processed, so the missing metadata does not prove fake or camera-original status.
Review action: Run the social copy and the original side by side, then document which signals disappeared after platform processing.
Camera-like support without proof
Situation: A photo-like JPEG includes plausible Make and Model metadata, but no C2PA credentials and no independent source verification.
Final assessment: Camera-like support, no verified AI provenance
Main driver: Camera metadata can reduce suspicion, but it does not prove the depicted event or rule out editing.
Review action: Use cautious language and pair the file report with source checks, reverse search, and publication history.
Asset-binding mismatch
Situation: A file includes a manifest, but the signed asset binding does not match the bytes submitted for review.
Final assessment: C2PA integrity mismatch
Main driver: The credential may describe a different version or a file that changed after signing.
Review action: Treat this as a high-risk integrity signal, request the original asset, and do not rely on the credential as trusted provenance.
Use the examples with the live checker
Run the file first, then compare the report wording with the closest example. If the evidence is marker-only, missing, or platform-stripped, ask for the original file before making stronger claims.