Skip to content
FactLockCam

Category distinction

Provenance vs AI detection

Asking “does this look fake?” is a classification guess. Asking “does this exact file match the record anchored at capture?” is a comparison against an earlier reference.

Verifies the exact file and signed record—not the physical truth of the scene.

What is being asked?
Detection Does this look synthetic?
Provenance Does this exact file match the record anchored at capture?
How is it answered?
Detection A classifier guesses from pixel statistics.
Provenance SHA-256 of the bytes is compared with the earlier signed record.
How does it age?
Detection Accuracy drifts as generators improve.
Provenance Deterministic for that file: identical bytes match; changed bytes do not.
What is the output?
Detection A probability score.
Provenance Three separate results: file, signature, public record.

Why the jobs differ

Detection models can be wrong as generators improve. A SHA-256 comparison against a signed capture record is deterministic for that file: identical bytes match; changed bytes do not.

FactLockCam does not publish an “AI detected” score. The Can You Tell challenge on /deepfake-defense/ shows an original match and an altered media_hash_mismatch from the shipping verifier.

Verifies the exact file and signed record—not the physical truth of the scene.

Next step

Skip the guess. Keep the reference.

FactLockCam does not publish an “AI detected” score. It records the original so a later copy can be compared.

Anchor your first capture How the comparison works