Media provenance and AI training data
Capture that is provably real – for newsrooms, archives and the models trained on them.
Why the recording is the weak point
Signed content credentials tell a reader that a file has not been altered since it was signed. They do not tell the reader that the scene was real – the standard itself says so.
For AI training, "guaranteed real" is becoming a premium property: datasets that are provably not contaminated with generated imagery.
Verified offline- Capture
- Media provenance
- Moment
- 2026-09-22 · 14:07:41 UTC (± 90 s)
- Presence
- device
- Level
- Basic
Photo desk to archive
Capture
The photographer's camera app produces a proof of real capture alongside the existing content credentials.
Proof
The proof travels with the file, inside the same open container. The level is visible to the picture desk and, later, to the reader.
Decision
The archive can label material as proven real capture – and license it as such.
What the proof establishes
- The image originated from a real scene, not from a screen, print or model.
- When it was captured.
- That it can be checked by anyone, offline, years later.
Publishers, agencies, archives, AI data providers.
Basic for volume, Witnessed for editorial evidence.
- C2PA Technical Specification 2.x, section 4.2.2.8: a valid manifest and signature do not prove that the captured content depicts a real scene; the specification names the recapture of a screen or print as out of scope.
Bring one workflow. We bring the proof layer.