This article was produced with AI assistance. Editorial standards apply — see our About editorial process.
Myth: you can always tell
You cannot reliably tell human video from AI video by watching once. Artifacts (hands, physics, on-frame text) still appear, and they are not a proof of authenticity when they are absent. Deepfakes, in the deepfake sense, depict a real person doing something they did not do; that is a different claim from a fully invented scene. Continue in the video cluster.
Provenance, not eyeballs
The C2PA 2.4 explainer defines Content Credentials as a cryptographically bound manifest (assertions, claim, signature). C2PA does not prove a statement is “true.” It shows whether provenance is well-formed, untampered, and signed by a trusted signer, and bound to the asset. Metadata can be stripped on re-upload.
OpenAI’s Sora system card (December 9, 2024) describes C2PA metadata on assets, visible Sora watermarks by default, and an option for paying users to download without the visible watermark while still embedding C2PA. The card states there is no single solution to provenance. Compare construction vs generation in CGI vs AI.
Parent silo: AI culture.
FAQ
Can you tell AI video from real video?
Not always. Visual tells help some of the time. Provenance signals are the designed answer, and they can be stripped.
Does C2PA prove a video is true?
No. It attests to a signed history of the file, not to whether the depicted event happened.
Did Sora ship without watermarks?
The system card describes visible watermarks by default, with a paid path to hide the visible mark while keeping C2PA.
