There is no single AI video signature
An AI video may carry several independent clues. A C2PA Content Credential can be attached to the video file as signed provenance. A generator may also place an imperceptible watermark in the frames, add a visible mark, or keep an internal record of the generation. A platform may add its own label after upload. The presence or absence of any one clue does not settle the whole question.
The distinction matters when a file is shared. Metadata can be lost when a service exports a new file; a watermark embedded in frames may survive some transformations. Google DeepMind describes SynthID as a watermark in generated video segments, while OpenAI describes multiple provenance signals for Sora videos. Those are vendor-specific systems, not a universal detector.
Where a Content Credential lives in MP4
MP4 and MOV are container formats made from boxes. The C2PA specification defines a designated uuid box for a manifest in BMFF-based assets. The manifest can record assertions about origin and editing and carries a digital signature. A compatible validator checks the signature and its binding to the asset; spotting a box alone is a narrower observation.
Free AI Wiper inspects supported MP4 and MOV metadata boxes, including recognized C2PA and XMP carriers. It can remove selected supported boxes and recheck the output. It does not validate a C2PA signature, scan frames for SynthID, identify a visible watermark, or query a platform's records. The video format page lists the supported workflow.
A useful way to investigate
Start with the original file, not a screen recording. Inspect its portable metadata and, if present, use a dedicated Content Credentials validator to assess the signature. Then consider the context of publication, visible marks and any platform label. A negative metadata result means only that this inspection found no supported portable record. For the broader limits of automated detection, read what AI detectors can prove.