What Metadata Reveals AI-Generated Video

By Calabi Labs Editorial Team · 2026-08-06

AI-generated video is easiest to spot in its metadata, not its frames. The file carries a C2PA Content Credentials manifest, XMP AI flags like trainedAlgorithmicMedia, generator/tool tags naming the model, software encoder fingerprints like Lavc and x264, and no real GPS or capture timestamp. Platform scanners read those signals in seconds. Cleaning them — and injecting an authentic phone-capture identity — is what makes a file read as a normal recording. Results vary by platform and source model.

The metadata that gives an AI video away

Open an AI-generated clip with a forensic tool like ExifTool and the tells stack up fast:

C2PA / Content Credentials. A cryptographically signed provenance manifest stored inside the file as JUMBF atoms — effectively a receipt saying "made by AI." A single generated file can carry 18 of these atoms and 16 C2PA references.
XMP AI flags. DigitalSourceType: trainedAlgorithmicMedia is the loudest one, plus generator and tool tags that name the model that made the clip.
Encoder fingerprints. Software pipelines leave markers a phone camera never writes — Lavc (libavcodec) and an x264 SEI string are the classics.
Missing capture identity. A real phone recording carries Make, Model, Software version, GPS, and a capture timestamp. AI video usually has none of it, and that absence is itself a signal.

Put together, that's around 144 metadata tags on a generated file — a fingerprint an automated scanner reads instantly.

What a clean phone-capture file looks like instead

The goal isn't an empty file — a stripped-bare clip with no camera identity is still suspicious. The goal is a file that carries the same signals a genuine recording does. Calabi does this in one automatic pass:

Upload your AI video.
Automatic strip — removes the C2PA/Content Credentials manifest, the trainedAlgorithmicMedia XMP flag, generator/tool tags, and encoder fingerprints like Lavc and the x264 SEI.
Automatic inject — writes an authentic phone-capture identity: Make, Model, Software version, GPS, capture timestamp, and a real-phone encoder name. Profiles include iPhone 17 Pro Max, 16 Pro Max, 15 Pro, Pixel 8 Pro, and Galaxy S24 Ultra.
Forensic proof card + download — an ExifTool report shows exactly what was stripped and injected before you download.

In a verified run, 18 JUMBF/C2PA atoms went to 0, 16 C2PA references went to 0, the trainedAlgorithmicMedia flag was removed, and 144 metadata tags collapsed to about 94 neutral, phone-consistent tags.

The honest boundary

Metadata is the layer Calabi works on — invisible signals, not pixels. It does not change how your video looks, does not recut or re-edit footage, and does not erase a visible watermark from the frame (cropping does that; Calabi handles the invisible layer that survives the crop). It does not touch invisible pixel-level watermarks, and no honest tool guarantees a platform won't flag you. What it fixes is the file-level metadata that reveals an AI origin.

FAQ

What single metadata field most reveals AI video?

The C2PA Content Credentials manifest is the strongest tell — it's a signed provenance record explicitly marking AI origin. The trainedAlgorithmicMedia XMP flag runs a close second.

Can I just delete the metadata myself?

You can strip it, but a video with zero camera identity is still an outlier. The harder part is injecting a coherent phone-capture profile — matching Make, Model, encoder, GPS, and timestamp — so the file reads as genuine.

How do I check what metadata my video carries?

A forensic tool like ExifTool lists every tag. Calabi's proof card runs ExifTool for you and shows the before/after so you can see what was stripped and injected.

Try Calabi free at calabilabs.com — 10 cleans, no card.