Calabi Labs · Guide · 2026-05-25
What metadata gives away an ai generated video
What Metadata Gives Away an AI-Generated Video
AI-generated videos leave traces — not always visible to the naked eye, but embedded in the file itself. Here's what forensic tools, platforms, and verification services look for.
1. C2PA Metadata (The Gold Standard)
The Coalition for Content Provenance and Authenticity (C2PA) is an open technical standard that embeds a cryptographically signed trail of metadata into photos, video, and audio. Think of it as a birth certificate for media — it records what created a file, when, and with which tool.
When a video is generated by Sora, Runway, Pika, or similar tools that support C2PA, the metadata chain typically includes:
- Generator software (name and version)
- Creation timestamp (precise, often in UTC)
- "AI-generated" disclosure flag
- Digital signature from the issuing organization
To verify, use the Content Credentials verifier at contentcredentials.org or the open-source C2PA tooling. A file with active Content Credentials will display a "CR" badge or a disclosure panel showing the tool that made it.
> Caveat: C2PA metadata can be stripped. Not all AI video tools support it yet, and re-encoding a video often removes the trail entirely.
3. XMP / IPTC Metadata Fields
The IPTC (International Press Telecommunications Council) published specific guidance for flagging AI-generated "synthetic media" in standard metadata. Fields to look for:
Iptc4xmpExt:DigitalSourceType— explicitly categorizes the media source (e.g., "trained AI model," "composite AI," "partial AI enhancement")photoshop:DateCreatedvs.xmp:CreateDate— inconsistencies between these fields are a red flag- Software/Creator tool strings — AI generation tools often write their name directly into the XMP block (e.g., "Runway Gen-3 Alpha" or "Pika 2.0")
If you open a video's metadata in a tool like ExifTool or MediaInfo and see software names that don't correspond to a physical camera or editing suite, that's a strong signal.
4. EXIF Anomalies in Video Containers
While EXIF is most associated with photos, video containers (MP4/MOV/WebM) carry their own metadata equivalents. AI-generated video often exhibits:
| Anomaly | What It Suggests |
|---|---|
| No camera hardware metadata | Video didn't come from a physical device |
| Timestamp in an unusual timezone with no offset field | Generated in a server environment |
| Frame rate that doesn't match standard camera presets | Created by a model (e.g., 12 fps oddball output) |
| Codec string referencing generation software | AI pipeline in the file header |
| Missing location (GPS) data when content claims real-world context | Potentially synthetic |
Legitimate camera footage consistently carries device-specific metadata. Synthetic or re-encoded video frequently drops or contradicts these fields.
5. Container-Level Software Signatures
Inside an MP4 or WebM file, the encoding software and pipeline are often named in technical headers. AI video tools write strings like:
MetaHuman,Gen-3,Stable Video Diffusion,Kling AI,Hailuo AI
These appear in codec or handler fields visible in hex editors or tools like MediaInfo. Finding one confirms the video was machine-generated.
6. What Strips or Hides These Traces
Forensic investigators also look for signs that metadata was removed — which itself is a signal:
- Metadata scrubbed — intentional stripping before upload
- Re-encoding artifacts — compressing a file can destroy C2PA signatures and watermark signals
- Platform re-processing — uploading to YouTube, Twitter/X, or TikTok re-encodes the video and removes most metadata flags
- "Clean" downloads — tools like Calabi remove AI flags so a file looks like it came from nowhere
If a video has had its metadata professionally cleaned, verification becomes harder — which is why combining metadata checks with visual/audio forensic analysis matters.
How to Check Yourself
- Content Credentials Verifier — upload at contentcredentials.org to read any embedded C2PA trail
- ExifTool — run in command line to dump all metadata fields from a video file
- MediaInfo — view technical metadata including codec and creation software strings
- Calabi — clean AI-generation metadata from files before sharing if you want to reset provenance
The short version: Look for C2PA Content Credentials first — they are the most authoritative trail. Then check XMP/IPTC fields for software signatures, EXIF/container metadata for anomalies, and be alert to signs of intentional stripping. No single flag is conclusive, but a cluster of them — especially a C2PA signature pointing to an AI tool — is a clear fingerprint.
Try Calabi free at calabilabs.com — 3 cleans, no card.