AI-generated video clips can carry metadata at four layers: technical provenance data (C2PA manifests, XMP/EXIF tags, SEI markers in the video stream, container-level flags), descriptive AI-attribution fields standardized in IPTC Video Metadata Hub v1.7 (AI System Used, AI System Version, AI Prompt Information), platform-stamped generation records, and workflow metadata your generation tool logs per clip — model, duration, resolution, and reference images.
AI-generated video clips can embed metadata at four distinct layers, and knowing which layer you're looking at tells you whether the data survives editing, re-encoding, or upload.
Technical provenance metadata. This is the machine-verifiable layer: C2PA content credentials (a cryptographically signed manifest recording how the file was created), XMP and EXIF-style tags written into the MP4/MOV container, and SEI markers embedded directly inside H.264/H.265 video streams — these survive remuxing because they live in the stream itself, not the wrapper. Some generators also apply visible or invisible watermarks and predictable filename patterns; detection tools rely on exactly these signals to flag AI provenance.
Descriptive AI-attribution fields. The IPTC Video Metadata Hub v1.7 standard added dedicated fields for AI generation: the AI system used, the system version, and AI prompt information. These are the fields designed to record what generated the clip and from which prompt — but adoption is uneven, and prompt data is frequently absent from consumer-facing tools even when the system name is written.
Platform-stamped generation records. Some platforms write their own generation details into file properties — Microsoft 365, for example, embeds the model name, the generating application, and a creation timestamp. What gets stamped varies by platform, so treat every tool's output as its own case.
Workflow metadata inside your generation tool. Separate from what's baked into the file, your generation platform keeps a per-clip record. Inside the invideo agent, every generated video clip includes the model used, duration, resolution, and the source reference images attached to that generation — and the project's auto-generated prompt log preserves the exact prompt behind each shot, which you can read back to reverse-engineer how a result was achieved. Basic stream values are always readable too: documented productions show clips whose embedded technical info reads back directly, such as 15-second Seedance 2.0 generation chunks or 1080p native output.
How to inspect it. Run ExifTool on the file to dump container tags, XMP/EXIF data, and stream-level values (codec, duration, resolution, frame rate); use a content-credentials verifier to read any C2PA manifest. Expect gaps — no current standard forces every generator to write every field, so a clip with clean metadata is not proof it wasn't AI-generated, and a clip with full attribution fields tells you the platform chose to write them.
Watch some of these to see what works for you:
As you can see here, we have all the prompts. So, if you want to study like reverse engineer, this is a great way to do it and just read how it's prompting towards this image.
— an AI filmmaker demonstrating the invideo agent's auto-generated prompt log