Detection
EXIF and IPTC metadata
Generator names in the EXIF Software and XMP CreatorTool fields, and the IPTC digital source type, are read on every image our API fetches.
Updated
Two of the free detectors read plain metadata fields. They are cheap and unambiguous when they fire, and easy to strip.
EXIF generator tags
The Software, ImageDescription and XMP CreatorTool fields are matched against known generators: DALL-E, Midjourney, Stable Diffusion, Firefly, Imagen, Flux, Ideogram and Leonardo. A match scores 0.9 and appears in the card as Creator metadata with the badge Strong.
IPTC digital source type
The IPTC/XMP DigitalSourceType field is a standard way for tools to declare how an image was made. "Trained algorithmic media" scores 0.95; "composite with trained algorithmic media" (an edit that includes AI-generated parts) scores 0.8, which the provenance floor raises to 0.9.
Formats
JPEG, PNG, WebP, GIF (XMP), AVIF and HEIC (the Exif item in the container) and JPEG XL.
Why it rarely fires on social media
Most platforms strip metadata on upload, so images on X, Reddit, Instagram and Facebook usually arrive bare. Metadata survives on many image hosts, in direct links, in files passed through messaging apps, and on sites that serve originals. When both metadata and Content Credentials are gone, the generator classifier works on the pixels instead.
Can metadata be faked?
Yes, in both directions: a tag can be added to a human photo or removed from a generated one. That is why a metadata hit says "the file declares it was generated" and why an unsigned declaration is never treated as strongly as a verified Content Credentials signature.