What this tool can and cannot tell you
It reads three things: an invisible watermark in the pixels, the IPTC digital source type in the metadata, and whether a C2PA container is attached. If any are present it reports them and what they say.
What it cannot do is prove a negative. An image with no marks may be an untouched camera original, or it may have been generated by a tool that writes nothing, or edited by one that does and then re-saved through a pipeline that stripped it. Those three cases look identical to any detector, and any tool claiming otherwise is overstating what it knows.
Why marks are layered
Metadata is easy to read and easy to lose. Almost every social platform, CMS and MLS pipeline re-encodes images on upload, and most drop everything except the pixels. A C2PA manifest is the strongest evidence when it survives, and it frequently does not.
That is why the marking has to be layered. An invisible watermark lives in the pixels themselves, so it survives re-encoding, resizing and moderate cropping — the exact transformations that destroy metadata. The metadata, in turn, is machine-readable by systems that will never run a watermark detector. Neither alone is sufficient.
The obligation this satisfies
EU AI Act Article 50(2) requires providers of generative systems to mark their outputs in a machine-readable format. The accompanying Code of Practice asks for at least two techniques and a publicly accessible way to check them. This page is that facility, and it is deliberately free and unauthenticated.