Many photos today are neither fully generated nor untouched but somewhere in between. When only part of a real photo was erased or filled with AI, the result is accurate only when read on two separate axes.
The AI generation axis asks whether the whole image was generated; the alteration axis asks whether a real photo was changed. A partly edited photo typically reads human or inconclusive on the generation axis and shows signs of editing on the alteration axis.
Some editing tools record AI edits as C2PA Content Credentials, metadata or a visible mark. When such a record exists, the ledger notes what was edited and when.
Edits that remove an object with the latest generative fill and leave no record were rarely caught in our own measurement. This limit is written on the landing page and in reports. A result with no edit signs does not guarantee the photo is untouched.
Getting both the captured original and the edited version, rather than a file re-saved by an editing app, lets you compare what changed. If editing matters, securing the original is the surest step.
This is general guidance, not legal advice. A scan result is a probabilistic assessment, not proof.