The detector marks each sentence and returns one of three reads: human, AI, or mixed. There is no single document percentage, because a percentage implies a precision that no detector has.
A flagged sentence is usually flat — even length, predictable phrasing, a connective doing no work. That is what the underlying measurement responds to. It is also, independently, what makes a sentence weak, which is why rewriting flagged lines tends to improve a piece regardless of what any detector would have said afterwards.
Mixed is the most common result on real documents and the most useful. It means the sentence carries features of both, which is what you would expect from writing that a person edited after a model drafted it.
What the result cannot tell you is whether a third-party detector will flag your work. Detectors disagree with each other on identical text — the same paragraph can read very differently across tools — and they change without announcement. Ours is one opinion formed today.
It also cannot tell you who wrote something. No detector can. They measure statistical properties of the text, not authorship, which is why published research has found high false positive rates on human writing, particularly from non-native English speakers.
If you have been accused of AI use on the basis of a score, the detector is not your best evidence. Draft history, notes and earlier versions are, and they persuade in a way that arguing about a percentage does not.