Text has no EXIF and no signature. All that remains is the sentences, and sentences lose their traces when they are rewritten. AI text detection cannot offer the kind of certainty image analysis can, and not hiding that is the first rule of this feature.
Sentence-length variance, vocabulary diversity and the frequency of certain connectives show statistical tendencies. A tendency is not a verdict on one document. People write that way too, and formal reports and translated prose sit especially close to the tendency.
Short texts, heavily quoted texts, machine-translated texts and writing by non-native speakers all produce false positives in the same direction. It is also clear who is harmed. So in those conditions a rule prevents stylometry alone from producing an AI verdict.
A document file is a different matter. Editing history, authoring tool, revision session identifiers and whether the cited sources actually exist can all be checked independently of style. A document citing a source that does not exist is a strong clue in its own right.
We use no definite phrasing about writing, and when the only evidence is stylometric the result is inconclusive. This was not built to judge a student or a candidate on its own. The result is where a conversation starts, not where it ends.
This is general guidance, not legal advice. A scan result is a probabilistic assessment, not proof.