Well, humans might be great at detecting AI (few false negatives) but might falsely accuse humans more often (higher false positive rate). You might be among a set of humans being falsely accused a lot, but that's just proof that "heuristic stylometry" is consistent, it doesn't really say anything about the size of that set.