The flaw in trying to detect AI by its use of particular idioms is that it would have learned these idioms from its training corpus, which consists of writings from actual human beings.
In other words, some people actually write like this.
In other words, some people actually write like this.
I have a friend that has used ems all his professional life and is livid that they're now a telltale for AI. So yeah, false positives.
Heck, anyone used to a word processor that automatically changes dash dash into em-dash.
There’s a lot of us that knew how to use em-dash.
The valley girl using "like" every other word, for example?
Or I had a colleague who would use the expression "we can say" (in French, because we were speaking in French) basically every couple sentences for a bit.
Humans also repeat speech/linguistic patterns, therefore "repetition of the same pattern" is not sufficient to mark text as produced by AI :)