I'm curious - why does it sound like a fallacy to you? As far as I am aware, detection models are rarely looking at the underlying meaning, and most claim to scan sentence by sentence and calculate scores on that basis - which is what the note is about. in fact many humanizers rewrite the text so that even humans have difficulty making sense of it, and yet many detectors will give a 100% human result on those humanized texts.

about your note - it's a purely mathematical eq. if some text is 100% AI, how it is shuffled/sliced (get AI to slice/re-order sentences so that it still makes sense; if you want to control for human input) should not change the score significantly.

and - Yes ofcourse, to humans reading - it would make a dramatic difference. But when tools claim 'we give you a sentence level breakdown', it implies a degree of isolated analysis. So without breaking sentence units, it would seem fair to play around to check how much a score changes. For instance, I tried a 75 word piece on a popular detector, the last 25 words came back as AI and first 50 as assisted (roughly). When I clipped the 25 and put them at the front of the paragraph, they came back as human. Not sure if you get what I mean?

Word substitution does change meaning, but here's how I see it - if you change 2/100 words. but the score moves 25%, it sounds like a highly sensitive model, which could also be interpreted as a result based on weak signals. The point here is simply that gimmicks like percentages should be taken as gimmicks, unless you can find more robust ways to substantiate the scores.