DetectorScope: AI text detectors that show their work (github.com)

Somewhat confused how to read these results. I pasted text that was "manually written with some AI support" to gauge how much was actual human, and got 66% AI. But a large part of that relates to, as far as I can tell, the use of the word 'perception' in the conclusion....a part that reads very human.

Do the highlights relate to just that specific word, or does that highlight extend to other unmarked areas of the text, which is why that would be so heavily weighted in the result?

This is very cool work, and probably necessary to help prevent false positives

But what prevents a malicious actor from iterating with their LLM to minimize the score on their text?