I was being overly simplistic - I meant more that the patterns it's trying to detect were created by an LLM trained on the data they're inputting.

It's like how reddit comments from 2016 look generated. If you stick one of those into an AI detector, it'll give a false positive for the same reason