Comment by TimPC

It seems to me that recent changes have dramatically exacerbated this problem. Pangram is still claiming their extremely low false positive rate from before adding humanization detection and now very frequently detects actual human works as humanized AI.