Differences between predicted PR for human-written texts (HW), ChatGPT-polished texts (CP) and ChatGPT-generated texts (CG): HW and CP are directly from the HPPT testset where CG are from the HC3 testset.