Did you read the paper you are linking to? It records a 0% false positive rate for evaluation of human-authored controls, and fewer than 5% of the hybrid and humanized papers had their AI levels overestimated by pangram. It is not completely clear from the data, but it seems that depending on whether the n=2 overestimates were “100% ai generated” assessments, the study you link says pangram’s 100% AI assessments were right anywhere from 95-100%