The implication would be that GPT-4.5 was not judged to be human 27% of the time. You can't determine how often humans were judged correctly as humans from that data point.