Most AI detectors still struggle with false positives, so I’m more interested in the actual methodology and benchmark data behind this one.