He is suggesting that better methods of interviewing could increase “area under the curve”, improving both fase negative and false positive rates at once.
I understand that. But even if you come up with a classifier that gives better false positives and false negatives rates than the existing process, it's still going to be tuned towards minimizing false positives and leaving false negative to grow freely. My point was that only companies that do that became successful enough to attract a whole lot of applicants who fail interviews and complain about them on social media.