> but it empirically performs the worst with a roughly 50% correlation to on-the-job performance. You’re better off flipping a coin because at least then you’re not biased.

I was going to point out that a correlation of 50% is pretty good (specially for predicting job performance from a single interview), whereas flipping a coin has 0% correlation with anything that is independent of the coin flip (such as job performance).

You probably meant to say the probability you rank job-performance of two random candidates correctly based on an interview is about 50% (what your source calls "percent concordant"), a correlation of 0%.

Out of curiosity: do you remember which section of the book looks at "chat about tech" interviews compared to other kinds of tech interviews in regards to their job performance prediction capabilities?