Going back to whiteboarding: I don't like whiteboarding myself, but to be fair it does measure certain dimensions quickness-on-feet, memorization abilities, ability to exude presence, fluency in language, etc. While these are laudable abilities on their own, I agree they might not correlate with overall on-the-job performance (but it depends on the job).
I guess "job performance" is this amorphous latent variable y that is correlated with a bunch of direct predictors x which we can measure, u which we can't measure, so we use proxy variables z to stand-in for them: y = f(x, u, z).
The worry is that some candidates, who may be bad for the job, but just happen to be good at these proxy dimensions (or train for them) might get the job; on the flip side, we may exclude certain candidates who are potentially good for the job but perform badly at the proxy dimensions. Whiteboarding measures the proxy dimension z.
Edit: oh look, an article on HN's front page on this very issue:
My basic point here is that even 1% would be too much. If an interview process creates any inherent disadvantage for some groups, I'd say it deserves serious scrutiny. BTW, by "inherent" I mean beyond what can be addressed by bias training and such. That can enable interviewers to conduct any type of interview in as fair and kind a way as possible, but not to change the interview structure itself. If the structure is the problem, training isn't the solution. I think the in-person white-board algorithm interview is unavoidably weighted toward factors that have nothing to do with likely on-the-job performance, and thus should be avoided.