Comment by curiousllama

NAL, but have worked in this. Griggs is a bit more complicated than that, and its progeny modify application anyway.

The TLDR is that arbitrary tests are permissible if there's no disparate impact. Tests with disparate impact are permissible iff they are not arbitrary (i.e., "directly" assess job responsibilities).

So, for example, Leetcode may have disparate impact, but it's "direct" enough to be permissible. On the other hand, most "AI Assessments" are actually so badly implemented that they're effectively random - and a coin flip won't have disparate impact.