It's a bit of a mystery, truly, and bias can mean different things (e.g., slope vs. intercept predictive bias). Note also that predictive bias is not the same thing as mean subgroup differences (e.g., mean score differences of White vs. Black candidates).

This is a quick overview of the topic: http://www.siop.org/_Principles/pages31to34.pdf

An interesting thing is that the predictive bias is reduced for open-form response questions vs. select-one tests. It's an indication that there is more at work than just subgroup differences.

When my mother was doing sociology at Uni I read one of her texts (as you do) and it had an example of an IQ test been flawed, they gave the same test to two different groups of children and found that the poorer (working class as they where grouped in the study) children performed less well consistently.

One of the people looking at the results then looked at the breakdown of questions by group and noticed immediately that questions like "The cup goes on the a) saucer b) floor c) table d) shelf" where consistently "wrong" (correct answer was a) for the poorer group at which point he realised that working class children drink tea from mugs and saucers where a middle class thing.

The story might be apocryphal but its stuck with me since I was 12-13 whenever I run into any kind of standardised testing/results.

The latter. The races perform differently on "fair" cognitive tests.

How do we know the tests are fair? For a given test score, life outcomes like income and criminal conviction rate are the same.