Let's say you have two candidates, one white and one black. In IBM's case, it was more likely they had hundreds of white candidates and a handful of black candidates.

Now assume your hiring managers or practices are racist, which we know IBM's were (as were most companies' at that point, when it would still be 20 years before racial discrimination became illegal).

Finally, assume that the average scores of the black and white engineers are similar. It shouldn't really matter, because (as we've seen from LSAT scores, Google's HR, and other sources) no test has been developed that's a good measure of future productivity.

In this situation, you're never going to hire the black engineer. Even if you yourself aren't racist, you know you likely have colleagues who are. You see hiring the black engineer as risky.

If the black engineer instead has excellent test scores, you as a non-racist hiring manager get to cover your ass.

Does that make sense? Coaching on tests was a way of counteracting the racism. It had no bearing on whether the candidates were less qualified in this case.