I think what the author sneakily avoids is that what is being withheld is socio-behavioral measurements made along with the genetic samples. And the reason is that people who didn't actually gather data themselves are ignorant of the shortcomings of the measurements. This includes both "intelligence" testing and measuring/assigning "race". As the parent of a phenotypically "white" child and "black" child (who share both parents, to be clear) this later point is really relevant. It's possible that people can study these things carefully, but an obvious precursor is understanding that if someone doubts their dataset is appropriate for something, they probably have good reasons. Bias in deep learning models is a classic case of this....