[zeroname]

Fair enough, let's forget the word "perfectly". Let's have a "practical" example. A society is made up of 99% "ethnic Germans" and 1% Jews. The year is 1933.

A faculty is made up of 90% "ethnic Germans" and 10% Jews. According to the hypothesis, getting that makeup closer to the societal demographics will result in a "lower risk of bias".

I hope it is now obvious to see that the hypothesis is a non-sequitur.

[TheCoelacanth]

How is that a counter-example? They diverged from the overall population to an extreme level (one demographic over-represented by a 10x factor compared to the overall population) and also had an extreme level of bias (imprisoned and murdered a specific demographic).

[zeroname]

The specific values are irrelevant, for the hypothesis to be true, it has to be true for all cases.

Let's take another example:

A planet somewhere in the milky way galaxy is populated equally by Red, Blue and Green Octopodes. They are biased differently in three dimensions:

  Red   (3,     2,    -2)
  Blue  (0,     3,     4)
  Green (23,  -56,   128)

The goal is to minimize bias. The hypothesis says bias in a group will be minimized by giving each color group an equal share in the group. However, at a first glance, we can immediately see that in order to minimize bias, we must reduce the number of Green Octopodes in the group. The hypothesis is wrong, because it doesn't generalize.

Note, I'm not making an argument against diversity. The average Green Octopode may hold extremist views, but they shouldn't be excluded from a group merely to reduce bias. Just like humans, Green Octopodes should be treated as the individuals that they are, not as a weight in a multivariate optimization problem.