The odd thing is that many of the supposedly 'random' new boards appear to be almost as segregated as the final "totally segregated" arrangements, because even randomness creates clusters.
Apparently the final results score higher on whatever metric they're using to calculate 'segregation' (which they don't define, as far as I can tell), but, really, the clusters frequently look almost the same.
(Which reminds us, of course, to pay attention to what actually happens in a simulation, rather than relying on the conclusions presented in the narrative.)
So if you create an incredibly simplistic model of dualistic 'preference', your (already-clustered) game board tends to align more into slightly larger clusters that better satisfy that preference?
What this tells us about racism, though, is nothing whatsoever.