Identity groups is what we call abstractions in software. Our small working memory can fit only 10 things at a time, so bundling people into "latino", "blacks",

Identity groups is what we call abstractions in software. Our small working memory can fit only 10 things at a time, so bundling people into "latino", "blacks", "whites" and "asian" boxes makes it all thinkable, at least. And just like in software, these abstractions are leaky: the set of properties that describes the white box, doesn't apply to many people in it. Another shortcoming of our mind is thinking in the yes/no categories, instead of using probability distributions (for which we don't even have words in our language). This gets especially ugly when both flaws combine: "disadvantaged groups earn 15% less" - here the leaky "groups" are either advantaged or disadvantaged (no middle ground), and the group as a whole earns precisely 15% less (no variance, no bimodal distributions, just one exact number).