I bring this up to establish a standard of charity that I'm going to try to keep to in what follows.
I was catching up recently with an old friend going through something of an identity crisis. He's always thought of himself as a feminist: "the radical notion that women are people," check, all one happy motte here. Recently, though, he dated a third-wave feminist, and ran into a lot of cognitive dissonance, in that they were basically unable to talk about the kinds of problems Ozy Frantz brings up in this post. Ozy's opening paragraph is worth quoting in full, the sort of thing that ought to be cross-stitched into samplers and framed in bathrooms as a reminder that the map is not the territory:
Memo To The Social Justice Community At Large: the privilege/intersectionality model of how oppression works? Is a model. It’s an oversimplification that people use because the actual reality of how oppression works is way too complicated to talk about. It is not the Ultimate Truth Of How Oppression Works Forever and Ever.
There are lots of reasons why a model might not adequately explain real-world outcomes in its domain, but most of them boil down to "you took a shortcut somewhere." This usually happens because it's necessary. In machine pattern recognition, we grade the performance of systems on two metrics: precision, or "of all the times the system predicted outcome FOO, how often was it right?", and recall, or "of all the times the system should have predicted outcome FOO, how often did it?" Tuning a system that you plan to use for decision-making on real data often involves making tradeoffs between precision, recall, model size, and engineering constraints. For example, most of the bulk of any voice recognition application on your phone or tablet will be a few megabytes of precomputed matrices that the app uses to transliterate your speech sounds from your microphone into words on your screen -- that is, when your device can't reach the cloud servers where much more nuanced and accurate models that can run to the tens of gigabytes or more sit. Such is life in a world where device storage is finite. The hybrid system uses the remote, more accurate method when it can, and the local, does-the-best-it-can method when it has to.
It's hazardous to get too quantitative when comparing silicon to brains (the architecture's just so different), but I think the qualitative analogy -- that our mental models necessarily incorporate heuristic shortcuts -- should be pretty uncontroversial. It's which shortcuts people choose to take (or refuse to consider not taking), I think, that creates the sort of cognitive dissonance which my friend experienced on a micro level, and which on a macro level creates and perpetuates outgroups.
My brain is pretty meta sometimes, and occasionally goes about installing heuristics without warning me about it, which in combination is probably how I ended up with a heuristic aversion to people who cling to sloppy heuristics. (I am also kind of a jerk, though, don't forget.) This is going to be an interesting one to deal with, because look, sometimes sloppy heuristics are all somebody can afford. And, unfortunately, we're not silicon-based pattern-matching systems, so replacing a cheap-but-sloppy heuristic with a ballpark-as-cheap-but-less-sloppy one isn't as simple as a matter of pushing an app update. Just like my stress, the decision problems never go away; just like nicotine, the sloppy heuristic is a quick and cheap solution with long-tail side effects that we cheerily hyperbolically discount away, because the opportunity cost of switching to the better solution is higher than we're willing to pay up front.
Thus do we trap ourselves in coordination problems of our own making. So it goes.