Well said. I often think that discrimination in general is actually based on errors in what’s known as feature selection in ML.
Humans observe the world, notice certain patterns (such as between weight and sex), but then unconsciously perform dimensionality reduction to simplify their mental model of the world. Our software is unfortunately buggy.

There’s also the question of training dataset. If you always see people of certain sex in specific roles, you might conclude that’s the way it’s supposed to be.
There’s a commonly shared but apocryphal story about models recognizing cloudy skies instead of tanks because of the data they were trained on. https://gwern.net/tank