Modelers are modern separatists. Absent data for behaviors they're interested in predicting, they substitute with proxies. They’re toy versions of the future that sacrifice accuracy for efficiency. Models are just opinions embedded in equations.
Think subprime mortgages and the securities pop-up of the '80s that led to the housing crisis of '08 that specifically targeted poor minority borrowers. Think banks undercutting payday lenders to give low credit borrowers a slightly less bankrupting rate on money. Think teachers evaluated on standardized scores disincentivizing learning and encouraging teacher cheating. People build models, and people suck.
Think about how it starts. US News has ranked colleges since the '80s. Their metrics are based on entrance scores, student-teacher ratios and acceptance rates. It's a glorified density ratio that's created a nasty feedback loop for colleges to focus improving what they're measured on, ironically quantifying them the same way they quantify their students. It gave rise to the administrator-saturated staff fueling today's amusement park campuses that push costs onto ballooned tuition that permanently indebt its students. It also gave rise to the college prep industry which, paradoxically, is only accessible to the wealthy.
The mind models on bias. Judges seek the death penalty at 4x the rate and 20% longer sentences for minorities than whites for identical crimes because we generalize individuals to skin color. The stop and frisk campaign that grew out of the bullshit Reagan-era Broken Windows Theory put more cops in poor neighborhoods resulting in more minority lock-ups for petty crimes resulted in lower access to work resulted in higher recidivism resulted in the intentionally disproportionate racial makeup of US prisons.
O'Neil, herself a mathematician turned quant turned activist, captures capitalism's fetish with predictive analytics at the cost of context.
Test scores go down when poorer, minority students are introduced to more affluent schools. However, break those scores down by income level and you can see score improvements rise for each bracket. Context is not a confounding variable, it's crucial.