Teach & case
Help a near-peer reason through the idea — when you explain it, you understand it best — then read the real decision it comes from.
Help Nils check the model
What you've shown you can teach
A real case · you decide
Nils and the loan model that scores 95%
Nils — a student whose loan-screening model is 95% accurate for both groups but approves them at different rates
Nils built a model that decides who gets a small-business loan. It's 95% accurate for Group A and 95% accurate for Group B — equal accuracy, which sounds fair. But it approves 60% of Group A and only 40% of Group B. So is it fair? That turns out not to be a math question with one answer.
"Fair" isn't one thing. You could ask for equal SELECTION RATES (both groups approved at the same rate), or equal TRUE-POSITIVE rates (equally-qualified applicants approved equally often), or PREDICTIVE PARITY (a given score means the same real-world risk for everyone). Each is a reasonable meaning of fairness.
A well-known impossibility result says that when the base rates differ between groups, you generally CANNOT satisfy all of these at once — improving one fairness measure trades off against another. So "make it fair" isn't a computation the model hands you; it's a choice among competing values that a person has to make and defend.
Nils has to pick which fairness goal the loan model should serve. Which does the team aim for?
Nils's manager says "it's 95% accurate for both groups, so just call it fair and ship it." How should Nils handle that?
Nils's whole point is that "accuracy" and "fairness" are different questions. Which fairness you choose is a human decision with real trade-offs — the honest move is to name it, defend it, and show the numbers, not to let a high accuracy score quietly stand in for a value you never actually chose.