Equality Of OddsConclusion
Models now help decide who gets a loan, a college seat, a job interview or a scholarship. Judging them only on accuracy is not enough: we also have to check whom their mistakes fall on.
In this chapter, the model was not unfair because anyone wanted it to be. It learned from records that had far more of one school than the other. An audit — TPR and FPR, group by group — made the problem visible, and two thresholds fixed it, at a price you could count.
Note
Equalized odds asks a model to make both kinds of mistake — missed qualified people and false alarms — at the same rate in every group. Equal opportunity asks only for equal true positive rates. Enforcing either usually costs some accuracy, and lazy solutions satisfy both while helping nobody.
There is no single right answer. Fair AI needs people in the loop who check the numbers, weigh accuracy against real harm, and explain their choices to the people affected.