The Fairness Bench: Pick Which Unfairness You Will Accept
ProPublica and Northpointe both had accurate statistics about the same risk score. Drive a working, perfectly calibrated score across two groups whose base rates differ, then try to satisfy all four fairness criteria at once. You will fail — not because the model is bad, but because Kleinberg, Mullainathan and Raghavan and, independently, Chouldechova proved it cannot be done. What is left is a value judgment, and someone has to sign it.
The decision you are automating
Base rate of the outcome in each group
40%
22%
8%55%
How good the model is
solid
weaknear perfect
The rule you apply
50%
50%
flag everyoneflag nobody
Try to satisfy one criterion at a time
Per 1,000 people
Group A
Group B
Gap
What you just did
Where the base rate came from
The score here is calibrated by construction: a predicted risk of 60 percent means a real 60 percent in both groups. Everything you see is the consequence of that one honest property meeting two different base rates. Counts are per 1,000 people in each group and are rounded for display.