A bank tests its approval model across demographic groups. Overall accuracy is 91% for every group, but the false-negative rate (creditworthy applicants wrongly declined) is 8% for one group and 19% for another.
What should the team conclude?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Accuracy averages over both kinds of mistakes. Two groups can share 91% accuracy while one absorbs most of the costly errors — here, qualified people being turned away. Metrics sliced by group (false-negative rate, false-positive rate, calibration) are the minimum diagnostic; credit scoring is a high-risk use under the EU AI Act, and adverse-decision explanations are required under laws like Colorado's. Investigate the cause (proxy features, training-data gaps) before shipping.
Accuracy parity says nothing about which errors each group receives. The 19% vs 8% gap is the finding.
Zip code, employer, and shopping patterns can proxy for a protected attribute. Removing the column does not remove the disparity — measurement does.
A single global threshold shift changes both groups together and raises false positives. The fix requires understanding the cause; a blanket change is not analysis.