Popular methods for making credit-scoring AI fair are actually breaking the model's ability to predict financial risk.
April 25, 2026
Original Paper
Algorithmic Debiasing: Is There Really a Low-Cost Accuracy-Fairness Trade-off?
SSRN · 6630505
The Takeaway
Low-cost fairness improvements in the finance industry are a statistical illusion. While the rank-order metrics look good, the actual calibration accuracy of the model collapses. This means banks may be using fair models that are fundamentally wrong about who will default. The trade-off between accuracy and fairness is much more severe than researchers previously claimed. This discovery forces a total rethink of how we build ethical algorithms for lending.
From the abstract
This paper challenges a central premise of algorithmic debiasing (AD) for consumer credit scoring models-that less discriminatory alternative (LDA) models can deliver materially improved fairness at minimal accuracy cost. Using a simplified logisticregression credit scoring model, I show that this conclusion depends critically on how model accuracy is measured. Specifically, when a relative or rank-order accuracy metric (such as the area under the ROC curve (AUC)) is used, an LDA Model may exhib