Structural Gender Bias in Credit Scoring: Proxy Leakage

Removing sensitive data isn't enough: Even when financial institutions remove explicit gender information from credit scoring models, gender bias persists through other variables like marital status, age, and credit limit—a phenomenon called "proxy leakage" that allows discrimination to continue invisibly.

Standard fairness checks miss hidden bias: The study found that gender could be reconstructed from non-sensitive financial features with 65% accuracy (ROC AUC of 0.65), revealing that traditional fairness audits fail to detect these embedded discriminatory pathways, meaning models can appear fair while still being biased.

"Fairness through blindness" is a flawed strategy: The common industry practice of simply excluding protected attributes doesn't work because structural bias is woven into the broader dataset—not just in one or two variables—requiring deeper investigation beyond surface-level statistical measures.

Specific features act as gender proxies: Variables like marital status, age, and credit limit function as powerful stand-ins for gender in credit models, meaning decisions that appear to be based on financial factors may actually be driven by gendered patterns in the data.

Structural accountability requires new approaches: The researchers advocate moving beyond simple statistical fairness toward "causal-aware modeling," suggesting that truly equitable AI in finance requires understanding why bias exists in the data structure itself, not just removing obvious indicators.