Algorithmic Bias and Fairness in AI Credit Scoring: Evidence, Mechanisms, and Governance Responses

Abstract

Machine learning models have become widely adopted in credit scoring, but their deployment raises persistent concerns about discriminatory outcomes for protected demographic groups. This paper examines algorithmic bias and fairness in AI credit scoring, drawing on evidence from 15 empirical studies drawn from a systematically screened corpus of 55 peer-reviewed papers published between 2018 and 2025. The evidence shows that bias in AI credit scoring models is not primarily a technical failure, it originates from historically generated training data, is amplified through correlated proxy variables, and persists in part because institutional incentive structures do not penalise discriminatory outcomes absent regulatory compulsion. Gender and socioeconomic disparities are the most empirically documented, with digital credit markets in Kenya showing higher male participation rates and profit-optimised models disproportionately excluding marginal applicants. Technical mitigation strategies, causal inference corrections, SHAP-based feature audits, and discriminatory feature removal, show conditional effectiveness, but none resolves the structural misalignment between commercial model objectives and equitable lending. The paper argues that fairness governance requires both technical standards and institutional accountability mechanisms.

JEL Codes: G21, G28, C45, J71, K23