Human–Algorithmic Bias: Source, Evolution, and Impact.

This article investigates the sources, evolution, and impacts of human bias in loan approval decisions within a high-stakes microlending platform, focusing on two distinct types of bias: preference-based bias (inherent evaluator animus) and belief-based bias (subjective beliefs about group creditworthiness). Using a structural econometric model and a unique dataset featuring repeat loan applications, the study finds that human evaluators exhibit both biases favoring female applicants, with belief-based bias diminishing over time as evaluators learn from repayment behavior. Counterfactual simulations demonstrate that removing either bias improves fairness—measured by equal opportunity and demographic parity—and increases platform profits by raising approval rates for male borrowers who repay loans. Furthermore, machine learning (ML) algorithms trained on human decision data inherit these biases but generally reduce bias magnitude compared to humans; removing human biases from training data further enhances ML fairness, especially for new applicants. The findings suggest combining bias-reduced human judgment for new applicants with ML models for repeat applicants to optimize fairness and profitability in human-AI decision pipelines.