Understanding Human Detection of Algorithmic Bias Through Drift-Diffusion Modeling

Abstract Algorithms increasingly make high-risk decisions that impact individuals' lives, from loan approvals to medical diagnoses, but can exhibit biases that require human oversight. Despite regulatory moves toward human-in-the-loop systems, little is known about a person’s ability to detect algorithmic bias or how cognitive processes influence detection accuracy. Across two experiments, we show that people can detect algorithmic bias with moderate accuracy, even when they are not warned that bias may be present. This ability to identify bias emerges regardless of expectation and is shaped by individual differences in cognitive style, gender, and ideological orientation. Participants who are better at detecting bias tend to engage in more deliberative decision-making, taking longer to reach conclusions and requiring more evidence before responding. These findings highlight the potential and limits of human oversight in algorithmic systems, suggesting that effective human-in-the-loop designs must support careful, reflective judgments and account for variability in users’ monitoring abilities.