Effective Human Oversight of AI Systems

Abstract

Human oversight of artificial intelligence systems is increasingly mandated by regulation, yet empirical evidence on which interface design choices actually improve oversight quality remains scarce. This thesis investigates how two modifiable design factors, information signal granularity and intervention option range, affect the effectiveness and perceived workload of human overseers in an AI-assisted fraud detection task. A 3×3 between-subjects experiment was conducted online via Prolific (N = 144), in which participants reviewed 30 bank account applications flagged by a Gradient Boosting model under one of nine conditions, crossing three levels of information signal (plain case features, categorical risk level indicator, and continuous model confidence score) with three levels of intervention options (binary decision, decision with delegation, and decision with flagged delegation). Oversight effectiveness was operationalised as a composite score rewarding correct classifications and appropriate delegation decisions, and penalising both misclassifications and over-delegation. Perceived workload was measured using the NASA Task Load Index. Self-reported domain expertise was included as a covariate. Neither main effect reached the pre-registered Bonferroni-corrected significance threshold of α = 0.008, and no significant interaction was found on either dependent variable. A marginal effect of information signal granularity on oversight effectiveness was observed (F(2, 134) = 3.29, p = .040, η²p = .047), with a non-monotonic pattern in which the categorical risk level indicator outperformed both the plain information baseline and the continuous model confidence score. This reversal of the hypothesised ordering suggests that, for non-expert overseers, a well-designed categorical signal may be more actionable than a continuous probability score, as interpreting the latter requires complementary domain knowledge not uniformly present in a general population. The results are treated as preliminary, since the study was underpowered relative to the pre-registered target of N = 288, due to the planned expert-screened wave not being completed within the thesis timeline. Theoretical and practical implications for the design of human oversight interfaces are discussed, with particular attention to the relationship between signal granularity and overseer expertise.