When Accountability Reduces or Increases AI Reliance in Public Decision-Making

• In Event: AI, Algorithms, and Decision-Making in Government

Abstract

Introduction
As AI enters frontline government work, public agencies face a basic design problem: how should accountability be structured when decisions are made jointly by humans and AI? A common expectation is that stricter accountability will make frontline officials less willing to rely on AI. We show that this relationship is conditional. Drawing on Blame Avoidance Theory, with Prospect Theory as a complementary lens, we treat AI reliance as an observable administrative behavior and examine how accountability consequences, responsibility allocation, and task analyzability jointly shape it.

Research Question
We ask two questions: How do accountability consequences (High vs. Low) and responsibility allocation (Individual vs. Joint Accountability) affect frontline civil servants’ AI reliance? And how are these effects conditioned by task analyzability, defined as the extent to which a task can be resolved through clear rules and established procedures rather than substantial personal judgment?

Data and Research Design
We use a pre-registered mixed-methods design. Semi-structured interviews with nine tax officials in Guangzhou identified a prevailing “responsibility-as-task” logic and informed the experimental manipulations. Based on these interviews, we developed two tax approval cases that differed in task analyzability. We then fielded a 2×2 between-subjects survey experiment through Credamo. From 722 civil servants initially recruited, 660 valid responses were retained after excluding failed attention checks, failed manipulation checks, and extremely short completion times. Participants reviewed two tax approval cases, and AI reliance was measured using the Weight of Advice (WOA), which captures movement from initial judgments toward AI recommendations. We analyze group differences using ANOVA and estimate conditional effects with regression models.

Findings
ANOVA and regression analyses reveal a strongly heterogeneous pattern. Accountability consequences, more than responsibility allocation alone, shape AI reliance, but their effects vary across levels of task analyzability. In high-analyzability tasks, where rules are clear, high accountability consequences significantly reduce AI reliance: accountability pressure appears to encourage defensive caution and greater independent judgment in response to possible algorithmic error. In this context, joint accountability partially relieves pressure and significantly interacts with accountability consequences. In low-analyzability tasks, however, situational risk outweighs the direct disciplining effect of organizational pressure. When personal judgment is more exposed and accountability consequences are high, frontline workers become more likely to rely on AI advice as a risk-management aid under uncertainty. Joint accountability shows little direct effect overall. Interview evidence suggests that when accountability consequences are weak, nominal shifts in responsibility allocation do not materially change frontline incentives, because officials still remain practically responsible for completing the case and absorbing citizen-facing burdens.

Conclusion/Implications
This study shows that accountability does not simply suppress or encourage AI reliance. Instead, it changes how frontline officials exercise prudence in AI-assisted decision-making. For policymakers, the implication is clear: accountability rules should be matched to task type. High accountability consequences may strengthen independent review in routine, highly analyzable tasks, but in less analyzable tasks they may unintentionally encourage excessive dependence on AI advice. Designing accountable human-AI decision-making therefore requires aligning accountability structures with the judgment demands of specific administrative tasks.

Author

• Cui Wan, The Hong Kong University of Science and Technology (Guangzhou)
  Presenting Author