Selective Attention and Belief Formation in Public Administration: Experimental Evidence on Instability, Multimodality, and Human-AI Differences

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

Abstract

Behavioral public administration has documented pervasive deviations from rational decision-making among citizens, policymakers, and public employees. These patterns—over- and under-reaction to information, anchoring, and noisy judgment—have important implications for performance management, policy implementation, and service delivery. Yet the literature remains fragmented, with dozens of task-specific heuristics offering limited theoretical integration. This study proposes and experimentally tests a unifying framework, grounded in cognitive science, that attributes these anomalies to a single core mechanism: selective attention. Rather than stable distortions in the value function or distinct heuristics, belief instability and multimodality emerge because individuals and agents involuntarily attend to salient features of a problem while neglecting equally diagnostic ones. The framework generates two key predictions: (1) changes in framing or presentation alter which cues capture attention, producing systematic over- or under-reaction to equivalent data (instability); and (2) variation in attentional focus generates distinct clusters of responses around different anchoring points (multimodality). These predictions are tested in a series of pre-registered randomized experiments. A pilot study with approximately 2,000 subjects—comprising both public and private sector workers (primarily healthcare professionals)—randomly assigned participants to equivalent Bayesian problems that differ only in information framing. Participants reported, post-decision, which cues they attended to. The pilot explicitly compares patterns of instability and multimodality across public sector employees, private sector workers, and artificial agents. Findings from this pilot informed the design of a large-scale survey of more than 50,000 public workers currently underway. Results strongly support the selective-attention account. Self-reported attention to the prior, likelihood, or signal robustly predicts chosen anchors (logistic regressions: odds ratios 2.7–5.8, all p < .01). Distributional analyses confirm significant multimodality and framing-induced shifts. Comparative analyses reveal meaningful differences in instability and multimodality not only between humans and artificial agents, but also between public and private sector workers, highlighting how attentional mechanisms vary across organizational contexts and decision-maker types. By linking belief instability and multimodality to selective attention and contrasting patterns across public, private, and artificial agents, the study advances a parsimonious theory of cognitive bias and noise. These public-private and human-AI comparisons raise timely questions about the strengths and limitations of human, AI-only, and hybrid decision systems in the public sector, contributing to debates on algorithmic fairness, sector-specific moderators, and effective hybrid governance. The ongoing large-scale survey will further test generalizability across public-management settings. Overall, this research provides rigorous experimental evidence to the advancement of cognitive and behavioral public administration and lays foundational groundwork for understanding and improving decision processes in public and nonprofit organizations.

Author

• Nicola Belle, Sant'Anna School of Advanced Studies
  Presenting Author