Institutional Design and AI Delegation: Evidence from a Conjoint Experiment

• In Event: Artificial Intelligence and Governance

Abstract

As governments increasingly regulate high-stakes artificial intelligence (AI) such as autonomous driving, a central policy challenge is public willingness to delegate decision-making authority. While prior research examines institutional features in isolation, providing limited guidance for comprehensive governance frameworks, this study investigates how alternative institutional arrangements shape delegation preferences.

Building on principal–agent theory and Meaningful Human Control (MHC), we derive four institutional dimensions of AI governance and propose that they operate through three psychological mechanisms: perceived risk, algorithmic trust, and perceived controllability. To test these arguments, we implement a fully randomized conjoint experiment with 600 respondents in Wuhan, China—a pilot city for autonomous vehicles. We estimate Average Marginal Component Effects (AMCEs) and conduct mediation analysis using bootstrap inference.

The results highlight the importance of institutional design. First, human intervention mechanisms have the strongest positive effect on acceptance. Second, accountability arrangements exhibit an asymmetric pattern: compared to scenarios where users bear responsibility, assigning responsibility to manufacturers or operators substantially increases support, indicating strong aversion to personal liability. Third, government oversight is consistently preferred over industry supervision, whereas algorithmic explainability has a positive but statistically insignificant effect. Mediation analysis reveals that perceived risk, algorithmic trust, and perceived controllability serve as partial mediators, highlighting the cognitive pathways through which institutional design translates into delegation preferences.

These findings suggest that policymakers should prioritize maintaining human intervention capacity and avoid assigning responsibility to end-users, rather than solely relying on algorithmic transparency, to foster public trust and acceptance of algorithmic decision-making.

Authors

• Zirui Zhou, Beijing Normal University, School of Government
  Presenting Author

• Yue Guo, Beijing Normal University, School of Government
  Non-Presenting Co-Author