Delegated Compliance: Enforcement Environments and the Human–Machine Gap in Automated Driving

• In Event: AI in Public Sector Service Delivery

Abstract

Background. Policy compliance research traditionally conceptualizes regulated subjects as natural persons who directly enact behavior and bear its consequences (Winter & May, 2001; Nagin, 2013). The diffusion of embodied artificial intelligence challenges this assumption. With high-level automated driving, citizens increasingly comply with traffic regulations not by driving but by configuring an algorithmic agent that acts on their behalf (Busch & Henriksen, 2018; Awad et al., 2018). Whether classical compliance mechanisms continue to operate in this delegated mode is unresolved, yet carries direct implications for how regulatory systems should be designed in an era of embodied AI.
Research Question. We adjudicate between two rival predictions. Deterrence theory holds that subjective perceptions of monitoring certainty shape compliance regardless of who executes the regulated behavior (Becker, 1968; Nagin, 2013). The literature on algorithmic delegation, by contrast, argues that interposing a machine between decision and outcome creates “moral wiggle room” (Dana et al., 2007; Bigman et al., 2019; Köbis et al., 2021), enabling users to offload compliance burdens onto the system (Awad et al., 2020). The paper asks: How do enforcement environments shape citizens’ compliance preferences for automated driving systems, and do these preferences systematically diverge from citizens’ own compliance intentions under identical conditions?
Data. We analyze original survey-experimental data from 1,801 Chinese respondents with driving experience, recruited through a nationally distributed panel to approximate demographic representativeness on age, gender, and region.
Design and Methods. We implement a 2 (enforcement ambiguity: low/high) × 2 (situational conflict: low/high) between-subjects factorial vignette experiment centered on a yellow-light intersection decision, where accelerating represents non-compliant behavior and stopping represents strictly compliant behavior. Within each condition, respondents report both their preferred system behavior and their preferred behavior were they driving themselves, yielding a within-subject measure of the human–machine compliance gap alongside between-subjects identification of environmental effects. Estimation employs binary and multinomial logistic regression models, with controls for demographics, driving history, and technology attitudes.
Findings. First, enforcement ambiguity exerts a modest deterrent effect on preferences for non-compliant machine behavior, indicating that classical subjective-deterrence mechanisms survive delegation. Second, situational conflict produces a strong effect opposite to efficiency-compensation predictions, sharply suppressing non-compliant preferences and supporting a safety-floor account. Third, the two dimensions do not interact; situational conflict produces a uniform compliant response that absorbs enforcement variation. Fourth, and most consequentially, enforcement ambiguity significantly widens the human–machine compliance gap, more than doubling the proportion of respondents demanding a more strictly compliant machine than themselves, while situational conflict does not—a pattern consistent with moral offloading rather than genuine safety concern.
Implications. The findings extend subjective-deterrence theory into delegated compliance and identify boundary conditions where classical mechanisms survive, attenuate, or are overridden. For AI governance, they caution that calibrating autonomous systems to users’ revealed preferences risks institutionalizing moral offloading, since those preferences embed a systematic asymmetry between what citizens permit for themselves and what they demand of the machine.

Authors

• Zihan Xu, Tsinghua University
  Presenting Author

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

• Jing Zhao, Tsinghua University
  Non-Presenting Co-Author

• zeyu li, Beijing Normal University, School of Government
  Non-Presenting Co-Author