<p>On-demand platforms, such as Didi and Uber, have increasingly adopted powerful AI-based algorithms to support platform operations and governance. A prominent example of algorithmic management is the implementation of automated punishment when platform workers violate rules. While such algorithm-based penalties may trigger algorithm aversion among workers, it remains unclear whether punishment-related factors contribute to this aversion in on-demand contexts. Grounded in self-determination theory and self-serving bias theory, this study empirically investigates the role of punishment severity as an antecedent of algorithm aversion and its subsequent impact on service performance. Furthermore, we examine the moderating effects of punishment controllability on the antecedent path and algorithm trust on the consequence path, respectively. The results indicate that: (1) punishment severity is positively associated with algorithm aversion; (2) algorithm aversion is negatively associated with service performance; (3) punishment controllability moderates the relationship between punishment severity and algorithm aversion and the indirect effect of punishment severity on service performance through algorithm aversion; (4) algorithm trust moderates the relationship between algorithm aversion and service performance; and (5) algorithm trust moderates the mediating effect of algorithm aversion between punishment severity and service performance. This study enhances the understanding of algorithm aversion in platform-mediated human–algorithm interactions and offers practical insights into how organizations can design punishment mechanisms to mitigate algorithm aversion and enhance service performance.</p>