Understanding Proxy Failure: An Agent-Based Model of Shortcut Adoption in Social Systems

In complex social systems, efforts to optimize proxies that supposedly reflect system goals can paradoxically lead to outcomes that undermine those very goals. This phenomenon is known as proxy failure. Proxy failure stems from the fact that certain shortcuts can enable the optimization of proxies without truly advancing the system's goals. Building on recent attempts to provide a unifying framework for proxy failures across domains, the current study investigates a central question: Why do shortcuts that undermine system goals repeatedly emerge and spread in complex social systems? Using an agent-based model with reinforcement learning agents, this study simulates 7 experiments to investigate the individual, social, and network-level factors that contribute to the discovery and diffusion of shortcuts. Results revealed that, (1) at the individual level, agents with weaker prior knowledge, higher exploration tendencies, and greater planning depth were more likely to discover shortcuts; (2) at the social level, frequency-biased learners tended to abandon shortcuts when surrounded by many neighbors, and the early success (or failure) of a pioneer agent (those who discovered shortcuts in early stages) had cascading effects on shortcut adoption across layers of agents; and (3) at the network level, shortcuts diffused more readily in dense and highly modular networks, while clustering showed no significant impact. Crucially, when hub nodes acted as pioneers, shortcut adoption was both faster and more persistent. These results together suggest that proxy failure is not solely a consequence of flawed proxy design, but also a product of the cognitive, social, and structural dynamics. The findings, along with the computational framework developed, offer a foundation for the design of more robust and resilient social systems against proxy failure.