Evading the algorithm: increased propensity for tax evasion and norm violations in human-computer interactions

Abstract

Today's modern world is characterized by an increasing shift from human-to-human interaction toward human-computer-interaction (HCI). With the implementation of artificial agents as inspectors, as can be seen in today's airports, supermarkets, or, most recently, within the context of the COVID-19 pandemic, our everyday life is progressively shaped around interacting with automated agents. While our understanding of HCI is evolving, it is still in nascent stages. This is particularly true in the sphere of non-cooperative strategic interactions between humans and automated agents, which remains largely unexplored and calls for further investigation. A deeper understanding of the factors influencing strategic decision-making processes within HCI situations, and how perceptions of automated agents' capabilities might influence these decisions, is required. This gap is addressed by extending a non-cooperative inspection-game experiment with a tax-evasion frame, implementing automated agents as inspectors. Here, a within-subject design is used to investigate (1) how HCI differs from human-to-human interactions in this context and (2) how the complexity and perceived capabilities of automated agents affect human decision-making. The results indicate significant differences in decisions to evade taxes, with participants more likely to evade taxes when they are inspected by automated agents rather than by humans. These results may also be transferred to norm violations more generally, which may become more likely when participants are controlled by computers rather than by humans. Our results further show that participants were less likely to evade taxes when playing against an automated agent described as a complex AI, compared to an automated agent described as a simple algorithm, once they had experienced different agents.