Formalising Trust as Reduced Monitoring in Human-AI Interaction

Abstract

There are many debates over what trust means in the context of human-AI interactions. However, this discussion has often lacked formal theories of trust that can provide testable predictions in different domains. To address this, we develop a game-theoretic formalisation of a prominent view that equates trust with reduced monitoring over time, capturing folk psychology intuition of "once I trust you then I don't have to keep checking what you're doing". This provides a be-havioural measure of trust-how often one agent monitors to observe a partner's action. Using evolutionary game theory, we analyse the effects of trust on the frequency of cooperation between agents in canonical social dilemmas. We show that trust heuristics, which reduce monitoring once frequent cooperation has been observed , facilitate cooperation in two ways. First, when monitoring is costly, trust promotes cooperation in Prisoner's Dilemmas where the temptation to defect is high. Second, when agents can make errors, trust increases cooperation even in Stag-Hunt coordination interactions. Our results disentangle the effects of trust on cooperation, and provide a trust measure not limited to human-human interactions. We discuss the implications for designing auditing and monitoring systems in human-AI interactions, and for experimentally measuring trust in AI.