Exploring Responsibility Attribution in Realistic Scenarios Involving Intelligent Agents

Abstract

As artificial intelligence (AI) systems get increasingly involved in decision-making across industries, debates have emerged regarding responsibility attribution in scenarios with AI being the moral actor. This study contributes to the discussion by examining the influence of social perception on moral judgments to scenarios involving AI and other non-human entities, referred to as “agent”. Specifically, we investigate the impact of narrative framing, the human-agent distinction, and general attitudes toward AI on moral judgments and trust in AI-based decision-making, while emphasizing the mediating roles of anthropomorphic attributions and cognitive conflicts. Based on Jaeger’s (2020) anthropomorphism theory, we differentiate between near (surface-level, direct) and far (deep-level, inferred) attributes to analyze their independent impacts. Participants reported general attitudes toward AI and baseline cognitive conflicts before being randomly assigned to a scenario involving an advisor from a healthcare platform providing incorrect medical advice, resulting in financial loss for a human user. The advisor was either an AI described in humanized narrative framing, an AI described in mechanical narrative framing, or a human advisor. Participants then assessed blame attribution, financial punishment, anthropomorphism attribution, trust in the advisor, and situational cognitive conflict. We initially conducted ANCOVA and mediation analyses collapsing the agent conditions (as no significant differences emerged), followed by an exploratory mediation analysis excluding the human condition and adding non-minimal far-attribution as a mediator. Results indicated insignificant effect of narrative framing alone and significant effect of human-agent distinction on responsibility attribution and trust. Human advisors were associated with lower blame attributes and higher trust than agent advisors. Positive attitudes towards AI predicted lower advisor blames, while negative attitudes predicted reduced trust only in primary analyses. Induced cognitive conflict directly increased advisor trust. Lastly, non-minimal far-attribution was associated with less platform blame and higher trust in agent advisors. These findings support established literature in algorithm aversion and highlight the roles of deep anthropomorphism attributes, general attitudes, and cognitive conflicts in shaping moral judgments and trust with agents, providing insights for reliable and trustworthy AI practices.