Diet: Interpreting Ethical Tradeoffs in Disrupted Sociotechnical Systems

Abstract

Sociotechnical systems comprise human stakeholders and autonomous agents, where stakeholders may prioritize values (such as freedom, security, and welfare) differently. Social norms operationalize these values by specifying expectations about acceptable behavior. Disruptions in sociotechnical systems create situations in which these stakeholder expectations cannot always be simultaneously satisfied, forcing autonomous agents to make ethically consequential decisions and sometimes deviate from otherwise desirable behavior. We investigate this problem in a multiagent ride-sharing domain in which autonomous vehicles respond to stochastic disruptions such as vehicle breakdowns. Our contributions are threefold: (1) we formulate disruption response as a decentralized multiagent reinforcement learning problem grounded in a transit setting; (2) we analyze reward functions aligned with utilitarian and Rawlsian ethical theories, alongside a profit-maximizing baseline, and demonstrate how ethical objectives shape disruption-response behavior and rider well-being; and (3) we introduce an analysis framework that interprets learned policies by extracting their implicit priorities over ethically relevant system criteria informed by social norms. We find that ethical reward design improves aggregate well-being, induces distinct disruption-response behaviors under different ethical reward formulations, and may concentrate disruption-handling load on ethically motivated agents in heterogeneous populations.