Simulating Eutopia: Long-Term Fairness in AI Decision-Making

As AI-driven Decision Makers (ADMs) increasingly shape socioeconomic realities, their potential to both enhance efficiency and exacerbate social biases becomes critical. This research revisits the concept of long-term fairness in ADMs, particularly within the context of credit lending and its impact on wealth processes. Traditional approaches often overlook the performative nature of ADMs, where decisions can alter population behavior, and tend to focus on instantaneous prediction disparities rather than long-term equity. To address these limitations, the study formalizes the wealth dynamics induced by a loan-approving ADM as a performative Markov Decision Process, incorporating both ADM-level and social outcome-level reward functions. To facilitate testing, the researchers developed "Eutopia," a credit lending simulator equipped with a novel performative data generator. Experiments using Eutopia show that algorithms learning within performative dynamics, and those utilizing well-designed fairness-aware utilities evaluated on social outcomes, achieve superior long-term efficiency, equity, and inclusivity compared to classical methods.