As AI-driven Decision Makers (ADMs) increasingly influence our socioeconomic reality, their roles in enhancing efficiency while amplifying social biases have garnered significant attention. This paper revisits the nuances of long-term 'fairness' achievable by an ADM, specifically in the context of a credit lending induced wealth process. The existing literature on long-term fairness primarily (a) considers passive environments where the predictor's outcomes do not alter the population's behavior, and (b) measures bias in terms of disparity in instantaneous predictions rather than downstream equity, which does not hold true for modern ADMs like credit lenders.
To address these issues, we first formalize the wealth dynamics induced by a loan approving ADM interacting with a multi-demographic population as a performative Markov Decision Process (MDP) with ADM-level and social outcome-level reward functions. We then mitigate the absence of such a performative test-bed by developing Eutopia: a lending-process simulator enabled with a novel performative data generator to learn long-term fair strategies. Finally, we test performative and classical reinforcement learning (RL) algorithms with different fairness-aware and utilitarian utilities.
Experimental results show that (a) learning with performative dynamics leads to better long-term efficiency and equity, and (b) learning with well-designed fairness-aware utility evaluated on social outcomes induces better efficiency, equity, and inclusivity.
Blogger's Review: This paper's exploration of long-term fairness provides not only new theoretical insights but also practical applications through the Eutopia simulator, showcasing the complexities and dynamics involved in AI decision processes. Such research helps us better understand and design future AI systems for greater social fairness.