Research on algorithmic fairness has largely evolved along two complementary paths. In supervised learning, the emphasis has been on fair prediction, statistical notions of fairness, bias detection and mitigation, and rigorous evaluation of predictive models. In reinforcement learning, fairness arises in sequential decision making, where algorithms influence future states through exploration, interventions, and feedback loops, requiring consideration of long-term individual and societal outcomes. Although these communities address many of the same societal challenges, they have largely developed independently.

In this talk, I will argue that these two perspectives are fundamentally complementary. I will discuss the lessons each field offers the other across four dimensions: fairness notions (what it means for a system to be fair), algorithmic methods (how fairness can be achieved), evaluation (how fairness should be analyzed and measured in dynamic, adaptive, and interactive settings), and applications (where these ideas have the greatest impact). I will conclude by outlining a unified research agenda that bridges fair prediction and fair decision making, arguing that future progress in fair AI will depend on integrating insights from supervised learning and reinforcement learning, rather than advancing either in isolation.

Algorithmic decision-making systems in high-stakes domains raise fundamental questions about fairness, reliability, and the role of human judgment in shaping and evaluating these systems. In this talk, I discuss how humans enter the AI fairness pipeline at three distinct points - as model designers, as joint decision-makers, and as the baseline against which algorithms are judged.

First, I show how participatory input can help structure and navigate the space of model design decisions, enabling the representation of stakeholder values in the ML multiverse. Second, I introduce Conformal Multiverse Analysis (CMA), an auditing framework that integrates conformal prediction with multiverse analysis to assess how design decisions affect uncertainty and coverage disparities across groups in human-AI decision-making contexts. Third, I turn to the human baseline itself. Comparing human predictions to algorithmic predictions in a real-world case study illustrates that both can be biased in diverging ways, raising the question of how a fairer replacement should actually look like.

These perspectives jointly make the case for auditing not just models, but the system of interactions between humans and algorithms, and highlight the need for interdisciplinary work at the intersection of computer science, statistics, and social science.