proportional representation in artificial intelligence: clustering, alignment, and beyond

Proportional representation is a foundational concept in social choice theory, seeking to ensure that the preferences of distinct groups are reflected fairly in collective decisions. as algorithmic systems increasingly shape high-stakes decisions in society, there is a growing need for principled methods that enable artificial intelligence (ai) to account for heterogeneous human values and preferences. this article explores how proportional representation can be extended beyond its classical role in voting and elections to address key challenges in modern ai. we focus on two central domains. first, we examine clustering, when data points naturally represent individuals or agents with diverse characteristics or preferences. we review recent advances that reinterpret clustering as a representation problem, introduce formal notions as a representation problem for both centroid-based and non-centroid-based clustering, and highlight algorithmic guarantees ensuring that large, cohesive groups receive influence proportional to their size. second, we consider ai alignment, particularly reinforcement learning from human feedback (rlhf) in the presence of heterogeneous preferences. we argue that learning a single global reward function is fundamentally insufficient to capture population-level diversity and may violate basic social choice principles. to address this, we present a framework based on committees of reward functions, designed so that pairwise preferences induced by the committee proportionally reflect those of human annotators. we discuss theoretical guarantees showing that small committees suffice to achieve low proportionality error, as well as empirical evidence demonstrating substantial improvements over any single deterministic reward model.