AI Homogenization in Decision-Making and Alignment

ABSTRACT

As AI systems become more pervasive, their outputs in both decision-making and generative tasks often lack the diversity expected or desired. This thesis advances our understanding of AI homogenization by evaluating several distinct forms and proposing practical mitigation strategies. Part I studies outcome homogenization, or when certain individuals consistently end up on the losing side of AI decisions. I propose and evaluate two strategies to reduce outcome homogenization: model multiplicity and randomization. Part II examines homogenization in generative AI systems through the lens of three forms of misalignment: norm inconsistency, sycophancy, and mode collapse. These cases illuminate if and when homogenization is undesirable and motivate task-dependent strategies for mitigation. Overall, this thesis empirically challenges the assumption that promoting diversity trades off with utility, informing ongoing debates in AI alignment, system design, and safety.

COMMITTEE

Ashia Wilson (advisor), Manish Raghavan, Dana Calacci, Kathleen Creel

EVENT INFORMATION

Hybrid event. To attend virtually, please contact the IDSS Academic Office (idss_academic_office@mit.edu) for connection information.