ORC Seminar: Internal Pluralism and the Limits of Pairwise Comparisons

Bailey Flanigan Assistant Professor MIT Abstract: Across AI alignment and participatory design, pairwise comparisons are a workhorse for learning how people want automated decision rules to behave. We propose a preference model that contains standard pairwise-choice models as a special case, but which allows an individual to evaluate a decision rule according to multiple distinct priorities, and to experience irreducible conflict between them. This added generality exposes two ways in which standard learning pipelines can fail. First, some natural priorities—such as proportionality, egalitaria