Dynamic Preferences in AI Alignment: A Deliberative Democracy Lens

AI alignment usually focuses on ensuring that the outputs of a model are “aligned” with peoples’ preferences, either by personalizing to or aggregating over different peoples’ preferences. However, these approaches typically overlook how a system's actions might change people's preferences over time. What do we do when preferences are a moving target? I will argue that existing proposals which focus solely on peoples’ preferences all encounter various pitfalls. On the other hand, the field of deliberative democracy, which has always viewed preferences as inherently dynamic, can provide valuable insights. Unlike aggregative forms of democracy, the objective of deliberative democracy is defined not only by peoples’ preferences but also by properties of the deliberative environment. I will discuss how this dual perspective—focusing on both preferences and the environment—can be applied to AI alignment and present two examples of our recent work inspired by this approach. The first and second examples connect aggregated and personalized preferences, respectively, to the deliberative environment. In the first example, we develop ranking methods that optimize for deliberative ideals while incorporating representation constraints from social choice. In the second example, we investigate the broader impact of formalizations of personalized preferences, based on either revealed or stated preferences, on the deliberative environment, ultimately leading to a line of work on enhancing user control of recommender systems. Overall, I hope to illustrate how a deliberative democracy lens can be a useful and fruitful one for AI alignment and hope that it can spark some interesting ideas for the community.