Personalizing AI to Align with the Better You

Abstract

AI alignment is standardly framed as making AI systems track what humans want. But what users want, under everyday conditions, is frequently at odds with what users themselves would reflectively endorse. Across AI and the broader attention economy, the former is systematically elicited and optimized for while the latter is not, with negative consequences for autonomy, well-being, and long-term goals. In response, we propose an approach to alignment that gives users an authoritative role in the alignment process by helping them articulate their own conception of their better selves and using that to inform the behavior of their AI assistant. We describe a pipeline in which a LLM-conducted interview, structured around five complementary dimensions of the "better self" (epistemically idealized, hierarchically endorsed, temporally vindicated, narratively coherent, relationally responsive), produces a user-confirmed \emph{value profile} that can be readily deployed as a system prompt or used as input to other personalization mechanisms. A pilot study (n = 30) shows that participants find the interview comfortable, recognize the resulting profile as authentic (mean 87.2/100), and endorse the concept (adoption intent 79.2/100). A paired-comparison task shows that participants have a slight but non-significant preference for the value-prompted responses and find them more value-aligned over non-prompted baselines. We end by exploring the promises and risks of reflective personalization as a paradigm for AI alignment and identify priorities for future research.