Against Value Alignment: Toward a Social Understanding of Values in AI

The term value alignment has become prevalent in research fields such as AI ethics and fair machine learning. However, in most cases, the meaning of value is left implicit or operationalized through computational proxies. In this article, we argue that such approaches assume an atomistic perspective on values borrowed from disciplines like behavioral economics and cognitive psychology. Following our analysis of a small corpus of value-alignment papers, we argue that this understanding construes values as individualistic and mechanical preferences, akin to utility functions, which can be computed and optimized for. In contrast, drawing on research in the qualitative social sciences (including anthropology, sociology, and the Deweyan pragmatist tradition), we reconceive values as social, dynamic, and situational. Discussing the implications for future work, we argue that the term value alignment should be rejected; alignment may be a desirable goal, but the question of what to align with must remain open and subject to public debate.