New Paradigm Reframes AI Alignment as Preference Evolution Control.

Traditional approaches to AI alignment often assume human preferences are static targets for AI systems to infer and optimize. However, new research highlights that preferences are dynamic, layered, and actively shaped through interaction, especially with adaptive technologies like AI. As AI becomes more integrated into daily life, it increasingly influences what people value and endorse. The paper introduces "Constructive Alignment," a paradigm shift that redefines alignment as a control problem focused on managing the evolution of human preference trajectories, rather than merely satisfying static preferences. This framework models preferences as evolving state variables influenced by AI actions and interaction design. The core argument is that true alignment involves regulating how AI systems shape long-term human value formation, ensuring these values remain coherent, reflectively endorsed, epistemically grounded, and resistant to manipulation, while empowering users amidst uncertainty.