Constructive Alignment: Reframing AI Alignment as Value Co-Evolution

AAAI 2026 Workshop AIGOV Submission2 Authors

Keywords: AI Alignment

TL;DR: AI alignment shouldn’t lock in today’s values — it must continuously adapt to (and responsibly shape) how human values evolve over time.

Abstract: This position paper argues that AI alignment must be reframed as a dynamic process that evolves alongside human values—and inevitably influences them. Most approaches to AI alignment treat values as fixed, discoverable, and internally consistent, framing alignment as compliance with a predefined goal. But this view neglects a central empirical fact: human values are constructed through reflection, social interaction, and engagement with technology. As soon as we interact with AI systems, they become part of the value-formation loop. We introduce constructive alignment as a paradigm that treats alignment not as a one-time calibration to current preferences, but as an ongoing effort to support the ethical co-evolution of values over time. Drawing on evidence from behavioral economics, psychology, constructivism, and sociology, we model human values as trajectories shaped by interaction—and propose that aligned AI must predict, respond to, and respect this dynamic. We formalize this view through a value trajectory framework and propose concrete benchmarks to evaluate responsiveness, undue influence, and autonomy preservation. Ultimately, we argue that alignment should be understood not as static optimization, but as long-term moral stewardship: a continual effort to maintain ethical fit between AI systems and the evolving communities they serve.

Submission Number: 2