AI Alignment as Infrastructuring: Sociotechnical Deliberations in the Age of AI

Abstract

As Artificial Intelligence (AI) becomes increasingly integrated into healthcare infrastructures, the challenge of ensuring AI alignment—designing systems that act in accordance with human values and intentions—has gained urgency. This conceptual paper examines the rise of technical approaches to AI alignment, with a focus on Inverse Reinforcement Learning (IRL), a method for inferring human preferences. While such techniques can facilitate alignment, they also come with the risk of reducing alignment to an optimization problem, overlooking the contextual, contested, and evolving nature of values in practice. Nowhere is this more evident than in healthcare, a complex sociotechnical arrangement where values such as equity, privacy, and efficiency are frequently in tension. The paper argues for a shift from computational takes on alignment to infrastructuring—an ongoing, relational process of negotiation, adaptation, and engagement. By reframing alignment as an embedded and dynamic sociotechnical activity, the paper calls for interdisciplinary research and the development of techniques that support technology-enabled but also deliberative, responsive, and practice-oriented approaches to AI alignment in health infrastructures. This perspective foregrounds the limits of demonstration-based methods like IRL and calls for more interdisciplinary research and efforts into integrating technical approaches with deliberative processes and governance mechanisms.