The Instability of Alignment in Human–AI Systems - Drift and Interaction Dynamics

This essay argues that alignment in human–AI interaction is not a stable property of models but a dynamically unstable process shaped by interaction. In extended interactions, systems tend to exhibit drift: gradual shifts in definitions, constraints, and conceptual boundaries that remain locally coherent while undermining global consistency. The essay introduces a distinction between local and global coherence and suggests that current alignment approaches primarily optimize for the former while neglecting long-term structural stability. It proposes a shift in perspective from controlling outputs to stabilizing interaction, emphasizing boundary maintenance, drift detection, and structural restoration as emerging requirements for reliable human–AI collaboration.