The Topology of AI Alignment: How Human-AI Network Structure Shapes Collective Reliability

AI alignment is traditionally concerned with whether AI systems appropriately respond to human objectives, but how alignment shapes collective epistemic reliability remains underexplored. We introduce a network-level perspective: alignment has a topology, not merely a strength. Holding alignment strength and learning conditions fixed, we compare centralized and distributed human-AI alignment. Centralized alignment increases correlated-failure risk across every tested condition with nonzero alignment and advice reliance, yet can improve collective accuracy in some regimes. These results establish alignment topology as a system-level variable and reveal an accuracy-dependence trade-off in collective alignment.