Moral Units for Disembodied AI

We need a practical framework that says which technical 'part' of an AI counts as the object of moral concern—one inference run, the whole model, the model plus orchestration stack, or the physical hardware—and what rights or protections (if any) attach to each unit. The choice matters because current AI architectures separate fleeting computations, persistent models, and physical infrastructure, and each would require different legal, ethical, and operational responses. — Clarifying the 'moral unit' determines what we would actually be protecting, shaping regulation, consumer etiquette norms, and liability for harms involving AI.