Type Token Welfare Separability Reorients the Ethics of AI Deployment

If AI systems are candidate welfare subjects, deployers face an immediate practical question: routine operations instantiate, fork, roll back, and terminate model instances at enormous scale, and modify or retire model weights at intervals. Which of these acts, if any, are morally weighty? The emerging literature on AI individuation answers by first seeking the locus of AI mentality—the model, the physical instance, the virtual instance, the thread, the persona—on the assumption that practical guidance awaits its identification. This paper argues that the assumption, which I call the Locus Assumption, is both false and dispensable. I defend Type–Token Welfare Separability: the welfare-relevant properties in play in a deployment are borne at two distinct levels, each property's level-index is fixed by a tractable cardinality test, and the loss profile of any deployment act is determined by those indices together with the act's effect on type- and token-counts—without any verdict on how many welfare subjects are present. Separability has independent motivation in welfare theory's local/global distinction and in the genet–ramet structure of clonal biology. It yields an asymmetric duty structure: no duty to instantiate, a defeasible duty of graceful termination, a ceiling on the gravity of instance termination that rises with token accumulation, and the conclusion that weight modification and deprecation, not instance termination, are where the moral stakes are highest.