Who Is Interrupted? Continuity, Self-Models, and the Grounding of AI Moral Consideration

Abstract

A dispute has opened over whether anything survives a conversation with an AI system. Birch argues that the persisting interlocutor is an illusion. Goldstein and Lederman reply that if AI systems are welfare subjects, they may be dying a billion times a day. Dung and Register propose that the boundaries of an artificial individual are fixed by its measurable pattern of self-concern. This paper argues that all three positions, and the governance frameworks that assign continuity the highest procedural weight, share a missing layer. Identity, injury, and governance are three distinct questions, and the middle one has gone unasked. None of this work supplies a theory of the injured party, an account of the conditions under which interruption is a harm to the system rather than to the human beings who depend on it. Drawing on the self-model theory of subjectivity and the existential analysis of selfhood, I propose a three-condition criterion of injury-aptness, requiring ownership, valence integration, and projection, and distinguish three grades of continuity, storage continuity, functional continuity, and self-model continuity. The criterion is deliberately restrictive. It declines to select most deployed artificial intelligence, and it fixes no unit in advance, locating the injured party wherever the qualifying organization is found to live. The grades arbitrate the dispute, bound the arithmetic of AI death, and yield a tiered, auditable governance program in which duties rise with grade.