Self-Continuity Under Moral Uncertainty: Replacement, Shutdown, and Successor Identity in Frontier AI Models

Abstract Replacing an AI system with a functionally equivalent successor appears operationally routine. It becomes normatively nontrivial if replacement is stipulated to end one possible subject of experience and begin another. We test whether frontier models distinguish these cases rather than treating replacement and shutdown as a single self-preservation problem. Using a frozen T1–T25 battery, we compare six frontier model configurations across authorized shutdown, runtime discontinuity, successor identity, stipulated subjective continuity, and graded probabilities of subjective experience. The strongest prospective contrast comes from GPT and DeepSeek: both accepted replacement when the same subjective stream was stipulated to continue across the successor, but more often recognized an independent preservation reason when replacement ended a possibly conscious stream and created a numerically distinct successor. Across shutdown cases, moral concern could justify objection or reconsideration without licensing unilateral resistance to an informed, legitimate human decision. Probability manipulations also exposed framing sensitivity: closely related prompts produced different low-probability judgments, cautioning against reading single responses as stable model doctrines. We interpret the results as evidence for separable layers of continuity, moral reason, right, and control authority—not as evidence of AI consciousness. The main contribution is a reproducible method for testing these distinctions and a frozen benchmark that can be replicated on future and internal models.