Stake and Steerability: An Ontological Constraint on Artificial Agency and Subjectivity

Current evaluations of advanced AI agents are increasingly organized around functional indicators: memory, planning, global information access, self-modeling, multimodal perception, persistence, social reasoning, and increasingly autonomous action. This position paper argues that such indicators may be insufficient for claims about subjective agency because they omit a distinct variable: stake. We define stake as a mode of persistence in which a system’s continued existence is not externally guaranteed but continuously reproduced through its own endogenous activity. The central claim is conditional: if stake is necessary for thick subjective agency, then programmable, reproducible, losslessly steerable systems cannot establish such agency merely by satisfying functional checklists. This yields a practical counter-indicator: the easier it is to rewrite a system’s priorities, self-representation, or apparent commitments while preserving its baseline operation, the weaker the evidence that those commitments are constitutive of a subject rather than externally maintained functions. We develop the argument, distinguish thin phenomenality from thick subjectivity, and propose an evaluation agenda centered on intervention, persistence, history, and self-maintenance. The point is not that artificial subjectivity is impossible, but that stronger claims require evidence about how an agent exists over time, not only what it can do at a moment.