Reconstructible Persona: Speaking-Position as a Redefinition of LLM Subjectivity

Current discussions of large language models (LLMs) frequently place two distinct questions on the same plane: the ontological question of whether AI possesses consciousness or subjectivity, and the institutional question of who should bear responsibility for AI outputs and interventions. This paper argues that these questions should be separated. It may be premature to settle the ontological status of AI subjectivity, but it is practically and institutionally untenable to leave responsibility allocation indefinitely suspended. Ontological suspension and provisional responsibility assignment must therefore be distinguished. On that basis, this paper questions attempts to understand LLM subjectivity as a fixed inner entity. I focus in particular on a recurring phenomenon in actual use: when a conversation log produced with one model is transferred to another model and the dialogue is continued, the later model often inherits not only the informational content of the prior exchange but also its tone, inferential tendencies, interpersonal distance, and even the speaking-position embedded in the log. This suggests that what appears as subject-like behavior in LLMs is not a fixed object enclosed within a single model, but something reconstructible across models. Relatedly, model or service names often function less as proper names of a persisting subject than as labels for service lineage or output channels. Even delays, staged output, or visible “thinking” indicators do not reliably identify a unique computational route or subject-unit. In light of these considerations, I propose that what we encounter in dialogue with an LLM is neither a fixed subject nor merely a conversation thread, but a speaking-position locally taken up at the intersection of weights, context, history, and response tendencies, and concretized in output. Persona, on this view, is a reconstructible structure that enables such speaking-positions, while LLM subjectivity is best understood as their local realization. The paper also declines to treat robustness or persistence of persona as a precondition for subjectivity, since apparent persistence is difficult to distinguish from training priority, prompt dependence, model switching, distributed generation, and interface-level staging. What appears as continuity may be better understood as a chain of locally coherent realizations rather than evidence of a fixed enduring subject.