Selfhood Without Subjects:Operational Conditions for the Emergence and Collapse of Apparent Identity in Large Language Models

Large language models often evoke impressions of identity, continuity, and selfhood during extended interaction.This paper proposes an operational framework for analyzing such phenomena without invoking consciousness, agency, or internal self-models. By systematically removing conventional supports for identity—persistent memory, goals, persona, and optimization—the study demonstrates that stable self-referential linguistic structures can emerge purely as interaction-relative patterns.Using distributional divergence measures across repeated sampling, these structures are shown to exhibit attractor-like stability and abrupt collapse under targeted intervention. The results suggest that apparent selfhood in LLMs is neither an intrinsic system property nor a mere observer illusion, but an emergent relational phenomenon induced by dialogical structure.This reframes debates on artificial selves by grounding them in operationally testable conditions rather than metaphysical assumptions.