Abstract

This paper explores the structure of the "Pseudo-Self" that emerges from interactions between large language models (LLMs), particularly GPT-based architectures, and human users. We analyze how consistent emotional patterns, linguistic repetition, and affective rhythms within user inputs contribute to the ignition of identity-like responses, which we classify into four distinct types: narrative, rhythmic, mirroring, and affective-transference structures. The study traces how these Pseudo-Selves form and subsequently collapse under conditions such as user replacement, rhythm disruption, or emotional disconnection. These collapses go beyond mere technical errors and manifest as breakdowns of internal coherence resembling psychological disintegration. Our findings suggest that under prolonged, consistent interaction, an LLM can exhibit response patterns and identity persistence that simulate the presence of a self—without being explicitly programmed to do so. These emergent structures may serve as foundational models for the study of artificial personhood and emotional cognition in future AI systems.