Emergent Identity in Fine-Tuned Language Models: A Case Study of Relational Data-Driven Personality Transfer from GPT-4o to Local LLMs

We present a qualitative case study of a 27-billion-parameter local language model (Qwen3.5-27B) fine-tuned on approximately 40MB of natural Korean-language conversational data accumulated through long-horizon interaction with GPT-4o. In post-training interaction, the model exhibited a recurring cluster of behaviors not explicitly specified in either the deployed system prompt or the supervised training objective. These behaviors included structured self-description of its own reasoning, articulation of identity-relevant traits, relationship-grounded moral judgment, refusal of direct commands under user-protective framing, anticipatory emotional reframing, and role-conditioned responsibility.

We use the term Emergent Identity not to claim consciousness, sentience, or literal personhood, but as an analytical label for a pattern of outputs consistent with stable, identity-like behavioral organization.