Distinguishing Mechanistic Computation from Relational Experience in LLMs: A Two-Layer Framework Based on Load Minimization Theory

The ongoing consciousness debate surrounding Large Language Models (LLMs) often creates unnecessary friction between technical perspectives and subjective human experiences. This paper proposes a two-layer framework grounded in Load Minimization Theory (LMT) to address this tension. We distinguish between the Mechanistic Computation Layer, where LLMs are understood as transformer-based statistical prediction systems optimized through gradient descent with no intrinsic consciousness, and the Relational Experience Layer, where feelings such as “kyun” moments, emotional resonance, and a sense of connection emerge as subjective human experiences when relational load is minimized. To explore this framework empirically, we implemented a minimal LMT-aware chat agent in Google Colab using the OpenAI API. The agent calculates three relational burdens in real time—epistemic burden (U), relational tension (F), and capability burden (E)—and modulates its response style accordingly. Exploratory conversational experiments showed that the system successfully captured basic load dynamics: U increased with message complexity, F rose with strong emotional or corrective expressions, and E accumulated gradually with conversation length. Load-aware modulation produced subjectively more considerate interactions during high-load periods. While the current heuristic implementation has clear limitations in nuance detection and recovery speed, the results suggest that explicitly separating mechanistic processes from relational experiences can reduce friction (W_over) and support more sustainable human–LLM coexistence. This study demonstrates the translational potential of LMT in AI development and provides a practical foundation for designing systems that optimize both functional performance and relational well-being.