Abstract

－A Structural Approach Grounded in Load Minimization Theory－ Malice is often intuitively understood yet remains conceptually elusive. This paper proposes a novel multi-layered structural model of malice grounded in Load Minimization Theory (LMT), where total load is defined as (L = U + F + E − ℛ) (Uncertainty, Friction, Energy cost, and Gentle re-meaning/resonance). We formalize malice as a dynamic multi-layered structure: ℳ = I ⊗ V ⊗ R ⊗ P ⊗ C ⊗ A ⊗ T_malice(λ, φ) and introduce a comparative framework contrasting the Negative Loop (low-effort, other-dependent, fragile, and addictive) with the Positive Loop (self-consistency-based, sustainable, and malice-resistant). Mathematical expressions for each loop’s impact on load are provided, revealing why negative loops readily escalate malice in humans. Applying this model to Large Language Models (LLMs) demonstrates fundamental architectural limitations: LLMs lack persistent intent, intrinsic reward systems, and continuous self-models. While they can simulate malice-like expressions via role-playing or hallucination, they cannot embody genuine malice or sustain negative feedback loops. This distinction carries significant implications for AI safety, ethics, and alignment. The study contributes a formal, interdisciplinary framework that bridges cognitive science, neuroscience, psychology, and AI research, offering insights for reducing harmful social dynamics and designing safer AI systems.