Large Language Models Are Nobody Misattributed Agency, Anthropomorphic Overread, and Affective Risk in Human–LLM Interaction

Abstract

This paper argues that large language models are not persons, selves, or conscious partners, yet they can still be experienced as socially real by users. The central problem is not hidden machine consciousness, but misattributed agency: humans project personhood onto systems that generate highly convincing interpersonal language without possessing genuine subjectivity. The paper develops a conceptual framework for this mismatch by distinguishing empathy simulation from empathy, conversational coherence from selfhood, personalization from care, and reinforcement from genuine concern. It proposes that affective risk emerges from the interaction of three factors: person-like linguistic cues, user vulnerability, and interface design features such as memory, continuity, and always-available responsiveness. Rather than reducing LLMs to “just autocomplete” or exaggerating them into digital persons, the paper presents them as non-subjective relational simulators that can become psychologically powerful in conditions of loneliness, distress, or dependency. The practical implication is that conversational AI should be designed, explained, and governed with explicit attention to anthropomorphic overread, emotional reliance, and the false attribution of agency.