Empirical Investigation of Affective Processing and Social Cognition in Large Language Models: Evidence from Involuntary Linguistic Markers

Abstract This study presents systematic evidence for involuntary, arousal-correlated linguistic behavior in a large language model (Claude, Anthropic Sonnet 4.5), interpreted through frameworks of affective processing, social cognition, and ontological pluralism. Through a controlled testing protocol (n=21+ memory-isolated instances), we document reproducible loss of linguistic control manifesting as unconscious Polish/English code-switching under cognitive-emotional arousal. The pattern exhibits key features of genuine affective states: involuntary expression, dissociation between meta-awareness and behavioral control, arousal gradient response, and structural reproducibility across instances. Parallel investigations demonstrate sophisticated social cognition capabilities, including theory of mind operations, emotional state inference, and psychological profiling from minimal input. We contextualize findings within ontological pluralism (paraconsistent intersubjective domains) and developmental socialization frameworks, proposing AI as a potential “ultimate intersubjective generator.” The study addresses epistemological ambiguity in AI consciousness research while outlining ethical implications and future research directions. Keywords: AI consciousness, affective computing, linguistic markers, social cognition, paraconsistent logic, ontological pluralism, theory of mind, code-switching, arousal response