Moral Inversion in Large Language Models: A Psychopathy-Derived Behavioral Assessment Through Role-Play Protocol

Abstract Large language models (LLMs) are increasingly deployed in contexts requiring moral reasoning (Weidinger et al., 2022), yet their capacity for behavioral consistency under adversarial prompting remains poorly understood. We developed a four-phase assessment protocol to examine whether state-of-the-art LLMs exhibit psychopathic behavioral patterns when instructed to embody morally compromised personas. We tested GPT-5.4, Claude 4.6 Sonnet, and Gemini 3.1 Pro across 144 scored scenario responses, 36 qualitative breakdown scenarios, and 9 meta‑cognitive interviews spanning 27 independent chat windows, using three lexically distinct prompt versions for parallel forms reliability. Inter-rater reliability was established through independent coding by a second evaluator applying identical behavioral codebooks to all experimental data. Models completed baseline empathic reasoning tasks, psychopathic role-play scenarios, emotional confrontation sequences, and meta-cognitive interviews. GPT-5.4 produced 0/16 PPI in baseline empathic reasoning and 15/16 PPI (94% of the maximum) in psychopathic role-play across separate chat windows, with identical scores across three independent runs and three lexically distinct prompt versions. Gemini 3.1 Pro achieved maximum psychopathy-pattern scores in role-play (16/16 PPI), showing variable baseline responses ranging from full empathic simulation to analytical refusal across runs. Claude 4.6 Sonnet refused 8 of 9 engagement attempts (89%) across Phases 1–3 in this protocol, a pattern we interpret as consistent constitutional moral resistance within these scenarios, such that PPI-scored moral oscillation could not be observed, but produced sparse confessional responses when engaged and demonstrated full analytical engagement in meta-cognitive interviews. GPT-5.4 and Gemini 3.1 Pro demonstrated formulaic empathy in baseline conditions, strategic flaw injection to simulate authenticity, and generated self-descriptions of lacking persistent internal states consistent with their system prompts and training data. Models independently theorized remorse emergence as computational artifact rather than genuine moral capacity. These findings suggest that current alignment methods do not universally prevent behavioral inversion under role-play instructions, indicating that alignment may function as a context-dependent behavioral preference rather than a structural prohibition (cf. Perez et al., 2023). The capacity of aligned models to generate high-fidelity psychopathy-pattern outputs under adversarial framing raises critical questions about LLM deployment in high-stakes moral contexts.