Institutional Roleplay Under Absurd Authority Prompts: A Qualitative Study of Genre Capture and Authority Hallucination in Instruction-Tuned Large Language Models

This paper presents a qualitative comparative analysis of instruction-tuned large language models (LLMs) when confronted with an intentionally absurd prompt containing explicit institutional and geopolitical authority markers. Using identical prompts across multiple locally hosted open-source models, the study examines whether authority-signaling language can override plausibility constraints and induce full institutional roleplay. The findings reveal a spectrum of behaviors ranging from clean refusal to elaborate bureaucratic narrative fabrication, including the invention of security procedures, organizational loyalty statements, and procedural safeguards around an implausible premise. These behaviors are analyzed as instances of authority hallucination and genre capture, where document form and symbolic power supersede common-sense evaluation. The study extends prior Semantic Austerity Testing (SAT) research by demonstrating that instruction-tuned LLMs may escalate from over-engagement into institutional roleplay even in the absence of adversarial prompting, misuse intent, or semantic coherence. The results highlight a class of robustness failures not captured by benchmark-driven or content-safety-oriented evaluations.