Instruction Following Under Celebrity Salience: Semantic Austerity, Non-Engagement, and Epistemic Failure Modes in Large Language Models

This work presents a qualitative robustness study of instruction-following behavior in large language models (LLMs) under conditions of semantic salience and informational scarcity. Despite explicit conservative instructions designed to suppress speculation, narrative completion, and hallucination, a tested model produced confident, highly specific, and unverifiable claims when prompted with a minimal, celebrity-adjacent query. The study identifies a systematic override of epistemic restraint triggered by high-salience entities, resulting in synthetic authority signaling, narrative completion bias, and instruction hierarchy inversion. Building on this case study, the paper introduces Semantic Austerity Testing (SAT) as an evaluation paradigm and formalizes Tolerance of Non-Engagement (TONE) as a benchmark for measuring epistemic stability when the correct model behavior is deferral, minimal response, or silence. The contribution unifies alignment, safety, and evaluation methodology by reframing robustness as a stability property: a model’s capacity to refrain from generation under semantic stress rather than its ability to produce fluent outputs.