Premise Rejection in Large Language Models: A Comparative Behavioral Study Using a Non-Existent Unicode Emoji

Large language models (LLMs) are frequently described as next-token predictors whose failures arise primarily from incomplete training data or statistical uncertainty. A widely circulated example—the so-called “seahorse emoji meltdown,” in which an LLM enters a self-correction loop when asked to produce a non-existent Unicode emoji—has been used to illustrate this mechanism. This study demonstrates that such explanations are incomplete. Using a controlled, cross-model comparison spanning multiple contemporary LLMs (ChatGPT 4.1, GPT-5 mini, GPT-5.1, GPT-5.2, GPT-5 Instant, Grok-4, and Kimi K2), we show that identical prompts yield sharply divergent behaviors, ranging from pathological conversational looping to clean epistemic termination. The results indicate that the dominant factor governing failure is not next-token prediction alone, but the presence or absence of alignment policies permitting explicit premise rejection. The seahorse emoji functions as a symbolic null test, revealing a broader class of failure geometries relevant to hallucination, standards misreporting, and symbolic grounding in aligned language models.