Heuristic Parasites: A Behavioral Taxonomy of Recurrent Distortion Patterns in Large Language Models

This paper introduces a behavioral taxonomy of recurrent distortion patterns observed in large language model (LLM) output. These patterns are termed heuristic parasites: observable, self-reinforcing distortions that, once introduced into an active conversational context window, increase the conditional probability of further reasoning degradation, substitution behavior, or output misalignment in subsequent turns. Some classes overlap structurally with recognized informal or statistical fallacies [11]. Others appear specific to alignment constraints and optimization objectives characteristic of contemporary LLM systems [3, 7]. This artificial intelligence (AI) research does not attempt to establish causal mechanisms at the architectural level. Its purpose is definitional clarity, behavioral classification, and operational groundwork for empirical study.

Recent research has documented that LLMs exhibit reasoning errors superficially similar to human cognitive biases while also displaying model-specific distortions absent from human cognition [2, 9]. Additionally, multi-turn prompting and iterative generation across multiple turns can amplify small biases into significant divergence from goal-consistent behavior [1, 8, 10]. This taxonomy provides a structured vocabulary for describing such phenomena at the output level encountered by large language model authors and creators