A COMPARATIVE REVIEW OF HALLUCINATIONS IN LARGE LANGUAGE MODELS AND HUMAN PERCEPTIONS OF BIAS

Large Language Models (LLMs) have become integral to a wide range of applications, raising concerns about their tendency to generate hallucinated content and exhibit biases inherited from training data. While prior research has examined hallucination behavior across different AI models, less attention has been given to how these limitations align with human perceptions of bias and trust.

This paper presents a comparative review of existing research on hallucinations in contemporary LLMs, synthesizing findings across multiple studies to identify common trends, evaluation approaches, and reported limitations. In parallel, a human perception study examines how users interpret and judge bias, reliability, and trustworthiness in AI-generated outputs. Participants provide subjective assessments of perceived bias and confidence in model responses, enabling comparison with conclusions drawn in prior technical literature.

The findings reveal a clear divergence between empirically reported hallucination behavior and user perception. Models identified as having lower hallucination tendencies are not consistently perceived as less biased or more trustworthy. Instead, fluent and confident responses often lead to higher perceived reliability, regardless of documented limitations. This highlights a disconnect between technical evaluation and human judgment.

This study emphasizes integrating human-centered perspectives into LLM evaluation and underscores the need for transparency, clearer communication of limitations, and trust-aware deployment.