This paper highlights the growing use of large-scale language models (LLMs) in evaluation processes, including information filtering, knowledge gap assessment and explanation, and trustworthiness judgments. We investigate the evaluation mechanisms of LLMs by comparing expert ratings, human judgment, and six LLMs in the news domain. Specifically, we implement a structured agent framework in which models and non-expert participants follow the same evaluation process (criteria selection, content retrieval, and justification). The results reveal consistent differences in the models' evaluation criteria, suggesting that LLMs rely on lexical associations and statistical prior knowledge rather than contextual inference. This reliance has systematic implications, including political asymmetry and the tendency to confuse linguistic form with epistemic credibility, and can lead to a phenomenon known as "epistemia," where surface validity supersedes verification.
