As large-scale language models (LLMs) are increasingly integrated into workflows that involve evaluation processes, this paper addresses the need to investigate how these evaluations are constructed, what assumptions they rely on, and how they differ from human strategies. The study benchmarks six LLMs against NewsGuard and Media Bias/Fact Check (MBFC) expert evaluations and human judgments collected through controlled experiments. It implements a structured, goal-oriented framework in which both models and non-expert participants follow the same evaluation procedure (criteria selection, content retrieval, and justification generation), enabling direct comparisons. Despite their consistent outputs, LLMs rely on different mechanisms, such as lexical associations and statistical prior knowledge replacing contextual inference. This reliance produces systematic effects, such as political asymmetry, opaque justification, and a tendency to confuse linguistic form with epistemic validity. Delegating judgment to LLMs is therefore not simply automating evaluation, but rather redefining evaluation from normative reasoning to pattern-based approximation.
