The Epistemic Illusion of Large Language Models

Abstract

Large Language Models produce text whose argumentative structure, discursive coherence, and pragmatic register replicate the formal surface of epistemically competent discourse. This paper identifies and analyzes the epistemic illusion: the systematic phenomenon whereby such linguistically plausible outputs are perceived by users as expressions of knowledge, comprehension, or judgment. The illusion is relational: it arises from the encounter between an architecture optimized for syntactic plausibility and a human cognitive apparatus disposed to treat coherent language as a proxy for epistemic reliability. Three manifestations are distinguished: apparent transparency, textual indistinguishability, and sycophancy. The paper proposes a three-tiered analysis of truth (syntactic plausibility, structural satisfaction, ontological anchoring) and shows that LLMs satisfy the first level systematically, the second accidentally, and the third never. A distinction between functional and phenomenally qualified beliefs establishes that the doxastic condition of knowledge is met, if at all, only in a derivative sense. The analysis converges with the symbol grounding problem: the absence of bottom-up anchoring to the world is an architectural limitation of current LLMs, not a contingent defect. The strongest available objection, that behavioral evidence warrants the attribution of general intelligence, is addressed and shown to rest on an asymmetry between two epistemic situations: the human mind, whose generative mechanisms are not fully specified, and LLMs, whose architecture is known and sufficient to account for their outputs. The epistemic illusion is therefore shown to be a structural property of the relation between system and user, irreducible to a defect on either side.