TRUTH WITHOUT BELIEF: CAN LARGE LANGUAGE MODELS PRODUCE TRUTHFUL STATEMENTS WITHOUT KNOWING WHAT IS TRUE?

Abstract

The emergence of Large Language Models (LLMs) has challenged conventional assumptions about the relationship between language, truth, belief, and knowledge. LLMs can generate statements that are factually correct and linguistically convincing, yet this capacity does not necessarily imply that the system believes, knows, or understands the truth of what it produces. This study examines how users interpret truthful statements generated by LLMs and how such interactions influence their attribution of belief, knowledge, assertion, and epistemic authority to artificial intelligence. The study adopts a qualitative interpretive phenomenological approach combined with reflexive thematic analysis. The proposed participants are users in higher-education contexts who have experience interacting with LLMs for academic or informational purposes. The analysis is organized around four provisional themes: (1) convincing-looking truth, (2) ruptures in trust, (3) being correct does not mean knowing, and (4) renegotiating epistemic authority. The conceptual analysis indicates that factual correctness and epistemic possession should be treated as distinct categories. An LLM may produce a true proposition without possessing belief or knowledge in the human epistemic sense. Consequently, the apparent truthfulness of an output should not automatically be interpreted as evidence of epistemic agency. The study contributes to the philosophy of language and AI ethics by proposing the concept of truth without belief and by highlighting the distinction between epistemic outsourcing and epistemic transfer. These findings also have implications for AI literacy in higher education, particularly in developing practices of verification, calibrated trust, and critical engagement with generative AI.

Keywords: Large Language Models, Truth, Belief, Knowledge, Epistemic Authority, Speech Acts, AI Literacy