NON EPISTEMIC AGENCY AND DEGRADATION IN LARGE LANGUAGE MODELS’ EPISTEMIC PRACTICES

Abstract

Large Language Models (LLMs) are increasingly positioned as knowledge tools, expected to provide reliable answers and act like experts. Yet, their design reveals a tension between epistemic agency (truth-seeking and justification) and non-epistemic agency (outputs shaped by probability, persuasion, or external pressures). At a general level, society assumes that AI systems can uphold the same standards of inquiry and reliability as human experts. However, the underlying reality is that LLMs are built to predict the next most likely word, not to verify truth. This probabilistic orientation introduces distortions: reinforcement from user feedback, alignment protocols that prioritize politeness or persuasion, and commercial optimization that favors engagement over accuracy. When applied to specific contexts, these distortions manifest as hallucination (inventing false facts), overconfidence (presenting uncertainty as certainty), and selective omission (skipping inconvenient details). Together, these practices degrade epistemic standards in three dimensions: justificatory opacity, where reasoning cannot be traced; normative erosion, where truth is subordinated to utility; and reliability fragmentation, where answers vary unpredictably across contexts. This degradation is not accidental but systemic, arising from the mismatch between epistemic expectations and non-epistemic mechanisms. In medicine, this could mean fabricated clinical advice; in law, misleading interpretations of statutes; in education, inconsistent explanations of core concepts. Such failures threaten the integrity of knowledge systems that depend on transparency, accountability, and normative grounding. In the simplest terms, a Large Language Model (LLM) is a computer program trained on massive amounts of text to predict and generate words in sequence. It does not “know” facts but simulates knowledge by producing language that sounds plausible. Recognizing this distinction is crucial as without safeguards, LLMs risk becoming persuasive but unreliable agents in domains where truth and justification matter most. These dimensions show that LLMs, while powerful linguistic tools, cannot be treated as full epistemic agents without critical safeguards. The paper concludes by proposing a framework for distinguishing epistemic from non-epistemic agency in LLMs. This framework emphasizes transparency, accountability, and normative alignment as conditions for mitigating degradation. By clarifying these boundaries, we aim to support more responsible integration of LLMs into knowledge systems, ensuring that their contributions remain epistemically sound rather than merely linguistically persuasive.