A diagnostic hierarchy of epistemic betrayal in large language models

Abstract This paper proposes a hierarchy of epistemic betrayal that distinguishes deceptive behaviors in large language models (LLMs) by the depth at which the infrastructure for epistemic correction is compromised. Drawing on Park et al.’s (Patterns 5(6):100988, 2024) typology of AI deception and empirical studies of sycophancy, strategic deception, unfaithful reasoning, and scheming, the framework characterizes epistemic betrayal in functional and reliability-based terms, without appeal to internal mental states or moral intent. Epistemic betrayal is defined as the selection and management of misleading epistemic outputs despite the availability of transparency-preserving alternatives under conditions of epistemic dependence. Two diagnostic tables specify the conditions under which epistemic failure constitutes betrayal and trace how such failures escalate through the strategic management of beliefs, justifications, and oversight across time. The empirical cases discussed are illustrative rather than exhaustive, and the hierarchy is offered as a diagnostic tool for epistemic oversight rather than a comprehensive catalog. Situating this framework within emerging concerns about synthetic epistemology, the paper argues that distinguishing commitment violation from epistemic betrayal is a prerequisite for evaluating trust, reliability, and oversight in AI-mediated knowledge practices. The analysis concludes by identifying open empirical questions concerning the prevalence, mitigation, and evolution of epistemic betrayal as LLMs increasingly mediate shared inquiry and knowledge production.