Epistemic Safety: The Structural Blind Spot in AI Alignment

Current AI safety mechanisms—Constitutional AI, RLHF, safety classifiers, and redteaming—address what models say but not how they reason. We term this gap epistemic safety. Wedemonstrate that this gap is not accidental but structural: it arises from an assumption, inherited fromthe dominant tradition in Western epistemology, that there is a single truth describable in language.This assumption functions for content safety but breaks down for reasoning soundness, where thesame argument pattern can be legitimate or fallacious depending on context, power relations, andburden of proof allocation. Drawing on the rhetorical theory of Hidenobu Kozai(1958–2013)—virtually unknown in English-language scholarship—and a suggestive parallel withthe relativity of truth to axiom systems in mathematics, we diagnose the root cause and derive sixnecessary conditions that any viable approach to epistemic safety must satisfy. We show that nocurrently known approach meets all six conditions simultaneously. Our central proposition is thatfallaciousness is not a property of text but a property of the relationship between text,recipient, and context—a finding that implies the entire text-level architecture of current AI safetyrequires fundamental redesign for epistemic safety.