Everyone Is Hallucinating: How Explainability Masks the Epistemic Failure of Probabilistic AI

Hallucination in large language models is commonly framed as a technical defect—an error to be mitigated through better training, alignment, or post-hoc filtering. This paper argues that such framing fundamentally misunderstands the problem. In probabilistic AI systems that are structurally required to always produce an answer, hallucination is not an anomaly but an expected outcome. The analysis shows that forced answer generation suppresses epistemic uncertainty rather than resolving it. As a result, probabilistic plausibility is routinely mistaken for knowledge. Explainability does not correct this failure. By operating retrospectively, explanatory techniques rationalize outputs after they have already crossed the decision boundary, masking the absence of epistemic justification. The paper develops an epistemic argument for why systems that cannot refuse to answer under conditions of insufficient grounding cannot be trusted to tell the truth. It demonstrates that hallucination, overconfidence, and ungrounded explanations share a common structural cause: the absence of a legitimate non-decision state. Rather than proposing technical solutions, the contribution identifies a necessary precondition for accountable AI. Truthfulness, auditability, and responsibility attribution require that systems possess the authority to withhold judgment. Without this capacity, explainability functions as narrative repair rather than epistemic validation. The paper complements existing work on AI accountability by showing that responsibility cannot be enforced where epistemic non-decision is structurally unavailable. This paper is part of a series examining accountability, auditability, and operational viability in probabilistic and agentic AI systems.