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

The first paper in this series argued that explainability cannot establish auditability or enforce accountability in probabilistic AI systems, because retrospective interpretation is not equivalent to verifiable responsibility. It also identified non-decision as a necessary system state wherever responsibility must remain attributable. Building on that result, the present paper addresses a prior question: what happens when a probabilistic system is structurally required to answer even under conditions in which no epistemically justified answer is available. This paper argues that hallucination in large language models is not best understood as a technical defect, but as the expected epistemic consequence of forced answer generation. In systems that must always produce an output, uncertainty is not represented as uncertainty. It is converted into probabilistic plausibility and expressed as fluent language. As a result, plausible synthesis is easily mistaken for knowledge. The analysis shows that this failure cannot be resolved by explainability. Explanatory techniques operate after an output has already crossed the decision boundary. They may interpret how a response was generated, but they cannot justify whether that response should have been produced at all. Hallucination, overconfidence, and fabricated justification are therefore treated here as related manifestations of the same structural condition: the absence of a legitimate non-answer state. The contribution of the paper is to shift the debate from output correction to epistemic permission. Rather than asking how probabilistic systems can be made to hallucinate less, it asks under what conditions they should be allowed to answer in the first place. In this sense, the paper extends the accountability argument of the first publication by identifying an epistemic precondition without which responsibility attribution cannot meaningfully begin. It thereby prepares the transition to the next step in the series: if explainability cannot secure epistemic legitimacy, accountability will require a more formal language of system states, permissions, and decision boundaries.