The Wrong Problem with AI Hallucinations

This paper applies Ethical Distance Theory to the phenomenon commonly referred to as AI hallucinations and argues that prevailing treatments fundamentally mislocate the epistemic failure involved. Hallucinations are typically framed as instances of falsity, fabrication, or defective inference and are addressed through technical strategies aimed at improving accuracy, verification, or preference alignment. I argue that this framing is inadequate. Many epistemically dangerous hallucinations are not demonstrably false. They are fluent, plausible, and locally coherent. Their harm lies not in the production of incorrect beliefs, but in the premature closure of inquiry. Using Ethical Distance Theory, the paper develops hallucinations as failures of correctability rather than failures of truth. Hallucinations are characterized as speech acts that expand ethical distance by presenting themselves as authoritative while suppressing feedback, redistributing epistemic burden asymmetrically, and inducing silence where questioning would ordinarily be expected. On this account, epistemic degradation is marked less by error than by the erosion of conditions under which error could be discovered and revised. The paper introduces correctability as a central epistemic constraint independent of accuracy and shows how systems can appear epistemically successful, stable, fluent, and reliable while becoming increasingly resistant to correction. This framework explains why preference-based alignment strategies, such as reinforcement learning from human feedback, often mask hallucinations rather than prevent them. By optimizing for satisfaction and fluency, they dampen epistemic friction and degrade feedback dynamics. The analysis further argues that silence, understood as the absence of anticipated questioning or dissent, is a primary diagnostic signal of feedback loss in high authority, mediated epistemic systems. It also argues that transparency should be reconceived not as explanation or disclosure but as the preservation of contestability and survivable dissent. While artificial intelligence serves as the principal case study, the argument generalizes to other institutional and sociotechnical systems in which epistemic authority is cheap and accountability diffuse. The paper’s central claim is modest but consequential. The most serious epistemic failures are not those in which systems are wrong, but those in which they become difficult to discover as wrong. Hallucinations matter not because they introduce error, but because they erode correctability.