Naming as Intervention: The "Hallucination" Taxonomy as Epistemic Misdirection in Generative Transformer Research

The term hallucination is not a neutral label. It already carries the image of a mind gone astray: perception detached from the world, speech unmoored from reason. In debates about large language models (LLMs), the word has become a familiar name for fluent but false or misleading output. Yet the term does more than describe. In technical taxonomies, it marks a class of model failure. In corporate language, it can relocate failure into the model itself, making decisions about design, deployment, verification, and authorization less visible. In everyday use, the term stretches further still. A user may call an answer hallucinated because it is wrong, irrelevant, evasive, unsupported, or simply not what was needed. This looseness encourages the assumption that language models ordinarily retrieve truth and only occasionally depart into fabrication. The stakes are ethical and political: calling false output hallucination can make failures of deployment, oversight, and accountability appear as technical limitations of the model, to be managed rather than questioned. This position paper challenges that assumption. In the base generative process, transformer-based language models do not alternate between a mode of retrieval and a mode of hallucination. They generate utterances through a single probabilistic process. What changes is the later relation between the utterance and a reference state established outside it. An answer may be accepted, rejected, qualified, or left unverifiable, depending on context, evidence, and institutional procedure. The central distinction is therefore not between truthful output and hallucinated output, but between utterance and verification — a distinction the Oracle Model makes the basis of its accountability framework. In the Oracle Model, the LLM is not a failed knower. It utters; observers, systems, and institutions verify and authorize those utterances. Responsibility must therefore return to the verification chain.