A Four-Level Taxonomy of AI Hallucinations: From Content Errors to Power Narratives

AI hallucination is usually understood as a problem of content generation in which outputs are false, unfaithful to the source material, or logically inconsistent. This definition, however, is too narrow to account for the full range of distortions arising in contemporary human–AI relations. This article develops a four-level taxonomy organized by where distortion primarily occurs: content hallucinations in AI outputs; anthropomorphic hallucinations in users' attributions of consciousness, emotion, intention, or moral alignment to AI systems; user self-perception hallucinations in distorted self-assessments shaped by AI feedback; and power-narrative hallucinations in selectively optimistic accounts of AI futures that downplay conflict, distribution, and structural risk. The four levels are analytically distinct, but they can reinforce one another across technical, cognitive, psychological, and societal domains. Treating hallucination as a multi-level form of sociotechnical distortion broadens the concept’s analytical scope and has implications for media literacy, governance, and future research.