# Hallucination When an AI model generates information that is factually incorrect, fabricated, or unsupported by its training data or context — and presents it with the same confidence as accurate information. A model asked "Who wrote *The Great Gatsby*?" might correctly answer Fitzgerald, but asked about a less common topic might invent a plausible-sounding citation, a fake historical event, or a nonexistent person, with no textual signal that it's making things up. ## Why the metaphor matters "Hallucination" is a contested word choice. The metaphor imports assumptions from human psychology: hallucinations are *perceptual experiences* — you see or hear something that isn't there. Applying this to AI implies the model is *perceiving* reality and occasionally getting it wrong, rather than what's actually happening: the model is generating statistically plausible text with no mechanism for verifying truth. Critics (Emily Bender, [[Timnit Gebru]], and the "stochastic parrot" camp) argue that "hallucination" anthropomorphizes AI systems in a way that obscures the real problem. A model doesn't *believe* anything and isn't *perceiving* anything — it's predicting the next token based on patterns. "Confabulation" or "fabrication" would be more accurate but less evocative. The AI industry has largely adopted "hallucination" because it frames the problem as a tractable engineering challenge (reduce hallucinations through better training, RAG, grounding) rather than a fundamental limitation of the architecture (statistical text prediction has no truth-verification mechanism). The word itself does rhetorical work: it makes the failure sound like an imperfection in an otherwise intelligent system, rather than an inherent feature of systems that generate text without understanding it. ## The practical stakes Hallucination isn't just a terminology debate — it's the primary reliability failure of deployed AI. Lawyers have submitted AI-generated briefs citing fabricated cases. Students have turned in papers with invented sources. Medical AI systems have generated plausible but incorrect diagnoses. In military-intelligence contexts ([[Palantir AIP]], Pentagon AI systems), hallucinated intelligence could lead to targeting errors. [[RAG (Retrieval-Augmented Generation)]] reduces hallucination by grounding model responses in retrieved documents, but doesn't eliminate it — the model can still misinterpret, misquote, or selectively ignore the documents it retrieves. ## See Also - [[RAG (Retrieval-Augmented Generation)]] — the primary technical mitigation - [[Stochastic Parrot]] — the critical framing that disputes the hallucination metaphor - [[Guardrails]] — the engineering response - [[Large Language Model]] — the systems that hallucinate