Why Do LLMs Hallucinate

Question: Why do LLMs confidently state facts that are wrong? Key points: LLMs are probabilistic next-token predictors trained to produce plausible-sounding continuations of text, not database query engines with grounded fact-lookup. When the model lacks reliable knowledge of a fact (or the prompt nudges it toward a plausible-but-wrong continuation), it still generates fluent, confident-sounding text because fluency and factual correctness are optimized somewhat independently. Mitigations include retrieval augmentation (RAG), fact-checking pipelines, and calibration/uncertainty signaling.