Why Do LLMs Hallucinate?

Problem Why do LLMs confidently state facts that are wrong?

Be ready to discuss

• The core mechanism: an LLM is a probabilistic next-token predictor trained to produce plausible continuations, not a database with grounded fact lookup.
• Why confidence and correctness decouple: fluency is what the training objective directly rewards, so a low-knowledge answer still comes out in fluent, assured prose.
• Where hallucinations come from: gaps or noise in pretraining data, facts that changed after the cutoff, leading or underspecified prompts, and long-tail entities the model saw only a handful of times.
• The role of decoding and alignment: sampling temperature, and RLHF pressure to always be helpful and answer rather than say "I don't know".
• Mitigations: retrieval augmentation (RAG) to ground answers in fetched sources, citation and fact-checking pipelines, constrained or tool-based lookup for factual slots.
• Calibration and uncertainty: log-prob or self-consistency signals, abstention thresholds, and letting the model surface uncertainty instead of guessing.