Why it matters
Hallucinations cause real harm: legal citations to fictional cases, medical advice that's wrong, business decisions based on fabricated data. Deploying LLMs safely requires accounting for this.
The architecture
Training: LLMs are trained to predict next tokens. There's no explicit truthfulness objective. Coherent text is rewarded even when factually wrong.
Confidence miscalibration: LLMs don't reliably indicate uncertainty. Fabricated content is generated with the same confidence as accurate content.
How it works end to end
Mitigations: RAG (retrieval-augmented generation) grounds responses in retrieved documents. Model answers 'according to source X...' and can be checked.
Citations: require model to cite sources for factual claims. Check citations for existence and relevance.
Multi-model verification: ask multiple models; disagreement indicates uncertainty.
Domain-specific fine-tuning + guardrails: constrain output to verified sources for high-stakes domains.