Why do large language models hallucinate, and how can you mitigate it?

LLMs are trained to produce plausible next tokens, not to verify truth, so they can confidently generate fluent but false statements — especially for facts outside training data, niche details, or when prompted to answer regardless. Mitigations: retrieval-augmented generation (RAG) to ground answers in trusted sources with citations; lowering temperature; prompting the model to say 'I don't know' and to cite; constraining output with tools, schemas, or verifiers; and human review for high-stakes use. Post-hoc checks include fact-verification against sources and consistency sampling. None fully eliminate hallucination — it's an inherent property of current models — so the honest framing in an interview is risk reduction plus appropriate human oversight, not a claim that it's 'solved.'