Hallucination interview questions

A hallucination is output that's stated with total confidence but is simply not true. What makes it dangerous is that it arrives in the same fluent, assured tone as a correct answer — complete with plausible names, dates, or citations — so there's no nervous wobble to warn the reader. You can't catch it by feel; a bluff reads identically to a fact. That's exactly why you need grounding, citations, and automatic verification instead of trusting the output's confidence.

Its only training objective is to predict the next token that fits the pattern — nowhere does it check a claim against reality, and fluency and accuracy are separate skills it only optimizes the first of. So when it hits something it doesn't know, it doesn't fall silent; it emits the most statistically natural continuation, which often reads like a real answer with invented specifics. There's no built-in uncertainty signal, so the confident guess comes out in the exact same voice as a fact.

The ladder: ground with RAG (fetch real documents, answer only from them), allow "I don't know" so it stops feeling pressure to invent, lower temperature for factual tasks, and ask for citations so claims are traceable. Grounding is the biggest lever because it removes the root cause — answering about things the model was never given. The others tighten behaviour; grounding changes where the facts come from, swapping fuzzy memory for a real page it can even cite.