Why models "hallucinate"
A hallucination is when the model confidently states something false - a fake statistic, a book that doesn't exist, a wrong date. Last month, a major airline's chatbot hallucinated a refund policy and the court forced them to honor it. This happens because the model isn't looking anything up by default; it's generating the statistically most-likely next words based on patterns in its training data, and sometimes the most "plausible-sounding" answer just isn't true.
The model has no built-in fact-checking step - it doesn't distinguish between "I recall this precisely" and "this pattern feels right." Hallucination rates go up with obscure facts, exact numbers, quotes, and citations - anything requiring precision rather than pattern completion. This is exactly why tools like web search or RAG (Cluster C) exist: to give the model real, current text to ground its answer in, instead of relying purely on memorized patterns.
Ask Claude or ChatGPT (without web search enabled) for "the exact publication date and page number of a specific quote" from an obscure book. Then ask it to search the web for the same thing. Compare confidence and accuracy - a direct feel for grounded vs. ungrounded answers.
Why do models hallucinate?