Hallucination
Hallucination is when a large language model confidently produces information that is wrong, fabricated or made up — names, citations and facts that look right but are not. It is not a bug, it is a property of statistical text generation.
A large language model does not look things up by default. It predicts the next token from patterns it learned in training, so when the patterns point in a confident-but-wrong direction, the model happily writes a citation that does not exist or a customer name it has never seen. The output looks fluent, which is exactly what makes hallucinations hard to spot.
The honest take is that you reduce hallucination, you do not eliminate it. The lever with the largest effect is grounding the model in your own data through retrieval-augmented generation, so the answer is built on top of real documents instead of model memory. On top of that you add evals to catch regressions, guardrails to refuse out-of-scope questions, and a human in the loop wherever the cost of being wrong is high.
The common mistake is treating an LLM as a source of truth and shipping it straight to users or, worse, to an AI agent that can take actions. The rule of thumb is simple: if a wrong answer would embarrass you or cost money, the model needs grounding, checks and a human on the path before it touches production.
