LLM Hallucinations
Fluent first, true second
A hallucination is when an LLM states something false as if it were fact — a wrong date, a fake citation, an invented API. It happens because of what the model was trained to do: predict the next token that sounds right, not the one that is right. With no source to check against, it produces the most plausible-looking continuation — even when that's wrong.
Training rewards likely text, not verified truth. There's no built-in fact-checker.
Hallucinations are phrased just as confidently as correct answers — that's what makes them dangerous.
Most likely on niche details, recent events, exact numbers, and anything outside training data.
Watch a fact get invented — then grounded
Ask for a specific date the model is unsure of. It picks a plausible-but-wrong year with high confidence. Then we add retrieval: feed it the real source, and the answer corrects itself.
How to fight it
Retrieval puts real source text in the prompt so the model quotes facts instead of inventing them.
For factual tasks, a low temperature and "say if you don't know" prompting help.
Grounding, tools and careful prompting cut hallucinations sharply but never to zero. For anything that matters — legal, medical, financial — keep a human in the loop and demand sources you can check. Evaluation should measure factuality, not just fluency.
Why does the same model nail one fact and invent another? Training-data familiarity. Slide it: for facts the model barely saw, the plausible-but-wrong year wins — stated just as fluently. Then flip on grounding and watch retrieval rescue even the unfamiliar case.
The dangerous zone is the left half: the model picks 1887 with real confidence and the sentence reads exactly as authoritative as a correct one. Notice grounding makes familiarity irrelevant — the source in the prompt outweighs whatever the weights half-remember.