Hallucination
When a model produces confident, fluent output that is factually wrong or fabricated.
A hallucination is a plausible-sounding answer that is not true — an invented citation, a made-up figure, a policy that does not exist. It happens because a language model predicts likely text, and a fluent falsehood can be more likely than an honest 'I don't know'.
Prompting reduces hallucination without eliminating it: grounding the model in retrieved sources, giving it explicit permission to say the answer is not available, and asking it to cite evidence all shift the odds toward honesty. In extraction tasks, an explicit 'use null rather than infer' rule plays the same role.