Hallucination

Fluent output that is not grounded in any source. It is a property of next-token prediction, not a bug specific to one model, which is why grounding beats instructing the model to "only say true things" in the prompt.

Why exams ask this

Exams pair this with retrieval and with citation requirements. A common distractor claims a larger or newer model removes the problem outright, when it only lowers the rate. The fix an exam wants is grounding, not a bigger model.

Relevant to

Related concepts

Resources

No resources linked to this concept yet.