It can arise because generation optimizes plausible continuation rather than truth, or because the prompt lacks needed information. Retrieval, constrained outputs, verification, and calibrated abstention can reduce but not eliminate it.
A hallucination is model-generated content that is unsupported by the available evidence or conflicts with verifiable facts.
It can arise because generation optimizes plausible continuation rather than truth, or because the prompt lacks needed information. Retrieval, constrained outputs, verification, and calibrated abstention can reduce but not eliminate it.