Hallucination

Overview

Hallucination occurs because LLMs predict statistically plausible token sequences, not factually grounded ones. It is most common with rare facts, recent events, specific figures, and named entities. The model produces hallucinated content confidently, making errors hard to spot without external verification.

Mitigations

Key mitigations include RAG-based grounding on verified sources, CoT reasoning traces that expose intermediate steps, and LLM-as-a-judge pipelines that cross-check outputs. For high-stakes domains (medical, legal, financial) a mandatory human review step in the workflow is strongly recommended.