Is 'unfaithful' the same as 'hallucinated'? What is the relationship between these concepts?

**Faithfulness** asks one narrow question: is every claim in the answer supported by the context the model was given? In a RAG system, that context is the retrieved documents. If the answer says something the documents do not say, it is unfaithful, full stop. Hallucination asks a different question: is the claim actually false in the real world? A hallucinated claim is fabricated or wrong. The two reference points differ. Faithfulness compares the answer to the context. Hallucination compares the answer to reality. Because of this, the concepts overlap but are not identical. Every hallucination is unfaithful: if a claim is not real, the retrieved documents (assumed correct) could not have supported it. But the reverse fails. A model can add a true fact from its training memory that the documents never mentioned. That fact is unfaithful to the context, yet not hallucinated, because it is true. The correct answer captures this: hallucination is a subset of unfaithfulness. Unfaithfulness is the broader category, and a faithful answer is not automatically a correct one.

**Faithfulness** and hallucination are related but distinct. They use different reference points. Faithfulness is context-relative. A claim is faithful if the retrieved context entails it. Hallucination is reality-relative. A claim is a hallucination if it is false in the real world. These are two separate referees checking two separate things. The set relationship is clean. Every hallucination is unfaithful, because a false claim cannot be supported by correct context. But an unfaithful claim can still be true. A model may add a correct fact from parametric memory that no retrieved document contained. That answer is unfaithful, yet not hallucinated. So hallucination is a strict subset of unfaithfulness. Unfaithfulness is the superset. One sharp corollary follows. A perfectly faithful answer can still be globally wrong. If the retrieved context itself is wrong, a faithful answer dutifully repeats the error. Faithfulness measures grounding, not truth. That is why factuality is a third, separate axis. It compares claims to the real world regardless of what was retrieved. In practice, frameworks like RAGAS measure faithfulness with claim-level entailment against context, paired with a separate factuality check.