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

Attempt it

Quick reference

TL;DR

Unfaithful means not supported by the retrieved context; hallucinated means not true in reality. Every hallucination is unfaithful, but an unfaithful claim can still be factually correct.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Imagine an open-book exam where you must answer using only the textbook on your desk. Faithfulness asks one thing: did you stick to the book? If you write something the book never says, you broke the rule, even if it happens to be true from your own memory. Hallucination asks a different thing: is the statement actually false in the real world? So there are two separate referees. The faithfulness referee only checks the book. The hallucination referee checks reality. A made-up false fact fails both referees. But a true fact you recalled from memory, that the book never mentioned, fails the faithfulness referee while passing the reality referee. That gap is the whole point: faithful and true are not the same question.

Key concepts

Faithfulness Hallucination Factuality RAG eval metrics

Quick scan · 2–4 min

Everything important, quickly.

5 min: two reference points (context vs reality), the subset relationship, the faithful but wrong corollary, how factuality forms a third axis, and how to operationalise each in a RAG harness.

Side-by-side

Property	Faithfulness	Hallucination	Factuality
Reference point	Retrieved context	Real-world truth	Real-world truth
Question asked	Is the claim supported by context?	Is the claim fabricated or false?	Is the claim true?
True fact absent from context	Unfaithful	Not a hallucination	Factual
Faithful answer, wrong context	Faithful	Repeats a false claim	Not factual
Typical metric	Claim-level entailment vs context	Inverse of factual consistency	Atomic-fact verification vs source

Real-world examples

Real products, models, and research that use this idea.

RAGAS computes faithfulness as claim-level entailment against retrieved context, kept separate from its answer-correctness (factuality) metric.
TruLens scores groundedness (its name for faithfulness) independently of context relevance and answer relevance in the RAG triad.
Galileo and Patronus surface a context-adherence score distinct from a real-world factuality check in their hallucination dashboards.
A medical RAG copilot judged by Claude Opus 4.7 can add a true side effect from training that the retrieved leaflet omitted: unfaithful, not hallucinated.
FActScore decomposes long-form answers into atomic facts and verifies each against a knowledge source, measuring factuality rather than context faithfulness.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QIf your faithfulness metric is green but users still report wrong answers, where is the bug?

Sign in or go Pro to see the approach for this follow-up.

QHow would you operationalise a faithfulness metric in a RAG eval harness?

Sign in or go Pro to see the approach for this follow-up.

QWhy keep faithfulness and factuality as separate metrics instead of one combined score?

Sign in or go Pro to see the approach for this follow-up.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.