How do factual and faithfulness hallucinations differ, and why does the distinction change your fix?

## Two different reference points Both failures look the same on screen: a confident sentence that should not be there. They differ in what you compare the sentence against. **Factual hallucination** is judged against the world. The model says something reality does not support — a birthdate that is wrong, a statute that does not say what it is claimed to say, a person who never held the role attributed to them. Ground truth lives outside the system, so detecting it requires an external authority. **Faithfulness hallucination** is judged against the material in the context. The model was handed documents and produced a claim those documents do not entail. This includes outright contradiction, but far more often it is quiet extrapolation: a hedge dropped, a number rounded into a different number, a general rule stated where the source stated a conditional one, or background knowledge from pretraining fused into the summary as if it had come from the source. ## They are genuinely independent The four combinations all occur, and interviewers probe whether you see that. A claim can be **unfaithful but true**. A pharmacovigilance summarizer is given three sections of a product label and outputs an interaction warning that appears in none of them. Suppose the interaction is real and well documented in the literature — the summary is still unfaithful, and in a regulated pipeline that is a defect regardless of truth, because the artefact is supposed to be a faithful rendering of that label, and a reviewer downstream will attribute the warning to a document that does not contain it. A claim can be **faithful but false**. If the supplied source is stale, wrong, or a low-quality page pulled in by retrieval, a scrupulously grounded answer inherits its error. Grounding moves the trust boundary to the corpus; it does not create truth. And of course claims can be both faithful and true, or unfaithful and false — the case people picture first. ## Why the distinction changes engineering The two classes respond to different controls. Against **factual** error, the levers are informational. Put authoritative material in front of the model instead of relying on parametric recall. Narrow the question so the answer is a lookup rather than an inference. Resolve anything with a canonical registry — citations, product codes, ticket ids — against that registry and fail closed when it does not resolve. Curate the corpus, because a grounded system is now only as correct as what you grounded it on. Against **faithfulness** error, informational levers are already spent — the right document was in the context and the model still drifted. The levers here are constraint and verification. Instruct explicitly that the answer must come only from the supplied passages and that anything not covered must be reported as not covered. Give the model a real abstention output, so "these sections do not mention interactions" is a success state rather than a failed answer. Ask for span-level attribution, so every claim points at the text it came from. Then run a check: a second pass that re-reads the cited passages and labels each atomic claim supported, unsupported or not addressed, with unsupported claims stripped, rewritten or escalated. For high-stakes output, that check should be the gate, not an advisory. ## Where faithfulness errors come from Understanding the mechanism helps you predict them. Generation is always conditioned on both the context and the model's priors, and the two are blended, not switched between. When the context is thin on a point the priors are strong on, the priors leak in — this is why summaries of unusual documents drift toward the typical document of that genre. Instruction-tuned models also carry a strong pull toward being complete and helpful, so a source that only partly answers the question invites the model to finish the thought. And the more the task asks for synthesis across passages, the more room there is for an inference the sources do not jointly support. Practically, that means faithfulness risk rises with abstractive rather than extractive tasks, with longer and more numerous sources, and with questions the corpus only partially covers — precisely the cases where the answer looks most impressive. ## What to say in an interview Define both, stress that they are independent by giving the unfaithful-but-true example, and then map each to its fix: sources and hard identifier checks for factual error, constrained generation plus attribution and a claim-level verification pass for faithfulness. If pressed on measurement, say that faithfulness is the one you can score without external ground truth — because the reference text is in hand — which is why grounded systems are usually monitored on it first.