A colleague claims their model is 'grounded' because it cites sources. Explain why citation presence alone does not prove factual grounding.

Your colleague's reasoning has a gap. A model that cites sources has learned that citations are expected, but it has not necessarily learned to verify what those sources say. Models fabricate citations in two distinct ways. First, they cite real papers but attribute the wrong findings to them. Second, they cite papers that do not exist at all. Both look legitimate on the surface. A reader who does not check the source sees a well referenced answer and assumes it is trustworthy. True grounding means every claim in the output is actually supported by the cited source. Measuring this requires decomposing the output into atomic claims and verifying each one against the original text. Faithfulness metrics like RAGAS and FActScore automate this decomposition and entailment check. The practical takeaway: never use citation count as a proxy for grounding. Use a faithfulness metric that checks what the citations actually say.

**Citation presence** shows the model learned that referencing sources is expected, but it does not prove the cited content supports the claim. Models fabricate citations two ways: citing real papers with wrong findings, and citing nonexistent papers entirely. True factual grounding requires claim level entailment verification. The output is decomposed into atomic claims, and each claim is checked against the cited source. RAGAS faithfulness score and FActScore both implement this pattern. A model that cites five real papers but misattributes findings to each one scores high on citation presence and near zero on faithfulness. The gap between citation presence and verified grounding is exactly where hallucination hides in retrieval augmented systems. Production eval pipelines should measure faithfulness directly, not use citation count as a proxy.