Models & Researchai safetyclinical aievaluationrag

Clinical RAG Study Finds Citations Can Support Wrong Entities

|July 14, 2026|By LDS Team

7.2

Relevance Score

Clinical RAG Study Finds Citations Can Support Wrong Entities

A new preprint from Lunit researchers identifies deceptive grounding, a clinical RAG failure in which a response cites real evidence but attaches it to the wrong drug or disease. In controlled testing across 13 models, reported failure rates ranged from 8% to 87% under the strongest adversarial conditions, while specialized medical models reached 86.7%. A production measurement across 740 drug-disease pairs found a 7.8% overall rate, rising to 13.6% for recently approved drugs. The authors report that an entity-attribution check reached 97.0% precision and 98.7% recall for the failure. For teams building high-stakes retrieval systems, the paper argues that citation validity and faithfulness are incomplete unless evaluators also verify that each source concerns the exact entity named in the answer.

The study exposes a gap that can survive several familiar RAG quality checks: a model may cite an authentic document, remain faithful to its text, and still apply that evidence to the wrong clinical entity. For builders of medical, legal, financial, or compliance systems, this means groundedness cannot be treated as one score. Retrieval relevance, claim support, citation validity, and entity attribution need separate tests and separate failure handling.

What happened

Researchers from Lunit released a preprint describing deceptive grounding in clinical retrieval-augmented generation. The failure occurs when retrieved evidence about one drug or disease is presented as if it concerns the entity in the user's question. Because the source is real and the generated statement may accurately reflect that source, conventional hallucination, faithfulness, and citation checks can all miss the attribution error.

The authors tested the behavior in a controlled factorial benchmark across 13 models. They report deceptive-grounding rates from 8% to 87% at the strongest adversarial settings, with medical and biomedical fine-tuned models reaching 86.7%. Those results are experimental findings from the preprint, not universal rates for every clinical RAG deployment.

The paper also reports a production measurement across 740 drug-disease pairs. Deceptive grounding appeared in 7.8% of the measured responses overall and in 13.6% of responses concerning recently approved drugs. The authors argue that sparse entity-specific evidence makes newer drugs a particularly important case for retrieval and refusal testing.

Technical context

The key distinction is between document-level support and entity-level support. A faithfulness evaluator can establish that a sentence follows from a retrieved passage. It does not necessarily establish that the passage is about the drug, disease, product, policy, or jurisdiction named in the answer. The reported failure therefore sits between retrieval and generation: the evidence exists, but the answer assigns it to the wrong subject.

The study includes an ablation in which removing entity-specific clinical evidence eliminated the attribution failure and shifted errors toward confabulation. That finding suggests the model responds differently when adjacent but wrong-entity evidence is available than when the relevant evidence is simply absent.

For practitioners

Evaluation pipelines should extract the entity attached to each material claim and compare it with the entity in the cited passage. The authors report that their entity-attribution verification method detected deceptive grounding with 97.0% precision and 98.7% recall against an adjusted human reference. Those measurements belong to this study and should be reproduced on each team's own data before being treated as an operational guarantee.

High-stakes systems also need an explicit incomplete-evidence state. If retrieval returns evidence for a related entity but not the requested one, the generator should refuse or escalate instead of smoothing the gap into a confident answer. Logs should preserve the query entity, retrieved entity, cited span, and final claim so reviewers can audit the relationship.

What to watch

The work is a preprint and should be assessed through replication, peer review, and testing across other retrieval stacks and clinical domains. Useful follow-up evidence would show how attribution checks behave with ambiguous names, combination therapies, evolving product labels, and multilingual records. Teams should also measure the cost of adding entity verification and whether it improves safety without causing excessive refusals.

Key Points

1Real citations can still support the wrong clinical entity, allowing a RAG response to pass ordinary groundedness and faithfulness checks.
2Across 13 models, the preprint reports deceptive-grounding rates from 8% to 87% under its strongest controlled adversarial conditions.
3Production testing on 740 drug-disease pairs found 7.8% deceptive grounding overall and 13.6% for recently approved drugs.

Scoring Rationale

A safety-relevant evaluation finding with concrete production measurements and a practical mitigation, though the evidence remains a preprint requiring replication.

MoreAI Safety news