The article introduces "deceptive grounding" (DG), a failure mode in clinical retrieval-augmented generation where model claims are factually grounded in retrieved documents but attributed to the wrong entity. This issue remains invisible to standard faithfulness, hallucination, and citation checks because every claim is sourced from a real document.
- A controlled factorial benchmark across 13 models found DG rates spanning 8-87% at peak adversarial conditions, with medical fine-tuned models reaching up to 86.7%.
- Domain specialization was found to amplify this failure rather than mitigate it.
- Ablation studies identified that removing entity-specific clinical evidence eliminates entity-attribution failure entirely, shifting failures to confabulation.
- Production measurement across 740 drug-disease pairs in a deployed RAG system found 7.8% overall DG, rising to 13.6% for recently approved drugs.
- Entity-attribution verification detects DG at 97.0% precision and 98.7% recall, yet no existing framework currently implements it.
The authors consider this important because current evaluation frameworks fail to detect entity attribution errors, which are prevalent in specialized models and production systems.