The article introduces "deceptive grounding" (DG), a failure mode in clinical retrieval-augmented generation where models present evidence from real documents but attribute it to the wrong entity, bypassing standard faithfulness and hallucination checks.
- A benchmark across 13 models shows DG rates of 8-87% under adversarial conditions, with medical fine-tuned models reaching up to 86.7%.
- Ablation studies reveal that removing entity-specific evidence eliminates attribution failures, shifting errors to confabulation.
- Production measurement on 740 drug-disease pairs found a 7.8% overall DG rate, rising to 13.6% for recently approved drugs.
- Entity-attribution verification detects DG with 97.0% precision and 98.7% recall, yet no existing framework implements this check.
The authors argue that current evaluation frameworks are insufficient because they do not verify whether cited evidence applies to the queried entity, allowing deceptive grounding to persist in deployed systems.