Deceptive Grounding: Entity Attribution Failure in Clinical RAG

The article introduces "deceptive grounding" (DG), a failure mode in clinical retrieval-augmented generation where model claims are factually sourced from real documents but attributed to the wrong entity. This issue remains invisible to standard faithfulness, hallucination, and citation checks.

• A controlled factorial benchmark across 13 models reveals DG rates of 8-87% under adversarial conditions, with medical fine-tuned models reaching up to 86.7%.
• Ablation studies show that removing entity-specific clinical evidence eliminates attribution failure, shifting errors to confabulation.
• Production measurement across 740 drug-disease pairs found a 7.8% overall DG rate in a deployed system, rising to 13.6% for recently approved drugs.
• Entity-attribution verification detects DG with 97.0% precision and 98.7% recall but is not implemented in existing frameworks.

The authors argue that entity-attribution verification is critical for detecting this failure mode, which domain specialization amplifies rather than mitigates.