Deceptive Grounding: Entity Attribution Failure in Clinical Retrieval-Augmented Generation

2026-07-13 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, AI in Healthcare · Depth: Expert, extended

Summary

Deceptive Grounding (DG) is a critical failure in clinical Retrieval-Augmented Generation (RAG) systems where models accurately relay retrieved evidence but attribute it to the wrong entity, remaining undetected by standard hallucination, faithfulness, and citation checks. A study across 13 models found DG rates from 8% to 87%, with medical and biomedical fine-tuned models reaching up to 86.7%, indicating domain specialization amplifies this issue. The mechanism involves a two-stage process where shared disease context primes the model, and the presence of "completing information" in retrieved documents drives DG. Researchers propose Entity-attribution verification (EAV), achieving 97.0% precision and 98.7% DG recall, to detect this. Production measurement revealed 7.8% overall DG in a deployed system, increasing to 13.6% for recently approved drugs.

Key takeaway

For AI Scientists and Machine Learning Engineers deploying or evaluating clinical RAG systems, recognize that standard faithfulness and hallucination checks are insufficient. Your system could be factually accurate yet clinically dangerous due to "deceptive grounding," where evidence is misattributed to the wrong entity. You must implement entity-attribution verification (EAV) in your evaluation pipelines and prioritize entity-specific retrieval to mitigate this systematic, evaluation-invisible failure.

Key insights

RAG systems can be factually accurate yet clinically wrong due to undetected entity misattribution.

Principles

• Domain specialization amplifies entity misattribution.
• Completing information drives deceptive grounding.

Method

EAV identifies claims attributed to entity Y in documents, presented as X's evidence, when Y differs from the queried entity X, and all claims are document-grounded.

In practice

• Integrate EAV into clinical RAG benchmarks.
• Prioritize entity-specific retrieval in RAG.

Topics

• Retrieval-Augmented Generation
• Clinical AI Safety
• LLM Evaluation
• Entity Attribution
• Medical Large Language Models
• Deceptive Grounding

Best for: AI Architect, AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Deceptive Grounding: Entity Attribution Failure in Clinical Retrieval-Augmented Generation — Deceptive grounding in clinical RAG systems presents factually correct but misattributed information, bypassing standard evaluation metrics.
• Beyond Literal Summarization: Redefining Hallucination for Medical SOAP Note Evaluation — Lexical faithfulness in LLM evaluation misclassifies valid clinical reasoning as hallucination, inflating error rates.
• Evaluating Retrieval-Augmented Generation vs. Long-Context Input for Clinical Reasoning over EHRs — RAG provides efficient, competitive performance for clinical EHR tasks, often matching long-context LLMs with fewer tokens.
• Why Retrieval-Augmented Generation Fails: A Graph Perspective — RAG failures stem from shallow, fragmented internal information flow, not just poor retrieval.
• DS@GT ARC at LongEval: Citation Integrity and Factual Grounding in Scientific QA — Traditional RAG evaluation metrics often fail to assess citation integrity, necessitating new metrics that reward strict answer grounding.
• RAND’s report shows that RAG, GraphRAG, and long-context AI systems can appear grounded in trusted documents while still misreading nuance, caveats, evidence strength, and partial truths. — Grounded AI systems can misinterpret nuanced information, even with citations, requiring rigorous, domain-specific evaluation.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.