Auditing Institutional Heterogeneity for Generative AI in Patient Education: A Large-Scale Study of 102 US Transplant Handbooks

Synthesis: Li, Padman and Krishnan audit 102 US transplant-center patient handbooks that serve as grounding corpora for generative AI patient-education assistants. They show large institutional heterogeneity in the underlying education materials, undermining the premise that grounding a genAI assistant in local content yields consistent guidance: patients at different institutions can receive materially different AI-mediated answers to the same question. The study extends grounding-quality concerns familiar from Retrieval-Augmented Tutoring for Algorithm Tracing and Problem-Solving in AI Education into health education, and connects to AI-driven medical training work such as MedEasy: Designing AI Standardized Patients for Clinical Consultation Training, The Empirically Grounded Adaptive Virtual Patient for Psychotherapy Training: Disclosure That Responds to Therapist and MedGame: Storytelling Gamification Empowered by Large Language Models for Medical Education. The equity implication (institution-dependent quality of AI-mediated education) parallels Equity.

Key Findings

• The study audits 102 US transplant-center patient handbooks that serve as grounding corpora for genAI patient-education assistants.
• Large institutional heterogeneity in the underlying education materials undermines the premise that grounding a genAI assistant in local content yields consistent guidance.
• Patients at different institutions can receive materially different AI-mediated answers to the same question.
• The equity implication is that the quality of AI-mediated education becomes institution-dependent, mirroring broader concerns in Equity.

What this means for practice

• Institutions. Run the pairwise-judge self-consistency audit over your own patient-education corpus before deploying a grounded patient-facing assistant; the authors report the instrument as reproducible and tractable on a local corpus.
• Institutions. Treat silence as a fixable content gap rather than a prompt problem: reproductive health was the most-silent topic (82% absent) and pre-transplant handbooks were the most silent of all (mean 90% absence), precisely when fertility-preservation and pregnancy-timing decisions are made.
• Institutions. Do not rank centers or handbooks by raw disagreement rates; divergence may reflect legitimate medical variation, differing risk tolerance, or simply more comprehensive documentation, and a failed retrieval can mimic disagreement.
• Policymakers. Fold corpus-level self-consistency auditing into existing pre-deployment audit and monitoring expectations for clinical Generative AI, rather than treating grounding in local documents as sufficient assurance of consistent guidance.
• Policymakers. Ask deployers to report disagreement and retrieval failure separately, because the study shows apparent divergence between institutions can arise from RAG (Retrieval-Augmented Generation) retrieval behavior rather than from editorial conflict.

Limitations

• Single domain: the audit covers 102 handbooks from 23 U.S. solid-organ transplant centers, and the authors state the specific blind-spot topics, divergence themes, and institutional-voice effects should not be assumed to generalize to other clinical domains without replication.
• Judge bias was not eliminated: a 200-pair human-judge agreement study is informative for population-level kappa but is not powered to detect bias along sub-axes such as institution, organ, answer length, or clinical-significance rating.
• English-only and snapshot-in-time: the corpus and the 1,115 TransplantQA questions are English, and handbooks were collected over a single 2024-2025 window even though many centers also distribute Spanish and Mandarin materials and guidance evolves.
• Retrieval-induced apparent divergence: differences between two handbooks' answers may reflect one poorly matching retrieval rather than institutional disagreement, and the absence pre-screen only partially mitigates this without controlled passage baselines.

Citation

Yubo Li, Rema Padman, Ramayya Krishnan (2026). Auditing Institutional Heterogeneity for Generative AI in Patient Education: A Large-Scale Study of 102 US Transplant Handbooks.