When Guidelines Lag Evidence: ESM-Enhanced RAG for Bias-Resistant Clinical AI

As large language models (LLMs) integrate into clinical practice, guideline development, and patient decision support, their tendency to hallucinate (generating fluent yet factually inaccurate outputs) presents a serious risk to evidence-based care. Retrieval-Augmented Generation (RAG) offers a meaningful improvement by linking responses to external knowledge sources, typically reducing error rates by ~30–50% in controlled evaluations. RAG establishes a new standard for AI precision, enabling models like GPT to effectively navigate the nuances of advanced clinical decision-making. However, RAG alone is insufficient when retrieval yields selectively curated or biased references, which can reinforce inherited institutional assumptions rather than prioritizing unfiltered, high-veracity evidence essential for AI-enhanced applications.This article introduces ESM‑Enhanced RAG as a general‑purpose veracity framework for AI systems, using medicine as a proof‑of‑concept domain to both validate the approach and educate clinicians about bias‑reduced AI.ESM‑Enhanced RAG is proposed here as a paradigm‑level architectural framework rather than a fully instantiated, product‑specific system. Its performance estimates are derived from existing RAG enhancement benchmarks and retrieval‑pruning studies, conservatively extrapolated to a dual‑ESM configuration, and are intended as directional guidance for decision‑making rather than empirically validated results. The concrete specification of rules, oversight structures, and implementation‑level metrics will emerge through forward‑looking clinical studies and routine use in actual care settings.The ESM‑Enhanced RAG framework is structured around two principal aims. At the system level, it is designed to suppress hallucinations and strengthen factual reliability across use cases by incorporating formal epistemic bias checks both upstream and downstream of the retrieval‑augmented generation process. At the clinical level, it is operationalized in a high‑stakes scenario—menopausal hormone therapy in breast cancer survivors—to demonstrate performance under real‑world evidentiary tension and to support clinician education in bias‑aware AI–assisted decision making.