Mapping the Reliability–Readability Gap in AMD Patient Education Across Six Large Language Models (Preprint)

BACKGROUND Age-related macular degeneration (AMD) is a leading cause of irreversible vision loss globally, requiring patients to understand complex, long-term management plans. While large language models (LLMs) offer a scalable solution for patient education, their direct clinical use is hindered by two critical gaps: excessive readability (often at high-school or college level, far above the recommended 6th-grade standard) and an apparent trade-off where more reliable, comprehensive outputs tend to be less readable. Current evaluations lack head-to-head comparisons of state-of-the-art models in specialized domains like AMD under realistic “zero-shot” conditions that mimic patient queries. This study systematically benchmarks six leading LLMs to quantify this reliability–readability gap, providing an evidence base for the safe, informed integration of AI into ophthalmic patient communication. OBJECTIVE To address the critical gap between scalable AI communication and clinically safe patient education, this study aimed to benchmark state-of-the-art LLMs for AMD by jointly quantifying informational reliability and linguistic readability under a realistic zero-shot (naïve user) query scenario—a key but under-evaluated setting for clinical deployment. METHODS Thirty AMD-related patient questions were curated from Google Trends (Oct 10, 2020–Oct 10, 2025), the 2023 Chinese AMD guideline, and the 2024 AAO recommendations. Each question was entered verbatim into six publicly available LLMs (ChatGPT-5.1-auto, DeepSeek-v3.2, Gemini-2.5-Flash-Thinking, Grok 4, Claude-Sonnet 4.5, Qwen3-Max) during Oct 10–Nov 25, 2025. Two senior ophthalmologists, blinded to model identity, independently scored all responses using DISCERN, EQIP, GQS, and JAMA criteria, with adjudication for disagreements. Readability was assessed using six standard formulas against a ≤6th-grade benchmark. Between-model differences were analyzed using Friedman tests with Holm-adjusted pairwise comparisons. RESULTS RESULTS: Analysis of 180 responses revealed substantial to near-perfect inter-rater agreement (κ=0.72–0.97). Critically, no model met the recommended ≤6th-grade readability target, and a clear reliability–readability trade-off was observed: Grok 4 achieved the highest reliability (DISCERN 46.40±7.43; EQIP 74.33±9.07) while DeepSeek-v3.2 generated the most readable text (FRES 48.23±9.16; FKGL 9.95±1.87). Between-model differences were significant across all metrics (all P