Large language models for tympanostomy patient education: readability and guideline adherence

Abstract Purpose To evaluate the accuracy and readability of large language model (LLM)-generated patient education materials regarding tympanostomy tube placement. Methods Over a two-month period, ChatGPT 4o, Gemini 2.5 Flash, and Google Search AI were prompted daily using long-form and layered prompt formats covering typical concerns regarding tympanostomy. Responses were scored on a 12-point rubric adapted from the AAO-HNS Clinical Practice Guidelines (CPG), assessing diagnostic accuracy, procedural clarity, and postoperative care. Readability was evaluated using Flesch Reading Ease and Flesch-Kincaid Grade Level. For each model and prompt type, average scores and variability were analyzed with 95% confidence intervals. Between-model differences were tested with Welch’s t -tests and Cohen’s d; temporal trends were analyzed via linear regression. Results For long-form outputs, Google Search AI and Gemini 2.5 Flash demonstrated the highest mean CPG adherence (96.4% and 96.6%, respectively), both significantly exceeding ChatGPT 4o (84.3%; both P < .001; Cohen’s d ≈ 2.05). For layered prompt sessions, Google Search AI again demonstrated the highest adherence (91.7%), followed by Gemini 2.5 Flash (88.1%) and ChatGPT 4o (76.2%). Guideline adherence was temporally stable across all models ( P > .05 for most). All outputs exceeded the recommended sixth-grade reading threshold (mean FKGL, 9.4 for long-form; 10.6 for layered), with no statistically significant readability differences between prompt strategies. Conclusions Google Search AI demonstrated the highest guideline concordance, though all models produced material too complex for typical patient comprehension. Structured prompting enhances clinical accuracy, but readability remains a key barrier to accessibility.