Large language models match and occasionally surpass the SRS AIS handbook in adolescent idiopathic scoliosis education

BACKGROUND: Patients increasingly seek medical information from artificial intelligence (AI) tools. The accuracy and clinical alignment of large language models (LLMs) relative to established, surgeon-curated patient education resources remain incompletely characterized. PURPOSE: To compare LLM-generated responses with the Scoliosis Research Society (SRS) Adolescent Idiopathic Scoliosis (AIS) Handbook using a surgeon-adjudicated alignment scoring system. STUDY DESIGN: Questionnaire-based comparative study. METHODS: Four LLMs (Gemini 3, Claude Sonnet 4.5, GPT-5.2, and Grok 4.1) were prompted using all 19 SRS AIS frequently asked questions (FAQs) items. Three fellowship-trained spine surgeons, blinded to model identity, independently evaluated 76 LLM responses using a novel 6-point alignment scale assessing accuracy, completeness, and concordance with SRS guidance. Responses contradicting the handbook but judged more accurate by surgeon consensus received the maximum score of 6. Inter-rater reliability was assessed at the item and model level using intraclass correlation coefficients (ICC). Model performance was compared using repeated-measures analysis of variance (ANOVA) with Holm-adjusted pairwise comparisons. RESULTS: Mean scores (95% CI) were 5.23 (5.04-5.41) for Grok 4.1, 4.91 (4.66-5.16) for Gemini 3, 4.75 (4.44-5.07) for GPT-5.2, and 4.51 (4.25-4.77) for Claude Sonnet 4.5 (F (3, 54) = 7.63, p<0.001). Inter-rater reliability was moderate at the item level (ICC [2,1] = 0.45; ICC [2,3] = 0.71) and excellent at the model level (ICC [2,1] = 0.82; ICC [2,3] = 0.93). Overall, 82.5% of ratings (188/228) achieved scores ≥5, indicating strong concordance with or extension beyond guideline recommendations. Notably, 12.3% of ratings (28/228) received the maximum score of 6, reflecting surgeon consensus that the LLM output contradicted but exceeded the SRS handbook in accuracy. No ratings were graded as completely incorrect. CONCLUSIONS: Modern LLMs generate AIS patient education content that matches and occasionally surpasses the SRS AIS Handbook. Discrepancies between AI-generated responses and guideline-based materials may reflect evolving clinical evidence and highlight opportunities for updating patient-facing resources. Because LLMs increasingly reference authoritative educational materials, updating the SRS Handbook would directly enhance both traditional and AI-mediated patient education.