Can Artificial Intelligence Align with Evidence? Performance of ChatGPT-4o in Knee Osteoarthritis Surgical Guidelines

Abstract Artificial intelligence large language models (LLMs) such as ChatGPT are increasingly used in clinical settings, yet their reliability in reproducing evidence-based recommendations remains uncertain. This study aimed to evaluate the performance of ChatGPT-4o in addressing clinical practice guideline (CPG) recommendations for the surgical management of knee osteoarthritis and total knee arthroplasty (TKA). An observational cross-sectional design was conducted in September 2025. Twenty recommendations from the most recent American Academy of Orthopaedic Surgeons CPG on TKA were translated into structured clinical questions and submitted to ChatGPT-4o. Each query was entered three times in independent sessions to evaluate textual consistency. Two independent reviewers with expertise in musculoskeletal physiotherapy and orthopedics appraised the chatbot's answers, classifying them according to the CPG framework (“should do,” “could do,” “do not do,” “uncertain”). Agreement between reviewers and alignment with CPG recommendations were assessed using Cohen's and Fleiss' Kappa coefficients. ChatGPT-4o achieved an overall concordance of 60% with the CPG recommendations, representing fair agreement (κ = 0.392, p = 0.005). Internal text consistency across repeated trials was low, with several responses showing unacceptable similarity levels (