How Accurate Are Public Perceptions? A Comparative Analysis of Artificial Intelligence Responses on Physical Therapy for Gonarthrosis

Abstract

Purpose: This study aimed to compare the content quality, reliability, readability, and structure of responses from different artificial intelligence (AI) models to the most frequently asked public questions on physical therapy for gonarthrosis. Methods: Ten highly frequent questions were collected from Google Trends and  Google’s “People Also Ask” section in Turkiye (2022–2025) using the lay term “physical therapy in knee osteoarthritis.” Questions were submitted directly to ChatGPT, Google Gemini, and DeepSeek without additional prompts to simulate actual user behavior. Content reliability was assessed using mDISCERN (0–5) and overall quality with the Global Quality Score (GQS, 1–5). Readability was calculated via the Ateşman Turkish Readability Index (AOI), and structural features were analyzed by sentence count. Two independent raters performed evaluations, with disagreements resolved by a third rater (ICC=0.84). Inter-model comparisons were conducted using the Friedman test, followed by Wilcoxon signed-rank tests with Bonferroni correction (p<0.016). Finding: Significant differences were observed in mDISCERN (χ²(2)=9.80, p=0.007; W=0.49) and GQS (χ²(2)=11.10, p=0.004; W=0.56), with Gemini outperforming ChatGPT in reliability (p=0.005) and quality (p=0.006). ChatGPT showed higher readability than Gemini (p=0.005) and DeepSeek (p=0.013). DeepSeek responses were significantly longer (p=0.001). mDISCERN and GQS were strongly correlated (r=0.827, p<0.001), while quality and readability were negatively correlated. Conclusion:AI responses on physical therapy for gonarthrosis differ in quality and readability. Gemini excelled in reliability and overall quality, whereas ChatGPT offered higher readability. Response length did not reflect information quality. These findings emphasize the importance of AI model selection for patient education and clinical communication.