Can artificial intelligence respond as a retina specialist? A comparative evaluation of four platforms in age-related macular degeneration

Introduction: With the increasing use of artificial intelligence in medicine and patients’ growing need to obtain information about conditions such as age-related macular degeneration (AMD), ensuring the quality of responses generated by large language models (LLMs) has become a major challenge. Objective: To evaluate retina specialists’ perceptions of the quality of LLM-generated responses to patient-oriented questions about AMD. Methods: An observational, comparative study was conducted in 2025. Nine retina specialists evaluated responses generated by ChatGPT-5, ChatGPT-4, Gemini, and Grok to eleven frequently asked questions about AMD. Three dimensions – accuracy, clarity, and satisfaction – were assessed using 7-point Likert scales, along with intention to use (yes/no). Statistical analyses included the Friedman test, paired Wilcoxon tests with Holm adjustment, Cochran’s Q test, and McNemar tests ( P < 0.05). Results: Significant global differences were observed across the three evaluated dimensions ( P < 0.001). ChatGPT-5 consistently achieved the highest scores. In terms of accuracy, ChatGPT-5 outperformed Grok and ChatGPT-4 (Holm-adjusted P < 0.001). Regarding clarity, ChatGPT-5 scored higher than Grok, ChatGPT-4, and Gemini (Holm-adjusted P < 0.005). For satisfaction, ChatGPT-5 outperformed ChatGPT-4 and Grok (Holm-adjusted P < 0.001). Item-level analysis showed that model performance varied according to the specific question. Intention to use did not differ significantly among platforms after adjustment for multiple comparisons (Holm-adjusted P ≥ 0.375). Conclusions: ChatGPT-5 demonstrated the most consistent performance across all evaluated dimensions. The variability observed based on question type underscores the importance of clinical oversight before implementing these tools in patient education for retinal pathology.