Accuracy and Reproducibility of Different Artificial Intelligence Chatbots' Responses to Patient-Based Vitreoretinal Questions: A Comparative Study.

BACKGROUND

Generative artificial intelligence (AI) chatbots are increasingly used by patients and their reliability in complex ophthalmic conditions remains uncertain. This study aimed to compare the accuracy, comprehensiveness, and reproducibility of five AI chatbots-ChatGPT-5.o, DeepSeek R1, Meta AI, Grok 3.0, and Google Gemini 2.5 Pro-in responding to patient-centered vitreoretinal questions.

METHODS

A total of 135 questions covering diabetic retinopathy, floaters/flashes, age-related macular degeneration, retinal tear/detachment, and vitrectomy were sourced from the American Academy of Ophthalmology "Ask an Ophthalmologist" database. Each question was submitted twice to each chatbot under standardized instructions. Two board-certified vitreoretinal ophthalmologists independently graded responses for accuracy and reproducibility. Accuracy was calculated as the proportion of responses graded "Correct and Comprehensive" or "Accurate but incomplete"; reproducibility was defined as agreement between the two responses.

RESULTS

ChatGPT-5.o achieved the highest overall accuracy (94%, n=127/135, 95% CI: 89.9%-98.1%) with a reproducibility rate of 96.3% (n=130/135, 95% CI: 93.1%-99.5%). DeepSeek R1 demonstrated the greatest reproducibility (98.5%, n=133/135, 95% CI: 96.5%-100.0%) and high accuracy (92.6%, n=125/135, 95% CI: 88.1%-97.1%). Meta AI showed 91% (95% CI: 86.1%-95.9%) accuracy and 94% (95% CI: 89.9%-98.1%) reproducibility, whereas Grok 3.0 yielded the lowest accuracy (49.6%, n=67/135, 95% CI: 41.2%-58.0%) despite moderate reproducibility (88.1%, n=119/135, 95% CI: 82.7%-93.5%). Google Gemini 2.5 Pro recorded 72.6% (95% CI: 65.1%-80.1%) accuracy and the lowest reproducibility (77%, 95% CI: 69.9%-84.1%). By category, "Vitrectomy" scored the highest across all chatbots (94%, 95% CI: 87.2%-100.0%), followed by "Macular degeneration" (90%, 95% CI: 85.0%-95.0%). However, the category "Diabetic retinopathy" scored the lowest accuracy rate (64.7%, 95% CI: 52.1%-77.3%).

CONCLUSION

ChatGPT-5.o and DeepSeek R1 approached high accuracy and reproducibility comparable to clinical standards, indicating potential as patient-education tools in vitreoretinal care. However, variability across models and disease categories highlights the need for cautious clinical adoption and continued optimization to ensure safe, reliable information delivery.