Related Experiment Video
Updated: Jun 10, 2026
The bm12 Inducible Model of Systemic Lupus Erythematosus (SLE) in C57BL/6 Mice
Published on: November 1, 2015
Comparing the Accuracy of ChatGPT-4o, DeepSeek-V3, and Gemini 2.5 Flash in Answering Frequently Asked Questions About
Alvina Widhani1,2, Suzy Maria1,2, Aisha Putri Chairani1
1Division of Allergy and Clinical Immunology, Department of Internal Medicine, Faculty of Medicine, Universitas Indonesia, Diponegoro No. 71 Jakarta, Jakarta, 10430, Indonesia, 62 622-1390-4546.
Background:
Systemic lupus erythematosus (SLE) is a complex, fluctuating disease, creating a continuous need for reliable patient information. A prior study concluded that patients with SLE often turn to the internet, including artificial intelligence (AI) chatbots, for information regarding SLE. The rise of AI chatbots as a primary information source presents a critical challenge regarding the accuracy of the information they provide.
Objective:
This study aimed to evaluate the performance of the latest generation of AI chatbots (ChatGPT-4o, DeepSeek-V3, and Gemini 2.5 Flash) in answering frequently asked questions about SLE.
Methods:
Twenty-two frequently asked questions about SLE in Bahasa Indonesia (the Indonesian language) were posed to each chatbot. Responses were independently and blindly evaluated for accuracy by 5 clinical immunologists using a 4-point Likert scale. Readability was assessed using the Flesch reading ease score formula. Statistical comparisons for accuracy and readability were performed using repeated-measures ANOVA or the Friedman test, followed by the Bonferroni test for pairwise comparisons. The Spearman ρ was used to evaluate correlations among accuracy, readability, and word count.
Results:
Gemini 2.5 Flash demonstrated the highest accuracy, with a mean score of 1.25 (SD 0.53), significantly outperforming ChatGPT-4o (mean 1.71, SD 0.61; P<.001). Gemini 2.5 Flash significantly outperformed ChatGPT-4o in 2 evaluated domains. The interreliability analysis revealed a statistically significant level of agreement among the 5 evaluators across all responses (Kendall W=0.389; P<.001). Readability for all 3 chatbots was low (median Flesch reading ease score 42.22-46.66). Gemini 2.5 Flash produced the longest responses (8509 total words), followed by DeepSeek-V3 (5410 words) and ChatGPT-4o (3632 words). A significant negative correlation was found between word count and lower accuracy (ρ=-0.401; P=.001).
Conclusions:
Our study found that ChatGPT-4o, DeepSeek-V3, and Gemini 2.5 Flash provided overall satisfactory responses to SLE-related questions. The highest accuracy was demonstrated by Gemini 2.5 Flash; however, the absolute differences in scores among the 3 AI chatbots were relatively small. All 3 AI chatbots demonstrated low readability, which may limit accessibility for patient use. This finding highlights a critical "blind spot" in which clinical accuracy, as rated by experts, does not equate to patient accessibility. Thus, further research is required to develop more comprehensive evaluation frameworks incorporating safety, factuality, and calibration of AI chatbots across different medical fields and topics.
