What question did this study set out to answer?

The research aims to assess how well large language models respond to common patient questions in restorative dentistry regarding accuracy, empathy, and readability.

April 12, 2026Open Access

From algorithms to empathy: can large language models effectively answer patients’ questions in restorative dentistry?

Key Points

The research aims to assess how well large language models respond to common patient questions in restorative dentistry regarding accuracy, empathy, and readability.
Compared responses from four large language models: ChatGPT-5, Gemini 2.5 Pro, DeepSeek-V3.2, and Claude Sonnet-4.5.
Evaluated accuracy and empathy using Likert-type scales by experienced evaluators.
Assessed readability with various online indexes such as FRES and FKGL.
Utilized ANOVA/Kruskall-Wallis for group comparisons based on data distribution.
DeepSeek-V3.2 achieved the highest accuracy (4.8 ± 0.5) and empathy scores (1.96 ± 0.2).
Statistically significant differences in accuracy and empathy were found among models (p < 0.001).
DeepSeek-V3.2 demonstrated superior readability with mean FRES values of 60.85 ± 8.2.

Abstract

The aim of this study was to compare the responses given by the large language models (LLMs) ChatGPT-5, Gemini 2.5 Pro, DeepSeek-V3.2, and Claude Sonnet-4.5 in terms of accuracy, empathy, and readability, to frequently asked patient questions about restorative dentistry. Twenty-five open-ended questions were posed in English to ChatGPT-5, Gemini 2.5 Pro, DeepSeek-V3.2-Exp, and Claude Sonnet-4.5 models in individual, independent sessions. Accuracy was scored using a 5-point Likert-type scale and empathy a 3-point Likert-type scale by two experienced evaluators, and inter-rater reliability was calculated using the Intraclass Correlation Coefficient (ICC). Readability was evaluated through an online platform using the FRES, FKGL, GFI, SMOG, and CLI indexes. Group comparisons were performed using ANOVA/Kruskall-Wallis and post-hoc Dunn-Bonferroni tests according to the normality distribution of the data. The inter-rater reliability was found to be high score for accuracy and showed variability according to the model for empathy. Statistically significant differences were determined between the models in terms of accuracy (p < 0.001), with the highest mean value obtained by DeepSeek-V3.2 (4.8 ± 0.5), followed by Claude Sonnet-4.5 (4.28 ± 0.61) and ChatGPT-5 (4.12 ± 0.67), and the lowest accuracy was determined for Gemini 2.5 Pro (3.6 ± 0.5). Statistically significant differences were determined between the models in terms of empathy (p < 0.001), with the highest mean value obtained by DeepSeek-V3.2 (1.96 ± 0.2). The readability measurements showed that overall, DeepSeek-V3.2 produced more readable texts with higher FRES values (60.85 ± 8.2) (p < 0.001). Within the scope of this study, DeepSeek-V3.2 exhibited a comparatively more balanced performance profile across accuracy, empathy, and readability measures. While Claude Sonnet-4.5 and ChatGPT-5 showed high accuracy, the results for empathy and readability criteria were variable. No model can replace a clinical specialist, and these systems should be evaluated as decision-making tool under specialist supervision, especially at the stages of patient education and first providing information.

Bookmark

View Full Paper