To assess, in an exploratory manner, the performance of ChatGPT-5, a general-purpose multimodal large language model (LLM), in estimating dental age (DA) from panoramic radiographs (PRs), and to compare its estimates with those obtained using the Demirjian and Willems methods. A total of 900 PRs from children aged 5.00–13.99 years, stratified into nine balanced age/sex bands, were evaluated. The Demirjian and Willems DA were calculated via blinded stage scoring, while ChatGPT-5 produced DA estimates directly from the images. Performance was evaluated using mean absolute error (MAE), mean error (ME), mean squared error (MSE), root mean square error (RMSE), and Bland–Altman analysis. Age-group classification performance was also assessed using accuracy, precision, recall, and F1-score. Non-parametric tests were used for between-method comparisons. The overall MAE was lowest for the Willems method (0.63 years), followed by the Demirjian method (0.71 years) and ChatGPT-5 (0.80 years) (p < 0.001). The Willems method significantly outperformed both other approaches. Bland–Altman analysis showed the smallest bias and narrowest limits of agreement for the Willems method, whereas ChatGPT-5 demonstrated wider variability. Although the Willems method showed the highest overall accuracy, ChatGPT-5 demonstrated competitive performance in the 6 to 7-year age groups (Groups 2 and 3). However, the model showed significantly higher error rates in older age groups, particularly in Group 9. ChatGPT-5 showed exploratory potential for DA estimation, particularly in younger children; however, its overall performance remained inferior to that of the validated Willems and Demirjian methods. At present, general-purpose multimodal LLMs should be regarded as adjunctive decision-support tools rather than as replacements for established manual methods.
Çelik et al. (Thu,) studied this question.