Evaluating ChatGPT's Reliability in Analyzing Facial Asymmetry: A Comparative Study With Human Evaluators Using Cross Sectional Static Images.

OBJECTIVE

Facial asymmetry assessment is often subjective and time intensive, thus we aim to evaluate reliability of ChatGPT in analyzing facial asymmetry compared to human raters.

METHODS

Thirty patients with unilateral facial paralysis who underwent facial reanimation surgery were included in this study. Sixty static 2D frontal images (pre and postoperative) were obtained from our database. Facial asymmetry was assessed using the Sunnybrook resting symmetry scale and a 0-4 global asymmetry rating scale. Two human raters evaluated all images independently. ChatGPT Pro 5.0 accessed from September 2025 to October 2025, was prompted through standardized instructions to evaluate the same set. Agreement was assessed using intraclass correlation coefficient (ICC), Cohen's kappa, the Wilcoxon signed-rank test, and Bland-Altman plots.

RESULTS

ChatGPT Pro 5.0 showed no statistically significant difference compared with human raters in preoperative facial assessment (p = 0.701), indicating high reliability in detecting pronounced asymmetry. However, a significant difference emerged in the postoperative assessment (p = 0.0001), where the ChatGPT was less sensitive to subtle facial asymmetry in postoperative stage. Cluster analysis confirmed agreement in high asymmetry cases and divergence in mild cases.

CONCLUSION

ChatGPT reliably assesses pronounced facial asymmetry but is less accurate with subtle facial asymmetry features.