Multimodal Data Privacy Challenges in Chairside Generative AI

To the Editor, We deeply appreciate the transparency and informed consent principles articulated by Prof. Anand and Prof. Vaderhobli in “The Ethical Boundaries of Chairside Generative Artificial Intelligence in Dental Education” [1]. However, as chairside AI systems increasingly integrate multidimensional data streams, we believe the challenges to patient privacy protection in today's digital dental practice environment are far more complex than traditional HIPAA compliance, particularly when involving dental three-dimensional data, warranting deeper discussion. Current commercial chairside AI systems have achieved multimodal data integration. These systems no longer operate on single data types but integrate imaging data, clinical records, biometric information, and even patient behavioral data [2-4]. The unique nature of dental three-dimensional scanning data further compounds this challenge. Modern chairside AI integrated systems can fuse intraoral scanning, CBCT data, occlusal relationship analysis, and facial scanning information [5]. While the original paper correctly identifies that GenAI tools used without Business Associate Agreements may lead to inappropriate disclosure of protected health information, in contemporary dental school clinics, the data complexity students encounter extends far beyond traditional medical records. When students consult AI chairside following the workflow suggested in the original protocol, they often need to input or upload multimodal data containing high-precision biometric information. These datasets encompass combinations of highly accurate personal biometric features that possess not only identity-identifying attributes but can also reflect patients' health status, lifestyle habits, and even socioeconomic status. Building upon the “structured supervision” framework proposed in the original article, we suggest that dental schools should establish tiered privacy protection protocols specifically for multimodal data. First, expanding on the original emphasis of “disclosing GenAI usage intentions to patients,” institutions should clearly inform patients which aspects of their multidimensional biometric data will be used for AI consultation, including imaging characteristics, geometric measurements, and associated health information. Patients should have the right to selectively consent to the use of different data types. Second, regarding the “prompt engineering” training recommended in the original paper, we believe it should specifically include “multimodal data de-identification” skills. Students need to learn how to construct effective AI queries without revealing patient-specific biometric features. In conclusion, Prof. Anand and Prof. Vaderhobli have established an important foundation for the ethical application of GenAI in dental education. In today's era of rapidly advancing multimodal AI technology, we must recognize that patients' multidimensional biometric data represents a special category of privacy asset requiring more stringent and nuanced protection measures than traditional medical information. Only through such approaches can we safely embrace AI technological innovation. The authors have nothing to report. The authors declare no conflicts of interest.