WHO DOES THE ALGORITHM SEE? DEMOGRAPHIC DIVERSITY AND EXTERNAL VALIDATION OF ARTIFICIAL INTELLIGENCE MODELS APPLIED TO FACIAL AND SMILE ESTHETIC ANALYSIS: A METHODOLOGICAL SYSTEMATIC REVIEW

Abstract

This article aimed to assess the reporting of demographic diversity and the performance of external validation in studies on artificial intelligence (AI) models applied to facial and smile esthetic analysis. This is a methodological systematic review, with searches of PubMed/MEDLINE, Europe PMC, OpenAlex and SciELO in October 2026, reported according to PRISMA 2020 and restricted to studies with freely available full texts. Of the 264 unique records, 14 studies were included, published between 2022 and 2026 and originating from 12 countries. Participants' sex was reported in 78.6% of the studies, age in 57.1% and ethnicity or ancestry in 35.7%; among these, three of five used samples from a single population. None of the eight models developed by the authors underwent external validation, and only one of the six ready-made tools and language models had its training population described. No study assessed Brazilian patients. AI models for esthetic analysis are developed and tested in poorly described and poorly diverse populations, which limits their generalizability to admixed countries such as Brazil and reinforces the need for standardized reporting and external validation before clinical use.