Impact of Digital Advertising Policy on Harmful Product Promotion: Natural Language Processing Analysis of Skin-Lightening Ads

Introduction: Starting June 30, 2022, Google implemented its revised Inappropriate Content Advertising Policy, targeting discriminatory skin-lightening ads that suggest superiority of certain skin shades. This study evaluates the ad content changes from 2 weeks before to 2 weeks after the Methods: Text ads from Google searches in eight countries (Bahamas, Germany, India, Malaysia, Mexico, South Africa, United Arab Emirates, and United States) were collected in 2022, totaling 1,974 prepolicy and 3,262 post-policy ads, and analyzed in 2023. A gold standard database was established by two coders who labeled 707 ads, which trained five natural language processing models to label the ads, covering content and target demographics. The descriptive statistics and multivariable logistic models were applied to analyze content before versus after policy implementation, both globally and by country. Results: Vertex AI emerged as the best natural language processing model with the highest F1 score of 0.87. There were significant decreases from pre- to post-policy implementation in the prevalence of labels of "Racial or Ethnic Identification" and "Ingredients: Natural" by 47% and 66%, respectively. Notable differences were identified from pre- to post-policy implementation in India, Conclusions: The study observed changes in skin-lightening product advertisement labels from pre- to post-policy implementation, both globally and within countries. Considering the influence of digital advertising on colorist norms, assessing digital ad policy changes is crucial for public health surveillance. This study presents a computational method to help monitor digital platform policies for consumer product advertisements that affect public health. Am J Prev Med 2024;67(6):811-819. (c) 2024 Elsevier Inc. All rights are reserved, including those for text and data mining, AI training, and similar technologies.