On the Harms of Online Ads

Targeted advertising systems use personal data to deliver customized advertisements, raising significant concerns about privacy violations and discriminatory practices. While most existing solutions focus primarily on privacy protection, this approach may overlook the broader social consequences of algorithmic targeting on individuals and communities. In this thesis, we introduce a framework for evaluating targeted advertising systems through distributional harm—disparities in ad exposure across user groups based on irrelevant demographic features—and propose advertising utility as a complementary metric to estimate the alignment of ads with user profile characteristics, thereby providing a foundation for the systematic evaluation of such systems. We implement the harm detection component using web crawling techniques to collect advertisements from profiles generated from real browsing, use clustering algorithms to group similar ads, and analyze their distribution patterns to detect potential bias. We applied our methodology to browsing data from 36 German users, collecting 704 ads across categories such as gambling, fashion, technology, and finance. We find evidence of a significant gender bias in gambling advertisement targeting, with 77% of gambling ads delivered to male users, many of whom had not previously visited gambling-related websites. These preliminary results suggest targeting based on inferred demographics rather than interest. This framework can be used to evaluate both existing and future designs, and would be especially interesting to apply to privacy preserving advertising systems, in order to explore whether privacy interventions alone are sufficient to reduce harm in targeted advertising.