Identifying and measuring manipulative user interfaces at scale on the web

Abstract

Powerful and otherwise trustworthy actors on the web gain from manipulating users and pushing them into making sub-optimal decisions. While prior work has documented examples of such manipulative practices, we lack a systematic understanding of their characteristics and their prevalence on the web. Building up this knowledge can lead to solutions that protect individuals and society from their harms.

In this dissertation, I focus on manipulative practices that manifest in the user interface. I first describe the attributes of manipulative user interfaces. I show that these interfaces engineer users' choice architectures by either modifying the information available to users, or by modifying the set of choices available to users---eliminating and suppressing choices that disadvantage the manipulator.

I then present the core contribution of this dissertation: automated methods that combine web automation and machine learning to identify manipulative interfaces at scale on the web. Using these methods, I conduct three measurements. First, I examine the extent to which content creators fail to disclose their endorsements on social media---misleading users into believing they are viewing unbiased, non-advertising content. Collecting and analyzing a dataset of 500K YouTube videos and 2 million Pinterest pins, I discover that ~90% of these endorsements go undisclosed. Second, I quantify the prevalence of dark patterns on shopping websites. Analyzing data I collected from 11K shopping websites, I discover 1,818 dark patterns on 1,254 websites that mislead, deceive, or coerce users into making more purchases or disclosing more information than they would otherwise. Finally, I quantify the prevalence of dark patterns and clickbait in political emails. Collecting and analyzing a dataset of over 100K emails from U.S. political campaigns and organizations in the 2020 election cycle, I find that that ~40% of emails sent by the median campaign/organization contain these manipulative interfaces. I conclude with how the lessons learned from these measurements can be used to build technical defenses and to lay out policy recommendations to mitigate the spread of these interfaces.