Abstract

We develop a measurement device that allows eliciting public perceptions and preferences using synthetically generated visual vignettes. In a world where societally relevant human behavior is increasingly taking place on platforms, it is important to be able to assess the effects of exposure to online content in a systematic and controlled fashion. However, exposure to content is selective and content features are not independently distributed. Our approach couples the randomized combination of visual and verbal elements of online content, such as speech, images, or conversational context, with a conjoint design. Content is synthetically composed on the basis of human input, resulting in an arbitrarily high number of treatment states. We present an application to online content moderation preferences, which allows us to causally identify message-, context-, and citizen-level determinants for perceptions of and preferences for action against hate speech. The application uses multi-modal input (text and images) and is fielded in 11 countries (Brazil, Colombia, Germany, India, Indonesia, Nigeria, Philippines, Poland, Turkey, United Kingdom, United States), totalling more than 19k respondents. Our approach is scalable to evaluate public perceptions of any kind of verbal and visual content online, such as misinformation, political speech, and AI-generated content.

Biography

Professor of Data Science and Public Policy

Director of the Data Science Lab