Effects of AI-Generated Counterspeech

Online hate speech poses an urgent challenge in Western democracies and beyond. The consequences for individuals, digital discourse, and society at large are profound. Counterspeech can help mitigate these harms, and AI could make producing counterspeech easier and more widespread. However, counterspeakers themselves worry that AI might undermine perceptions of authenticity, potentially diminishing the impact of their efforts. To date, little is known about how users perceive AI-generated counterspeech, and whether AI-labeled counterspeech affects individuals’ perception of online hate speech incidents and their motivation to intervene – relative to unlabeled counterspeech. Moreover, evidence is lacking on whether the effects vary across different counterspeech forms. Two approaches in particular — social norm appeals and empathetic target support — have been theorized and empirically shown to be effective when deployed by human users, yet their effects under AI-labeling remain untested. In this study, we investigate how different forms of counterspeech labeled as AI-generated and unlabeled (i.e., user-generated) in reaction to online hate speech will affect users’ perceptions of counterspeech as well as their perception of the incident and their own intervention. To test this, we will use an online survey experiment (2 x 2 between participant design with control group).