Whom to pity, whom to scold? Effects of empathetic and normative AI-assisted interventions on aggressive Reddit users with different activity profiles

During a six-month experiment conducted on Reddit, we studied the impact of counter-speech interventions against personal attacks sent by 440 users regularly attacking others. We used two types of interventions, normative—which referred to social norms, and empathetic—which referred to emotions and encouraged perspective-taking. We employed a collective intelligence approach—the collaboration between human and machine intelligence. Artificial Intelligence was used to detect verbal aggression and notify human volunteers, who then performed the interventions, providing a level of context understanding and realistic human involvement not achieved by potentially automated responses. We analyzed the data from three perspectives. We used time series models of (1) the short-term impact of individual interventions, (2) of the cumulative impact of interventions received as the experiment progressed. We also (3) used aggregated data for a long-term before/after analysis. The short-term effect of interventions is damaging: users tend to be on average around 26% more aggressive the next day, but the effect does not last beyond two days. The cumulative effect of interventions is helpful: each intervention (up to around 8–10 total, the effectiveness of more interventions tends to be lower) decreases daily aggression by 4% on average, and the effects accumulate and balance out the short-term effect in the long run. The effectiveness of normative interventions seems overall higher, except for the less aggressive offenders, for whom empathetic interventions might be equally or more useful.