Abstract
As social media usage reaches record highs, there are growing concerns that personalization algorithms can radicalize users by reinforcing their existing beliefs. However, quantitative evidence on how algorithms and user preferences jointly shape harmful online engagement is limited. I conduct an individually randomized experiment with 8 million users of a prominent TikTok-like platform in India, replacing the feed-ranking algorithm with random content delivery. I focus on hateful content targeting minority groups, given its prominence on Indian social media, and establish a trade-off: random post recommendation lowers exposure to harmful (“toxic”) content by 27%, but at a substantial cost to the platform as overall usage falls by 35%. Strikingly, treated users share a larger proportion of the toxic posts they view, mitigating the decline in the number of toxic posts shared. Users with a higher interest in toxic content at baseline drive this result as they reduce overall platform usage, and seek out posts the algorithm does not show them. I collect survey evidence to trace users’ behavior beyond the platform and show that the most affected users switch to other applications. I rationalize these findings with a dynamic model of a revenue-driven algorithm that faces users choosing which posts to consume according to their fixed preferences. Counterfactual simulations evaluate alternative interventions that cannot be implemented in the field, using elasticities estimated from the model. My results underscore the limits of piecemeal algorithmic regulation intended to moderate harmful content online, especially when user behavior is immalleable.
The seminar will be held in room 1249 (12th floor) at Eilert Sundts Hus. The address is Moltke Moes vei 31.