User-Side Interventions Reduce Harmful Content Exposure in Algorithmic Feeds

Abstract

Recommendation algorithms on social media platforms optimize for user engagement, which can inadvertently amplify exposure to harmful content such as violence, sexual material, and hate speech. Platform-level moderation is often delayed, opaque, and uniform, motivating the need for complementary user-side interventions that allow individuals to reduce unwanted content in their feeds without relying on platform cooperation. Prior work largely relies on single-session or human-subject studies, limiting the ability to capture recursive recommendation feedback loops, control for users’ baseline preferences for harmful content, or systematically compare intervention strategies across harm types. To address these gaps, we propose a sock puppet simulation framework that models 30 rounds of iterative recommendation and interaction. We evaluate two user-side interventions: Downranking and Replacement, on YouTube’s Homepage and Up-Next interfaces, controlling for users’ baseline harm exposure levels (0%, 25%, 50%), yielding 18 experimental conditions with 1,000 puppets each. Our results show that user-side interventions are effective relative to the baseline, with effects concentrated on the Homepage interface. In particular, Downranking emerges as the most robust and consistent strategy, producing statistically significant improvements in both final-state outcomes (net change) and cumulative harmful exposure across baseline preference levels. For example, Downranking reverses baseline increases in harm into significant decreases (e.g., from a +0.6 percentage-point increase to a -0.9 percentage-point reduction), and yields durable reductions in cumulative exposure over time. Replacement shows weaker and less consistent effects on final-state outcomes. We further find no strong evidence of heterogeneous intervention effects across harm types, and observe that the overall reduction in harmful recommendations is largely driven by declines in Physical harm. Our work establishes the long-term efficacy of user-side interventions and provides guidance for their design.