SPILLOVER study: Cyberbullying spreads like a contagion on social media

Researchers led by Arslan Bisharat and colleagues from several universities have released the SPILLOVER study, a deep dive into how cyberbullying (CB) propagates across social media comments. Using data from 32,754 consecutive comment pairs across 430 Instagram sessions, they found that a preceding CB comment significantly increases the likelihood of the next comment being hostile—an effect confirmed by session fixed-effects controls. The spread is primarily driven by cross-user influence, not repeated behavior from the same user. Notably, CB-to-CB pairs show higher textual similarity than non-CB-to-CB pairs, and even non-aggressive replies grow more negative as the preceding CB becomes more severe, suggesting a subtle emotional contagion below overt aggression.

The findings were replicated across three independent datasets—Reddit, Wikipedia Detox, and the SOCC corpus—with spillover rates varying by platform visibility design. The most striking practical insight: a simple binary feature (whether the prior comment was CB) improves prediction accuracy over both session-level baselines and a fine-tuned HateBERT classifier. This means platforms can deploy an ultra-lightweight signal to intercept bullying chains in real time, targeting the spreading pattern rather than individual offenders. The study offers a concrete, scalable anti-bullying tool at a time when social media moderation remains a pressing challenge.