New AI Tracks Harmful Posts on Social Media Before They Spread

Researchers introduced a new diffusion model for signed social networks that accounts for the fact that real-world relationships are not uniformly supportive. The model treats awareness as continuous and non-monotonic, allowing beliefs to weaken or reverse under competing influence—something classical models don't allow. It also factors in differences in vulnerability, prioritizing protection of those most at risk. On top of this, they formulated a Harm Minimization problem and proved it admits a greedy algorithm with a bounded approximation ratio. Across six signed networks, their model was the only one tested that could let awareness reverse after activation, and it achieved the highest harm reduction of any method evaluated.