Shaping Opinions in Social Networks with Shadow Banning

TL;DR

Shaping Opinions in Social Networks with Shadow Banning presents an optimization-based framework to compute shadow banning policies that target the mean or variance of user opinions in large directed networks. By modeling continuous-time opinion dynamics with latent opinions and a bounded confidence update, and by formulating a greedy linear program that maximizes the instantaneous reward derivative, the approach yields scalable edge-level censorship policies. Experiments on synthetic paths, SBM structures, and real Twitter data show that shadow banning can meaningfully shift mean, modulate polarization, and even appear neutral while biasing outcomes over time. The work underscores both the potential for mitigating harm and the danger of covert manipulation, emphasizing the need for edge-level monitoring, transparency, and safeguards against abuse in content moderation systems.

Abstract

The proliferation of harmful content and misinformation on social networks necessitates content moderation policies to maintain platform health. One such policy is shadow banning, which limits content visibility. The danger of shadow banning is that it can be misused by social media platforms to manipulate opinions. Here we present an optimization based approach to shadow banning that can shape opinions into a desired distribution and scale to large networks. Simulations on real network topologies show that our shadow banning policies can shift opinions and increase or decrease opinion polarization. We find that if one shadow bans with the aim of shifting opinions in a certain direction, the resulting shadow banning policy can appear neutral. This shows the potential for social media platforms to misuse shadow banning without being detected. Our results demonstrate the power and danger of shadow banning for opinion manipulation in social networks.

Figures (15)

• Figure 1: Path network with edges untouched by the initial shadow banning policy for different objectives. The node colors indicate the opinion (lower are blue, higher are red). The direction of the edges indicates the flow of information on the network. The objectives are (top left) no shadow banning, (top right) maximize mean, (bottom left) minimize variance, and (bottom right) maximize variance.
• Figure 2: Opinion distributions and mean shadow ban strength versus time under shadow banning control policies for different objective functions on a path network. For the opinions, the purple region is the 25th to 75th quantiles, and the pink region is the 5th to 95th quantiles. The objectives are (top left) no shadow banning, (top right) maximize mean, (bottom left) minimize variance, and (bottom right) maximize variance.
• Figure 3: Stochastic block model network with edges untouched by the shadow banning policy for different objectives. The node colors indicate the opinion (lower are blue, higher are red). The direction of the edges indicates the flow of information on the network. The objectives are (top left) no shadow banning, (top right) maximize mean, (bottom left) minimize variance, and (bottom right) maximize variance. For the no shadow banning policy, the node colors correspond to opinions at time $t=0$. For the other objectives, the node colors correspond to opinions at time $t=10$.
• Figure 4: Opinion distributions and mean shadow ban strength versus time under shadow banning control policies for different objectives on a stochastic block model network. For the opinions, the purple region is the 25th to 75th quantiles, and the pink region is the 5th to 95th quantiles. The objectives are (top left) no shadow banning, (top right) maximize mean, (bottom left) minimize variance, and (bottom right) maximize variance.
• Figure 5: Bar plots of terminal objective values with (blue) no shadow banning versus (orange) shadow banning for the U.S. election and Gilets Jaunes datasets, with objectives being (left) maximize mean, (middle) minimize variance, and (right) maximize variance. For the variance objectives, the terminal variances are reported. The objective improvements by shadow banning compared to no shadow banning are (by U.S. election and Gilets Jaunes) 9% and 12% for maximizing mean, 7% and 23% for minimizing variance, and 40% and 60% for maximizing variance.