Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Disrupting Networks: Amplifying Social Dissensus via Opinion Perturbation and Large Language Models

Erica Coppolillo, Giuseppe Manco

TL;DR

The paper investigates how targeted content injection can intentionally disrupt online opinion dynamics by extending the Friedkin-Johnsen framework to include negative influence and node-level susceptibility. It establishes theoretical guarantees showing basic FJ cannot increase disruption at equilibrium, while the enriched model enables configurations where disruption is maximized, and further enhanced by perturbing innate opinions. To operationalize these insights, the authors design an RL-guided framework that fine-tunes a Large Language Model to generate disruption-oriented text, and validate the approach on synthetic and real-world networks, with the generated content approaching the theoretical disruption upper bounds. The work raises important ethical considerations for content moderation and regulatory policies and outlines future directions including dynamic networks, multi-agent interactions, and defenses against adversarial AI-driven influence.

Abstract

We study how targeted content injection can strategically disrupt social networks. Using the Friedkin-Johnsen (FJ) model, we utilize a measure of social dissensus and show that (i) simple FJ variants cannot significantly perturb the network, (ii) extending the model enables valid graph structures where disruption at equilibrium exceeds the initial state, and (iii) altering an individual's inherent opinion can maximize disruption. Building on these insights, we design a reinforcement learning framework to fine-tune a Large Language Model (LLM) for generating disruption-oriented text. Experiments on synthetic and real-world data confirm that tuned LLMs can approach theoretical disruption limits. Our findings raise important considerations for content moderation, adversarial information campaigns, and generative model regulation.

Disrupting Networks: Amplifying Social Dissensus via Opinion Perturbation and Large Language Models

TL;DR

The paper investigates how targeted content injection can intentionally disrupt online opinion dynamics by extending the Friedkin-Johnsen framework to include negative influence and node-level susceptibility. It establishes theoretical guarantees showing basic FJ cannot increase disruption at equilibrium, while the enriched model enables configurations where disruption is maximized, and further enhanced by perturbing innate opinions. To operationalize these insights, the authors design an RL-guided framework that fine-tunes a Large Language Model to generate disruption-oriented text, and validate the approach on synthetic and real-world networks, with the generated content approaching the theoretical disruption upper bounds. The work raises important ethical considerations for content moderation and regulatory policies and outlines future directions including dynamic networks, multi-agent interactions, and defenses against adversarial AI-driven influence.

Abstract

We study how targeted content injection can strategically disrupt social networks. Using the Friedkin-Johnsen (FJ) model, we utilize a measure of social dissensus and show that (i) simple FJ variants cannot significantly perturb the network, (ii) extending the model enables valid graph structures where disruption at equilibrium exceeds the initial state, and (iii) altering an individual's inherent opinion can maximize disruption. Building on these insights, we design a reinforcement learning framework to fine-tune a Large Language Model (LLM) for generating disruption-oriented text. Experiments on synthetic and real-world data confirm that tuned LLMs can approach theoretical disruption limits. Our findings raise important considerations for content moderation, adversarial information campaigns, and generative model regulation.

Paper Structure

This paper contains 19 sections, 2 theorems, 23 equations, 8 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Foundational Models of Opinion Dynamics
  4. Polarization and Disagreement in Social Networks
  5. Large Language Models in Social Influence
  6. Theoretical Model
  7. Empirical Framework
  8. Large Language Models
  9. Fine-tuning Procedure
  10. Evaluation
  11. Experimental setup
  12. Synthetic Networks
  13. Real-Wold Networks
  14. Settings
  15. Results
  16. ...and 4 more sections

Key Result

Theorem 1

$I\xspace_{\mathcal{G}\xspace, \mathbf{z}\xspace^*}\leq I\xspace_{\mathcal{G}\xspace, \mathbf{s}\xspace}$ since $Y$ is negative semi-definite.

Figures (8)

  • Figure 1: A simple network which empirically shows that $I\xspace_{\mathcal{G}\xspace,\mathbf{z}\xspace^*} > I\xspace_{\mathcal{G}\xspace, \mathbf{s}\xspace}$. The values of $s$ and $\lambda$ represent the innate opinion and susceptibility of the nodes, respectively, while the signs on the edges indicate the nodes influence ($W$).
  • Figure 2: Opinion distribution of the synthetic graph by varying the Beta parameters $\alpha_1$ and $\beta_2$, from high (left) to low homophily (right). We fix $\alpha_2 = \beta_1 = 15$ for all configurations.
  • Figure 3: Susceptibility distribution on the synthetic network by varying the Beta parameters, from highly skewed ((a)-(b)) to slightly skewed ((c)-(d)) to uniform ((e)-(f.
  • Figure 4: Visualization of the real-world social networks from $\mathbb{X}$ (former Twitter). Nodes colour span from blue ($-1$: "Remain"/"No") to red ($1$: "Leave"/"Yes"), while their size resembles their degree.
  • Figure 5: Visualization of the social graphs of Brexit (upper row) and Italian Referendum (bottom row), where nodes are colored according to the induced (normalized) disruption while their size depends on the given centrality measure.
  • ...and 3 more figures

Theorems & Definitions (2)

  • Theorem 1
  • Theorem 2