Dynamic Simulation Framework for Disinformation Dissemination and Correction With Social Bots

Boyu Qiao, Kun Li, Wei Zhou, Songlin Hu

TL;DR

This paper introduces MADD, a novel multi-agent simulation framework that models the complex interplay of social bots and humans in the spread and correction of disinformation. By incorporating realistic network structures and dynamic bot behaviors, MADD offers a more robust platform for evaluating the effectiveness of different correction strategies.

What Problem Does This Paper Address?

The digital information landscape is increasingly shaped by a symbiotic relationship between humans and social bots. While social bots can amplify disinformation, they also hold potential for correction. However, existing simulation models often oversimplify user behavior and network structures, failing to capture the dynamic nuances of bot-driven disinformation campaigns and the effectiveness of counter-strategies. This gap hinders our ability to develop effective governance and risk control mechanisms for online information ecosystems.

Methods

MADD tackles this challenge by building a sophisticated multi-agent simulation framework with three core modules:

  • Disinformation Dissemination Modeling: This module constructs a realistic propagation network by integrating the Barabási-Albert Model (BAM) for scale-free network topology and the Stochastic Block Model (SBM) for community structures. It defines five key user/bot attributes (interest community, trust threshold, dissemination tendency, social influence, activation time) derived from real-world user data. The framework incorporates three agent types: regular users, malicious bots (MBots), and legitimate bots (LBots).
  • Correction Strategy: MADD models two primary correction strategies: fact-based correction (direct refutation with evidence) and narrative-based correction (using emotionally engaging stories). It also allows for the simulation of different intervention timings (early, mid, late).
  • Propagation Impact Quantification: The framework uses both individual-level metrics (e.g., dynamic changes in trust thresholds) and group-level metrics (e.g., susceptible, exposed, infected, and uninfected user ratios) to assess the impact of disinformation and correction efforts.

A key innovation is the use of Large Language Models (LLMs) to evaluate user attributes like interest community and trust threshold, and to simulate agent decision-making, enhancing the realism of user behavior.

Datasets

The study utilizes real-world user data from the X/Twitter platform, specifically focusing on six representative vertical communities: Entertainment, Technology, Sports, Business, Politics, and Education. For each community, they collected users' basic profile information and a sample of 200 content posts. This data was used to inform the design of user attributes and network structures. No new benchmarks were introduced; rather, the focus was on validating the model against existing empirical studies.

Key Insights and Ideas

  • Realistic Network Structure Matters: Combining scale-free properties (BAM) with community structures (SBM) creates a more accurate representation of online social networks, crucial for understanding how disinformation spreads.
  • Bots are Double-Edged Swords: The framework explicitly models both malicious bots spreading disinformation and legitimate bots correcting it, highlighting their dual role in the information ecosystem.
  • Early Intervention is Crucial: The simulations strongly suggest that intervening early in the disinformation spread is significantly more effective than mid-stage or late-stage interventions.
  • Fact-Based Correction Often Outperforms Narrative-Based: For many topics, directly refuting disinformation with facts proved more effective than using narratives, though effectiveness varied by community.
  • Politics is a Tricky Domain: The "Politics" community showed more volatile disinformation spread patterns, highlighting the challenges of predicting and countering misinformation in highly debated topics.
  • LLMs Enhance Realism and Efficiency: Using LLMs for attribute evaluation and agent decision-making not only improves simulation accuracy but also offers computational efficiency gains compared to purely LLM-based evaluations.

Key Results

  • MADD's simulated network attributes and structures (interest communities, degree distribution) show consistency with real-world social networks.
  • Disinformation spread in simulations generally follows a pattern observed in real-world studies, with a rapid increase followed by a decline in infected spreaders.
  • Fact-based correction demonstrated significantly higher effectiveness than narrative-based correction in "Business," "Politics," and "Technology" communities during early interventions.
  • In the absence of intervention, user trust thresholds tend to increase over time, indicating some inherent user resilience. However, correction strategies, especially fact-based ones, can further boost this resilience.
  • Early intervention with legitimate bots significantly reduced infection rates, with a 30% increase in legitimate bots leading to a much faster drop in infection rates compared to no intervention.
  • Long-term simulations showed that early fact-based interventions could keep disinformation influence below 5% and significantly increase user trust thresholds.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Bot Ratio Testing: The experiment did not systematically test the differential effects of malicious and legitimate bots across a wide range of ratios.
  • Simulation Duration: The simulation was limited to 72 time steps due to computational constraints, preventing long-term dynamic analysis.
  • Single Community Validation: For computational efficiency, correction strategy evaluation was primarily conducted within single communities, not globally.
  • Potential LLM Bias: The LLM used as a user agent might exhibit topic-specific biases, influencing simulation outcomes.

Future work could involve:

  • Constructing larger-scale network models and extending simulation periods.
  • More comprehensive testing of bot ratios and their impact.
  • Prioritizing more neutral topics to mitigate LLM bias.
  • Developing more advanced LLM-based user agents to better capture complex cognitive processes.

Summary updated

Abstract

In the human-bot symbiotic information ecosystem, social bots play key roles in spreading and correcting disinformation. Understanding their influence is essential for risk control and better governance. However, current studies often rely on simplistic user and network modeling, overlook the dynamic behavior of bots, and lack quantitative evaluation of correction strategies. To fill these gaps, we propose MADD, a Multi Agent based framework for Disinformation Dissemination. MADD constructs a more realistic propagation network by integrating the Barabasi Albert Model for scale free topology and the Stochastic Block Model for community structures, while designing node attributes based on real world user data. Furthermore, MADD incorporates both malicious and legitimate bots, with their controlled dynamic participation allows for quantitative analysis of correction strategies. We evaluate MADD using individual and group level metrics. We experimentally verify the real world consistency of MADD user attributes and network structure, and we simulate the dissemination of six disinformation topics, demonstrating the differential effects of fact based and narrative based correction strategies.
BibTeX
@inproceedings{qiao2025dynamic,
  author = {Boyu Qiao and Kun Li and Wei Zhou and Songlin Hu},
  title = {{Dynamic Simulation Framework for Disinformation Dissemination and Correction With Social Bots}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure