Simulating Misinformation Propagation in Social Networks using Large Language Models

TL;DR

The paper tackles how misinformation propagates in social-like networks by modeling user biases with persona-conditioned LLMs and tracking fidelity with a QA-based auditor. It introduces an auditor-node framework, formalizes the Misinformation Index (MI) and Misinformation Propagation Rate (MPR), and classifies drift into factual errors, lies, and propaganda. Across homogeneous branches, identity- and ideology-driven personas amplify misinformation while expert/neutral personas help preserve factual fidelity; when branches are heterogeneous, misinformation almost universally escalates to propaganda. The framework yields interpretable, domain-aware insights and lays groundwork for using LLM-based proxies to study, predict, and mitigate misinformation diffusion in digital ecosystems.

Abstract

Misinformation on social media thrives on surprise, emotion, and identity-driven reasoning, often amplified through human cognitive biases. To investigate these mechanisms, we model large language model (LLM) personas as synthetic agents that mimic user-level biases, ideological alignments, and trust heuristics. Within this setup, we introduce an auditor--node framework to simulate and analyze how misinformation evolves as it circulates through networks of such agents. News articles are propagated across networks of persona-conditioned LLM nodes, each rewriting received content. A question--answering-based auditor then measures factual fidelity at every step, offering interpretable, claim-level tracking of misinformation drift. We formalize a misinformation index and a misinformation propagation rate to quantify factual degradation across homogeneous and heterogeneous branches of up to 30 sequential rewrites. Experiments with 21 personas across 10 domains reveal that identity- and ideology-based personas act as misinformation accelerators, especially in politics, marketing, and technology. By contrast, expert-driven personas preserve factual stability. Controlled-random branch simulations further show that once early distortions emerge, heterogeneous persona interactions rapidly escalate misinformation to propaganda-level distortion. Our taxonomy of misinformation severity -- spanning factual errors, lies, and propaganda -- connects observed drift to established theories in misinformation studies. These findings demonstrate the dual role of LLMs as both proxies for human-like biases and as auditors capable of tracing information fidelity. The proposed framework provides an interpretable, empirically grounded approach for studying, simulating, and mitigating misinformation diffusion in digital ecosystems.

Figures (5)

• Figure 1: Diagram of system architecture for misinformation propagation, illustrating the flow of information across persona-conditioned LLM nodes, the auditor’s intervention, and data recording modules. This design allows tracing factual drift step-by-step across heterogeneous or homogeneous branches.
• Figure 2: Heatmap visualization of Misinformation Propagation Rates of each homogeneous branch across 21 LLM agents and 10 news domains. The color bar depicts Misinformation Severity from factual errors (green) to lies (orange) to propaganda (red). The bottommost row gives the average MPR over the 21 agents for specific domains, and the rightmost column gives the average MPR over the 10 domains for specific agents.
• Figure 3: Node-level heatmap visualization of misinformation propagation, showing misinformation indices (Eq. \ref{['eq:MI_bk']}) calculated after each rewrite of the given domain by the identical set of agents placed at the 30 nodes of the branch. The branches or agent-domain pairs are chosen from Fig. \ref{['fig:Homo_Heatmap']}, with highest 10 (top) and lowest 10 MPRs (excluding education1; bottom).
• Figure 4: Same as Fig. \ref{['fig:Homo_Heatmap']} but for heterogeneous branches: 30 nodes each with controlled random assignment from 21 agents. Misinformation Propagation Rates (numerical values) and Severity (3 color scales), as well branch-wise and domain-wise average of MPRs are shown.
• Figure 5: Same as Fig. \ref{['fig:Homo_Heatmap_node']} but with a controlled-random set of 21 agents placed at the 30 nodes of each branch.