Overstating Attitudes, Ignoring Networks: LLM Biases in Simulating Misinformation Susceptibility

1. Problem Statement

Large Language Models (LLMs) are increasingly utilized in computational social science as proxies for human judgment, particularly in simulating social networks and misinformation dynamics. However, it remains unclear whether LLMs can accurately reproduce the complex empirical relationships between misinformation susceptibility (belief in and willingness to share false claims) and key social predictors.

Prior research suggests human susceptibility is shaped by:

• Individual factors: Demographics, cognitive traits, and attitudes (e.g., trust in science, political leaning).
• Network factors: The composition and structure of personal social networks (e.g., discussion partners, tie strength).

Existing LLM simulation studies often focus on sociodemographic predictors while neglecting network features. This paper investigates whether LLMs can recover the relational structure of misinformation susceptibility or if they introduce systematic biases that distort these dynamics, potentially leading to flawed conclusions in downstream social simulations.

2. Methodology

Datasets (Ground Truth)

The authors utilized three distinct online survey datasets as ground truth, covering different domains and national contexts:

1. Public Health (US, 2023): 486 participants, 5 misinformation items (e.g., vaccines, chronic diseases).
2. Climate Change (US, 2025): 377 participants, 3 image-based memes converted to text.
3. Pandemic Politics (South Korea, 2020): 708 participants, 5 politically charged items (e.g., election fraud, mask shortages).

Each dataset included four categories of variables:

• Misinformation Susceptibility: Belief ratings and sharing intentions.
• Personal Network Measures: Ego-centric network data (discussants, alter attributes, ties).
• Demographics: Age, gender, race, education, income.
• Attitudinal/Behavioral: Political leaning, trust in science, media use, health literacy.

Experimental Procedure

• LLM Simulation: The authors prompted various LLMs (including reasoning models like GPT-5, DeepSeek-V3, and chat models) to impersonate survey respondents.
• Input: Structured profiles containing the participant's demographics, attitudes, behaviors, and personal network information were fed into the models.
• Task: Models were asked to rate the accuracy of specific false claims and their willingness to share them, mimicking the human survey response format.
• Robustness Checks: The study tested different prompt configurations, including:
  • Alternative ordering of profile blocks (network first vs. last).
  • Composite scores vs. raw item-level data.
  • Chain-of-Thought (CoT) reasoning elicitation.

Analytic Pipeline

The study employed a six-step analysis to compare LLM outputs against human data:

1. Distributional Similarity: Measured using Jensen–Shannon Divergence (JSD) and Earth Mover's Distance (EMD).
2. Rank Correlation: Spearman's $\rho$ between human and simulated responses.
3. Belief-Sharing Coupling: Analyzed the correlation between belief and sharing in both datasets.
4. Predictive Modeling: Fitted Elastic Net regression models to predict susceptibility from profile features.
   • Compared $R^2$ (explained variance) between human and LLM models.
   • Performed ablation studies (removing feature blocks) to assess feature importance.
5. Interaction Analysis: Pooled human and LLM data to test if the effect of specific predictors differed significantly between the two groups.
6. Root Cause Analysis:
   • Reasoning Traces: Analyzed CoT outputs to see which variables models invoked.
   • Training Corpus: Used OLMoTrace to search open-source training data for co-occurrences between predictors and misinformation concepts.

3. Key Contributions

• Systematic Evaluation of Network Features: Unlike previous studies focusing on demographics, this work explicitly tests LLMs' ability to simulate the impact of personal network structures on misinformation susceptibility.
• Identification of Structural Biases: The paper demonstrates that LLMs do not merely add noise to human data but introduce systematic, patterned distortions, specifically over-relying on attitudinal factors and ignoring network context.
• Methodological Framework: Proposes a comprehensive pipeline for diagnosing LLM biases in social simulations, combining distributional metrics, regression analysis, and training data tracing.

4. Key Results

RQ1: Distributional Alignment

• LLMs capture broad distributional tendencies and show moderate correlation with human responses.
• However, no single model or prompt configuration consistently outperformed others. The alignment is sufficient for coarse-grained trends but insufficient for precise individual-level prediction.

RQ2: Recovery of Association Structures

• Exaggerated Belief-Sharing Coupling: In human data, belief and sharing are correlated but distinct ($\rho \approx 0.34–0.61$). In LLM simulations, this correlation is near-monotonic ($\rho \approx 0.75–0.96$), suggesting LLMs fail to capture the nuance that people often share content without believing it (e.g., for entertainment or signaling).
• Inflated Explained Variance: Regression models on LLM data yielded substantially higher $R^2$ values (often an order of magnitude higher) than human models. This indicates LLM responses are overly deterministic and tightly coupled to input profiles, lacking the stochastic noise inherent in human behavior.
• Ignoring Networks:
  • Human Data: Both attitudinal/behavioral and network features significantly predicted susceptibility.
  • LLM Data: Attitudinal/behavioral features dominated prediction. Removing network features from the model caused little to no drop in predictive power for LLMs, whereas it significantly reduced performance for human data.

RQ3: Systematic Feature Effects

• Salient vs. Relational Predictors: LLMs placed disproportionate weight on well-documented, salient individual predictors (e.g., trust in science, political leaning, social media use).
• Underweighting Networks: The estimated effects of network characteristics (e.g., network density, alter attributes) were significantly weaker in LLM simulations compared to human data.
• Root Causes:
  • Reasoning Traces: LLMs frequently invoked attitudinal variables in their CoT reasoning but rarely mentioned network attributes.
  • Training Data: Analysis of the training corpus revealed that attitudinal variables co-occur with misinformation concepts in directional, non-uniform patterns, whereas network variables are less explicitly linked to misinformation in the text data.

5. Significance and Implications

• Limitations as Proxies: LLM-based survey simulations are not interchangeable with human data, particularly when the research question involves social context, peer influence, or network diffusion.
• Risk of Over-Structuring: Relying on LLMs may lead researchers to conclude that misinformation susceptibility is primarily driven by individual attitudes, while underestimating the role of social networks and structural factors.
• Diagnostic Utility: While poor substitutes for human data in predictive modeling, LLMs are valuable for diagnosing systematic deviations and identifying where current models fail to capture complex social dynamics.
• Future Directions: The findings suggest a need for architectural interventions (e.g., Graph Neural Networks, explicit relational inductive biases) rather than just prompt engineering to better model relational social phenomena.

Conclusion: The paper concludes that while LLMs can approximate the "what" of human judgment (broad distributions), they fail to capture the "how" (the structural relationships and stochastic nature of social behavior), specifically by overstating the role of individual attitudes and ignoring the critical influence of personal networks.