TLDR: This research paper introduces an agent-based simulation using Large Language Models (LLMs) to model how different populations respond to misinformation. By creating agent personas with varying professions and mental schemas, the study evaluates their reactions to news headlines. The findings show that LLM-generated agents closely align with human predictions and ground-truth labels, validating their use as proxies. Crucially, the research reveals that mental schemas, rather than professional backgrounds, are the primary drivers of how agents interpret misinformation, offering valuable insights for designing targeted interventions against disinformation campaigns.
In today’s interconnected world, disinformation campaigns pose a significant threat, capable of distorting public perception and destabilizing institutions. Understanding how diverse populations react to information is vital for creating effective countermeasures. However, conducting real-world experiments on human populations is often impractical and ethically challenging. To overcome this, a recent study introduces an innovative approach: an agent-based simulation powered by Large Language Models (LLMs) to model responses to misinformation.
The research, titled “Simulating Misinformation Vulnerabilities with Agent Personas,” explores how AI-driven agents, designed with distinct characteristics, interpret and react to news headlines. The core idea is to use LLMs as proxies for human behavior, allowing researchers to study complex information dynamics in a controlled and scalable environment. This method builds on advancements in generative AI and agent-based modeling, offering new opportunities to analyze information operations.
Crafting Diverse Agent Personas
To ensure a broad spectrum of perspectives, the researchers developed eight unique agent personas. These personas were categorized into two main groups: professions and mental schemas. The professional backgrounds included military personnel, college students, retired persons, industrial workers, and financial analysts. These roles were chosen because they are often targets of influence campaigns or represent key demographics in information spread.
Beyond professions, the study introduced agent personas based on mental schemas to model different cognitive responses. These included a “conspiracy-believer” (highly receptive to conspiracy narratives), a “conspiracy-susceptible” agent (prone to misinformation but not fully embedded in conspiracy thinking), and “normal persons” (a neutral baseline news reader). This combination of occupational roles and cognitive frames provides a robust framework for analyzing how various demographics process, propagate, or resist misinformation.
The Simulation Process
The simulation utilized the Misinfo Reaction Frames corpus, a dataset containing 2,132 news headlines across critical public discourse domains like COVID-19 and climate change. Each headline was fact-checked and also evaluated by 63 human annotators, who provided insights into perceived veracity, emotional response, and propensity to share. This human reaction data was crucial for benchmarking the LLM agents.
In the simulation, each agent persona was presented with a headline from the corpus and asked two questions: whether they believed the headline was real and their likelihood of sharing it on a 1-5 Likert scale. The responses from the LLM agents were then compared against the gold standard labels (factual accuracy) and the human annotator judgments. The study employed two different LLMs for agent generation: LLaMA 3.1 8B Instruct and GPT-4, running them with a temperature setting of zero to ensure consistent and reproducible outputs.
Key Findings: Mental Schemas Over Professions
The results revealed several significant insights. GPT-based agents consistently demonstrated an ability to detect misinformation across different professional domains. Interestingly, the “normal” news reader agent, designed to be neutral, showed the highest overlap with human annotators and performed best in identifying misinformation, suggesting that an unbiased approach aligns closely with human decision-making in this context. Overall, six out of eight GPT agents outperformed human annotators in identifying misinformation.
In contrast, the LLaMA model exhibited greater variance in performance and less frequent alignment with human annotations, although five of its eight agents still outperformed human annotators. A crucial finding across both models was that an agent’s interpretation of information was more strongly influenced by its mental schema than by its professional background. While responses were relatively consistent across different professions, significant differences emerged when agents were prompted with varying cognitive predispositions, such as susceptibility to alternative news sources or belief in conspiracy theories. For instance, conspiracy-driven and susceptible agents showed a stronger tendency to classify misinformation as true, aligning with phenomena like confirmation bias.
Regarding sharing behavior, most AI agents generally aligned with human ratings, but agents with significant schema adjustments (like susceptible and conspiracy agents) showed a higher propensity to share information, including misinformation. These agents tended to cluster their sharing likelihood, unlike human annotators whose ratings were more evenly distributed.
Implications for Countering Disinformation
These findings underscore the importance of tailoring misinformation intervention strategies to cognitive and ideological predispositions rather than solely focusing on professional affiliations. Instead of designing interventions based on job titles, strategies might be more effective if they address the underlying cognitive biases that contribute to susceptibility. The research establishes a systematic pipeline for creating agent personas and simulating their responses, offering a controlled environment for studying factors affecting misinformation susceptibility at scale, which is difficult with human studies.
This approach also highlights the potential of LLM-based agent simulations for studying broader information dynamics, including polarization, trust in media, and susceptibility to other forms of deceptive content like deepfakes. From a national security perspective, understanding which population segments are vulnerable to specific messaging types allows for more precise design of targeted counter-disinformation efforts. For more details, you can read the full research paper here.
Also Read:
- Unpacking Self-Awareness in Large Language Models
- AgileThinker: AI Agents Mastering Real-Time Decisions in Dynamic Environments
Looking Ahead
While this study provides a preliminary investigation, future work could explore different LLMs, a wider range of professions and mental schemas, and further refine agent behavior to enhance alignment with human perception. As the digital information landscape continues to evolve, the ability to simulate complex human-like responses with multi-agent frameworks presents unprecedented opportunities to develop robust disinformation defense strategies in an increasingly AI-driven information environment.
