From Skepticism to Acceptance: Simulating the Attitude Dynamics Toward Fake News

1 Introduction

• Prior micro-level models numerically represent opinions and messages, which can miss linguistic nuance, reasoning, and personality-shaped differences in reactions.Earlier macro-level models tracked infected-population totals without detailed analysis of attitude dynamics.Source: §1 · p. 1
• FPS represents each individual as an LLM agent, enabling simulation of varied personas, complex reasoning, and dynamic opinion shifts.The authors describe the framework as offering semantic material for both macro- and micro-level study.Source: §1 · pp. 1–2
• Agents exchange information daily, use short- and long-term memory, and reflect on interactions; an official agent supports interventions and population-level tracking.The framework also tracks individual viewpoints as they evolve.Source: §1 · pp. 1–2
• Political fake news propagates notably faster than the listed other topics, and specific agent traits are associated with greater susceptibility to believing it.The reported topic comparison includes terrorism, natural disasters, science, urban legends, and financial information.Source: §1 · p. 2
• Early and consistent corrections work best to maintain low fake-news propagation; addressing misinformation once is not enough.The authors present this as guidance for timely information management and policy formulation.Source: §1 · p. 2

Fake News Detection

• Earlier detection work considered available news text, while later work modeled finer-grained reasoning; this paper instead addresses propagation after detection [Jin et al., 2022].The passage identifies detection as an essential step in combating misinformation [Guo et al., 2021].Source: §2 · p. 2

Fake News Propagation Modeling

• Propagation models include macro-level epidemic models, stochastic point processes, and individual numerical-threshold diffusion models.FPS is presented as a textual alternative to traditional numerical calculations.Source: §2 · p. 2

LLM-based Agents for Social Simulation

• LLM-agent research has simulated social media and generated human-indistinguishable social-media content [T¨ornberg et al., 2023; Park et al., 2022].The cited work also includes social simulations in digital environments [Park et al., 2023; Kaiya et al., 2023; Li et al., 2023].Source: §2 · p. 2
• The authors describe LLM-based fake-news simulation as a pioneering effort in this field.Figure 2 depicts the framework's feedback loop between dynamic opinion agents and an agent interaction simulator.Source: §2 · p. 3, Figure 2 · p. 3

3.1 Problem Formulation

• The problem formulation initializes N agents with personas and initial attitudes; each day they interact with c peers, reflect, update beliefs, and repeat for T days to track individual and population trajectories.The resulting trajectory summarizes the dynamics of fake news within the network.Source: §3.1 · p. 3

3.2 Our Simulation Framework

• FPS combines a Dynamic Opinion Agent for LLM-based decisions with an Agent Interaction Simulator that plans encounters, distributes official announcements, and updates belief states daily.Agent roles include attributes such as education level, gender, and personality traits.Source: §3.2 · p. 3

3.3 Dynamic Opinion Agent

• Agents use persona-informed reasoning and daily peer interactions; short-term memory summarizes current conversations, while long-term memory integrates them with prior history.Memory updates balance retaining historical context with managing daily interaction volume; the short-term memory is cleared at day’s end.Source: §3.3 · pp. 3–4

3.4 Agent Interaction Simulator

• The modified SIR model tracks susceptible, infected, and recovered agents, with recovered agents able to become infected again, unlike standard SIR.Replacing the recovered label with susceptible makes the model equivalent to SIS.Source: §3.4 · p. 4
• In the SIS equations, transmission and recovery rates determine changes in susceptible and infectious populations over time.The passage defines β as transmission probability per contact and γ as the recovery rate.Source: §3.4 · p. 4
• An official spokesperson issues refutations to agents on designated days, and the simulation monitors their influence on beliefs.The AIS distributes the official information as part of the intervention mechanism.Source: §3.4 · p. 4

4.1 Implementation Details

• Agents use Big Five traits with positive or negative versions sampled at 50% probability; the trait comparison selects high versus low agreeableness and neuroticism.The implementation uses gpt-3.5-turbo-1106 and Mesa.Source: §4.1 · pp. 4–5, §4.3 · p. 5

4.2 Metrics

• The study tracks susceptible, infected, and recovered trajectories, plus metrics for belief level and diversity, infection and recovery, peak infection, and time to half infection.The fitted β and γ parameters describe transmission and recovery dynamics.Source: §4.2 · p. 5

4.3 Macro-level Observation

• Political fake news peaks in four days, while science infections fluctuate around 10; politics has twice the fitted β and one-tenth the fitted γ.The political group also has a higher belief average, less variance, a larger infection rate, and zero recovery rate.Source: §4.3 · p. 5, Figure 3 · p. 6
• Credulous-trait infections rise faster, while skeptical-group recoveries trend slightly upward and credulous-group recoveries remain around 2.The comparison defines credulous traits as high agreeableness and neuroticism, and skeptical traits as low agreeableness and neuroticism.Source: §4.3 · p. 5, Figure 3 · p. 6

4.4 Micro-level Observation

• Michael frequently shifts his views through reasoned decisions, whereas Sandra seldom changes her opinion during the 15-day case study.Michael's shifts span periods of 3 to 14 days; Sandra remains unchanged despite recording differing viewpoints.Source: §4.4 · pp. 5–6

5.1 Fake News Intervention

• Day-one intervention initially reduces believers but may be followed by renewed belief; day-seven intervention yields a steadier decline but leaves many already affected.The intervention-strategy comparison reports that early and reasonably frequent regulation can significantly reduce propagation and influence.Source: §5.1 · p. 6, Figure 4 · p. 7
• Daily and every-three-day official news show no significant difference, supporting interval releases at reasonable cost without compromising control.Source: §5.1 · p. 7
• About 50% of infected agents remain infected throughout; this group shares high agreeableness despite variation in age and education.The authors suggest interventions may be more effective if tailored to these traits.Source: §5.1 · p. 7

5.2 Ablation Study

• Without long-term memory, short-term interactions alone are insufficient to persuade agents to change opinions, and effective interactions do not emerge.Source: §5.2 · p. 7
• Without short-term memory, opinion diversity is three-quarters of the full model's score; long-term memory instead accumulates daily opinions without deep analysis.Source: §5.2 · p. 7
• Removing reasoning has the least effect on overall performance, but generated opinions become descriptions of received views rather than thoughtful reasoning or personal opinions.Source: §5.2 · p. 7

• FPS is an LLM-based fake-news simulation framework incorporating short- and long-term memory and reasoning processes that mimic human cognition.Source: §6 · p. 7
• Its micro-level observations align with previous studies on fake-news topics and traits, while its macro-level results correspond with earlier numerical simulations.Source: §6 · p. 7
• FPS includes an intervention mechanism to curb fake-news propagation and provide practical strategies for real-world monitoring.Source: §6 · pp. 7–8