 |
| October 2026 |
 |
78 |
|
| 專輯論文Special Issue Articles |
| LLMs模擬意見動態和偏見信息的影響：基於「氣候變化」、「轉基因食品」和「疫苗安全」議題的多代理研究 |
| Simulating Opinion Dynamics and the Impact of Biased Information Using LLMs: A Multi-Agent Study on Climate Change, GM Foods, and Vaccine Safety |
|
|
|
 |
(4)
|
| 作者 |
陶煒、向安玲 |
| Author |
Wei TAO, Anling XIANG |
| 關鍵詞 |
語言模型、意見動態、偏誤資訊、社會現象 |
| Keywords |
language model, opinion dynamics, biased information, social phenomena |
| 摘要 |
在生成式人工智慧逐步介入資訊生產與傳播的背景下，經典傳播理論中有關有限理性、啟發式加工與意見極化的若干預設，亟需在新型認知主體條件下重新檢視。本研究探討大語言模型（large language models, LLMs）在模擬人類意見動態與科學議題公共傳播中的表現，特別關注偏誤資訊介入後群體科學共識的變化。研究以多個虛擬代理為對象，圍繞「氣候變化」、「轉基因食品」、「疫苗安全」三項具有科學共識的議題展開模擬測試。實驗一在無偏誤資訊條件下，觀察代理經由社交網絡進行日常互動後的信念狀態與傳播路徑；實驗二調整初始信念分佈，以分析初始條件對結果的影響；實驗三則引入10%代理發佈偏誤資訊，以檢驗其對信念動態的干擾。 50個代理於30天內共生成194,699條對話。研究結合依存語義圖譜、 Lotka-Volterra模型與混合效應模型進行分析。結果顯示，在無偏誤資訊情況下， LLM代理整體傾向與科學共識保持一致；即使在不同初始信念分佈下，各代理群體最終仍趨向科學共識，顯示模型具有一定穩定性。引入偏誤資訊後，意見分佈雖更趨多樣，且信念得分隨時間下降，但最終整體仍朝向科學共識發展。研究表明， LLM驅動的多代理系統可作為一種「可控社會模擬」工具，為理解科學共識形成及傳播理論重構提供方法論啟示。 |
| Abstract |
Generative artificial intelligence has introduced new cognitive agents into information production and dissemination, calling for a reexamination of classical communication theories on bounded rationality, heuristic processing, and opinion polarization. This study explores the performance of large language models (LLMs) in simulating human opinion dynamics and public communication on scientific issues. Using multiple virtual agents, we examined discussions on three topics with established scientific consensus: climate change, genetically modified foods, and vaccine safety. Across three experiments, we tested interactions under normal conditions, different initial belief distributions, and the introduction of misleading information by 10% of agents. In total, 50 agents generated 194,699 dialogues over 30 days. Results show that LLM agents generally tended toward scientific consensus. Although misleading information increased opinion diversity and lowered belief scores over time, the overall direction still moved toward consensus. The findings suggest that LLM-driven multi-agent systems can serve as a tool for controllable social simulation and provide insights into the formation of scientific consensus in intelligent media environments.
本文引用格式：
陶煒、向安玲（2026）。〈LLMs模擬意見動態和偏見信息的影響：基於「氣候變化」、「轉基因食品」和「疫苗安全」議題的多代理研究〉。《傳播與社會學刊》，第78期，頁1–45。
Citation of This Article:
Tao, W., & Xiang, A. (2026). Simulating opinion dynamics and the impact of biased information using LLMs: A multi-agent study on climate change, GM foods, and vaccine safety. Communication and Society, 78, 1–45. |
|
|
 |
| No.77 2026 July |
 |
| No.76 2026 April |
 |
| No.75 2026 January |
 |
| No.74 2025 October |
 |
| No.73 2025 July |
|
