EVENT DOSSIER
Researchers investigated whether large language models (LLMs) exhibited conspiracy theory tendencies. The team tested the models using validated psychological measurement questionnaires and employed various prompts and conditional strategies. The results showed that LLMs had some consensus on certain conspiracy theory elements, and conditional testing combined with sociodemographic attributes resulted in uneven outcomes, revealing potential group biases. Additionally, targeted prompts could easily steer model responses towards conspiracy theories, highlighting the risk of LLMs being manipulated and their potential hazards when deployed in sensitive scenarios.
研究人员调查了大型语言模型（LLMs）是否表现出阴谋论倾向、是否存在社会人口学偏差以及其被诱导采纳阴谋论视角的难易程度。研究团队向多模型施测经验证的心理测量问卷，并采用不同提示与条件化策略进行测试。结果显示，LLMs 在部分阴谋论要素上存在部分共识，且结合社会人口学属性进行条件化会产生不均匀效果，暴露出潜在的群体偏差；同时，针对性提示可轻易将模型响应导向阴谋论方向，凸显了 LLMs 易受操纵的风险及其在敏感场景部署的潜在危害。