Abstract
Large language models (LLMs) have demonstrated persuasive abilities across various contexts. However, it remains unclear whether this persuasive power favors accuracy or if bad actors can easily use LLMs to spread misinformation.
In this study, we investigate this question through four experiments where participants (N = 3996 Americans) discussed a conspiracy theory they were uncertain about with an LLM instructed to either argue against ('debunking') or for ('bunking') that conspiracy. Across several frontier models (with standard guardrails but prompted to allow lying), we found no consistent evidence of a truth advantage: LLMs were able to significantly increase and decrease average conspiracy belief. Participants in the bunking condition rated the LLM as more informative and collaborative and reported greater trust in AI than those in the debunking condition.
More encouragingly, debunking induced larger belief changes, and subsequent corrections reversed the bunking effect. Furthermore, simply prompting the model to provide accurate information dramatically reduced bunking effectiveness, with one powerful frontier model (GPT 5.2) almost entirely refusing to promote conspiracies, suggesting that appropriate guardrails can favor accurate beliefs.
Finally, we observed a stark truth asymmetry in information sharing: debunking had a large positive impact on mock social media posts composed by participants, while bunking had little effect. Overall, our findings indicate that people are not inherently less susceptible to misleading AI than to informative AI, but that potential technical solutions exist to mitigate this risk.
Blogger's Review: This article explores the dual nature of large language models in the propagation of conspiracy theories, revealing the complex relationship between these technologies and the accuracy versus misinformation dichotomy. It emphasizes the importance of appropriate technical measures in guiding beliefs, and future research should continue to focus on optimizing these models to reduce the spread of misleading information.