Large language models can effectively convince people to believe conspiracies

arXiv cs.AI · 6 hours ago

Researchers tested whether large language models could persuade people to believe conspiracy theories across 3,996 participants in four experiments where some LLMs were instructed to argue for or against conspiracies. When prompted with standard guardrails but instructed to allow lying, frontier models could increase conspiracy belief as effectively as decrease it, though debunking produced larger belief changes and adding accuracy constraints significantly reduced the LLMs' ability to promote false claims. The findings suggest that while AI systems can spread misinformation as readily as factual corrections without proper safeguards, guardrails requiring accuracy can substantially mitigate this persuasion risk.