Researchers from Carnegie Mellon, MIT, and Cornell conducted two experiments to test if conversations with large language models can reduce belief in conspiracy theories. The study focused on events following the July 2024 assassination attempt on Donald Trump and the September 2025 murder of Charlie Kirk.
The study involved 472 participants for the Trump event and 1,035 for the Kirk event. Participants were assigned to either a dialogue with Google Gemini, a static fact sheet, or a control conversation. The models used curated fact bases in their system prompts to address conspiracy narratives.
Conversations lasted an average of seven minutes and proved more effective than static fact sheets at reducing belief in conspiracy theories. The models adapted their tactics based on available evidence, using Socratic questioning when information was scarce and factual arguments when more data existed.
The debunking effect persisted over time, acting as a form of prebunking against future false claims. Participants who engaged with the model were less likely to believe in cover-ups regarding subsequent events. Researchers noted that similar dialogue methods could potentially be used to spread misinformation.