Short conversations with Google Gemini reduced belief in emerging conspiracy theories more than reading a static fact sheet in two online experiments, according to research from Carnegie Mellon, MIT and Cornell. The tests began soon after two violent political events, when verified information was still limited and speculation was spreading quickly.
The first experiment included 472 people who expressed conspiracy beliefs about the 2024 assassination attempt on Donald Trump. The second included 1,035 participants after the 2025 murder of Charlie Kirk. Participants were assigned to at least five dialogue rounds with Gemini, a cited fact sheet or an unrelated conversation about pets. The chatbot received a curated set of confirmed facts, debunked claims and explicitly unanswered questions.
Conversations averaged about seven minutes and produced larger reductions in each participant’s stated theory than either comparison. When evidence was scarce, the model relied more on uncertainty, source criticism and questions; with stronger evidence, it used more direct factual arguments. Some participants later showed lower belief in unrelated conspiracy claims, although one follow-up event produced no significant direct effect in the main analysis.
The authors describe the work as a case study, not a universal remedy. People must choose to engage, a real conspiracy could be dismissed incorrectly, and earlier research shows persuasive models can also move beliefs in the opposite direction. Any deployment would need carefully maintained evidence and safeguards against abuse.