Study: Seven-Minute Chatbot Conversations Reduced Conspiracy Beliefs Better Than Fact Sheets

Researchers at Carnegie Mellon, MIT, and Cornell tested whether short conversations with a large language model could reduce conspiracy beliefs soon after fast-moving real-world crises. The study focused on two events: the July 2024 assassination attempt on Donald Trump and the September 2025 murder of Charlie Kirk. Participants who expressed conspiracy beliefs were assigned to either a Google Gemini dialogue, a static fact sheet with citations, or an unrelated control chat.

The chatbot group had at least five rounds of back-and-forth, with the model instructed to reduce conspiracy beliefs through evidence-based conversation. Because the events happened after the models’ training cutoffs, researchers supplied a curated fact base covering confirmed facts, debunked claims, and open questions.

The chatbot conversations averaged about seven minutes and reduced belief in participants’ own conspiracy theories more than both the control chat and the fact sheet. Agreement with broader claims about a “cover-up or conspiracy” and “hidden or undisclosed factors” also dropped.

The article reports that the model adapted its approach to how much was known. When facts were scarce, it leaned on caution, source criticism, Socratic questioning, and acknowledgment of uncertainty; when more information was available, it relied more on factual arguments and the societal harms of conspiratorial thinking.

The debunking effect appeared to carry into later events. After a later incident involving another armed man on Trump’s property, participants who had taken part in the debunking dialogue were less likely to expect the truth would be hidden from the public or known only to a few powerful people.

For a later shooting and arson attack in Grand Blanc Township, Michigan, the main analysis found no significant direct effect. A secondary analysis suggested some remaining influence on conspiracy narratives and broader conspiracy beliefs.

The findings suggest that short, sourced, interactive AI conversations may help reduce false beliefs when public facts are still developing. That matters because early crisis periods are often when conspiracy narratives spread quickly and official information can be incomplete.

The authors also caution that the work is a case study, not a universal solution. The same persuasive dynamics could be misused to push conspiracy narratives, and the approach depends on people being willing to discuss their beliefs with a language model in the first place.