Chatbot conversations beat fact sheets at reducing conspiracy beliefs, study finds

September 5, 2026 at 05:39 PM2 min

This digest was compiled by AI from multiple sources — links to the originals are below.

Researchers at Carnegie Mellon, MIT, and Cornell found that seven-minute conversations with a large language model reduced conspiracy beliefs more effectively than static fact sheets in two online experiments. The studies, involving 1,507 US adults, tested beliefs about the Trump assassination attempt and the killing of UnitedHealthcare CEO Brian Thompson. The effect persisted beyond the immediate event window.

Key Facts

The two experiments involved 472 participants for the Trump assassination attempt and 1,035 for the killing of UnitedHealthcare CEO Brian Thompson.
Participants who chatted with Google Gemini for about seven minutes showed greater reduction in conspiracy beliefs than those who read a static fact sheet.
The models used had training cutoffs before both events, so researchers built a curated fact base into the system prompt.
In the Trump experiment, trust in the official explanation did not increase, while in the Kirk experiment it rose slightly compared to the control chat.

Study Design

Researchers from Carnegie Mellon, MIT, and Cornell recruited US adults through a survey platform and used GPT-4o to filter for participants expressing conspiracy beliefs about the event. Participants were randomly assigned to one of three conditions: a dialogue with Google Gemini, a static fact sheet with source citations, or an irrelevant control chat about cats versus dogs. The Trump experiment kept 472 participants, while the Kirk experiment kept 1,035. Both events occurred after the training cutoff of the models used, so the researchers built a curated fact base directly into the system prompt, split into confirmed facts, debunked claims, and open questions.

Dialogue Effectiveness

The conversations averaged about seven minutes and reduced belief in the participant's own conspiracy theory in both experiments, against both the control condition and the fact sheet. Agreement with statements about a 'cover-up or conspiracy' and 'hidden or undisclosed factors' dropped as well. In the Kirk experiment, where authorities had already shared details about the perpetrator, trust in the official explanation rose slightly compared to the control chat, but the difference compared to the fact sheet was not significant. The Kirk dialogue had no measurable effect on support for political violence.

Model Adaptation

To understand how the model persuaded people, the researchers broke its responses into individual sentences and found the model clearly adapted to the available evidence. For the Trump assassination attempt, the model leaned on epistemic humility, source criticism, and Socratic questioning. For classic conspiracy theories, it relied more on factual arguments.