A 7-Minute Chatbot Talk Cut Conspiracy Beliefs More Than a Fact Sheet

Two weeks after the Butler, Pennsylvania shooting, nearly half of Americans had heard the Trump assassination attempt was staged. Eleven percent believed it. Researchers from Carnegie Mellon, MIT, and Cornell saw that window as a test case: could a short conversation with a large language model actually move people off a conspiracy theory in real time?

The answer across two experiments was yes. Participants who spent about seven minutes chatting with Google Gemini showed significantly lower belief in their own conspiracy theory compared to both a control group and people who read a static fact sheet with citations. The effect held for the Trump shooting and for the September 2025 murder of far-right activist Charlie Kirk, where theories about Mossad involvement and false flags spread within days.

Both events happened after the models’ training cutoffs, so the researchers built a curated fact base directly into the system prompt. The model couldn’t rely on internal knowledge. It had to work with confirmed facts, debunked claims, and questions explicitly marked as open. That constraint turned out to be a feature. For the Trump shooting, where the shooter’s motive was still unknown, the model leaned on Socratic questioning and urged caution about jumping to conclusions rather than hammering rational arguments. For Kirk’s murder, where more details were public, it shifted toward emphasizing the societal harms of conspiratorial thinking.

What’s striking is the spillover. Two months later, people who’d had the debunking conversation were less likely to believe another incident on Trump’s property would be covered up. The effect on a subsequent church shooting in Michigan was weaker and didn’t reach significance in the main analysis, but general conspiracy beliefs stayed lower. That suggests the conversation functioned less like a one-off fact check and more like a cognitive vaccine. Whether that protection lasts beyond a few weeks, or whether people simply got better at telling researchers what they wanted to hear, remains the open question.