Interactive AI reduces conspiracy beliefs in two experiments

GPT-4o conversations cut conspiracy beliefs more than fact sheets in two U.S. crisis scenarios

Researchers from Carnegie Mellon, MIT, and Cornell tested whether short conversations with an AI model reduced conspiracy beliefs after two real-world events. In the first experiment (post-July 2024 Trump assassination attempt), 472 U.S. adults who reported conspiracy theories about the event engaged with GPT-4o version 1.5. In the second (post-September 2025 Charlie Kirk murder), 1,035 participants conversed with GPT-4o version 2.5. Both groups received at least five rounds of evidence-based dialogue averaging seven minutes per session. The AI reduced participants' conspiracy beliefs more than static fact sheets or control chats—effects persisted for two months after the Trump event. The model adapted tactics: emphasizing epistemic humility for low-information events, and societal harms for well-documented ones. Crucially, it could not draw on internal knowledge for either event as they occurred after the model’s training cutoff. Effects showed no significant impact on political violence support or trust in official explanations. This approach may help address misinformation without overwhelming users during crises, but works best when participants engage willingly and context matters. The source notes this is a case study with potential abuse risks for emerging conspiracies where debunking could be incorrect.

Source: The Decoder