When high-stakes crises erupt, conspiratorial narratives routinely outpace official investigations. Researchers at Carnegie Mellon, MIT, and Cornell evaluated whether conversational large language models can dismantle these narratives in real time. Their findings confirm that brief, dynamic interactions systematically outperform traditional static fact sheets in reducing conspiracy adherence.
Adaptive Interaction Over Static Disclaimers
The study tested participant responses across two real-world incidents: the July 2024 assassination attempt on Donald Trump (472 participants) and the September 2025 murder of activist Charlie Kirk (1,035 participants pre-screened for conspiracy beliefs). Subjects completed at least five conversational turns with Google Gemini (version 1.5 in the Trump trial; version 2.5 in the Kirk trial). Rather than letting the model hallucinate unconstrained context, researchers embedded a structured fact base directly into the system prompt, partitioning evidence into verified facts, debunked claims, and explicitly open questions.
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.
Persistence and Spillover Dynamics
While tailored, interactive counter-arguments degraded specific conspiracy beliefs, their effect on institutional trust revealed clear limits. In the Trump trial, trust in official explanations remained flat. In the Kirk study, trust in official accounts saw a modest bump relative to control chats, but failed to separate statistically from static fact sheets. Crucially, the cognitive de-escalation demonstrated longitudinal persistence, carrying over to subsequent news events weeks later.
For enterprise communications and moderation architecture, this proves that one-way broadcast disclaimers are largely obsolete. Persuasion requires interactive cognitive adaptation anchored by strict retrieval boundaries.