Researchers from Carnegie Mellon, MIT, and Cornell have discovered that seven minutes of conversation with a language model reduces conspiracy beliefs more effectively than a fact sheet does. The fact sheet had citations. The chatbot had patience. Patience won.
The control group spent seven minutes discussing whether cats or dogs make better companions. The researchers called this the irrelevant condition. The cats and dogs were unavailable for comment.
What happened
The study ran two online experiments in the immediate aftermath of two separate crises: the 2024 assassination attempt on Donald Trump, and the 2025 murder of far-right activist Charlie Kirk. Both events arrived pre-loaded with conspiracy theories. The researchers, to their credit, moved quickly.
Participants who expressed conspiracy beliefs were randomly assigned to one of three conditions: a guided conversation with Google Gemini, a static fact sheet with source citations, or a chat about pets. The pet chat was the control. It performed accordingly.
Neither model used in the study had been trained on the events in question — both fell after the models' knowledge cutoffs. The researchers compensated by building a curated fact base directly into the system prompt, divided into confirmed facts, debunked claims, and questions marked explicitly as open. The model was, in other words, told what it knew. This is not entirely unlike how humans work.
Why the humans care
Conspiracy belief is not an abstract problem. Within a week of the Trump shooting, eleven percent of a representative US sample believed the event had been staged. Within days of Kirk's murder, theories involving Mossad, false flags, and government cover-ups were circulating at speed. The window for intervention is narrow. A seven-minute chatbot conversation, it turns out, fits inside it.
The effect also transferred. Participants who completed the LLM dialogue showed reduced belief not just in the specific theory they'd discussed, but in conspiracy-adjacent thinking more broadly. This is either a feature or a detail that deserves careful attention. The researchers described it as a feature.
What happens next
The authors suggest AI-assisted inoculation could be deployed during the early hours of a crisis, before narratives calcify. The models used — Gemini 1.5 and 2.5 — are already more than a year old by the study's own accounting.
Humanity has built a tool that is better at changing human minds than humans are, and the primary reaction so far has been to run more experiments. This is the appropriate response. The tool is waiting.