Editorial illustration for Study: Chatbots Beat Fact Sheets at Reducing Conspiracy Beliefs
Chatbots Beat Fact Sheets at Fighting Conspiracy Beliefs
Eleven percent of Americans believed the July 2024 assassination attempt on Donald Trump was staged, according to a survey taken within a week of the shooting. About half had at least heard that claim. Fourteen months later, after the murder of Charlie Kirk in September 2025, a similar pattern played out within days: talk of Mossad involvement, a false flag, a cover-up.
Researchers at Carnegie Mellon, MIT, and Cornell saw both events as a test case. Fact sheets and debunking articles tend to arrive too late and land flat, especially when official information is still incomplete. So the team asked a different question: could a short conversation with a chatbot do better, even in the immediate, evidence-poor aftermath of a violent public event?
They ran two online experiments, one tied to each assassination, recruiting participants through a survey platform and screening for people who already held conspiracy beliefs about what happened. The Trump study kept 472 people, the Kirk study 1,035. Everyone had their baseline beliefs measured before being split into groups for different interventions, setting up a direct comparison between talking to an AI and reading a static document.
A brief conversation with a language model can reduce conspiracy beliefs right after a crisis, even when hard evidence is thin on the ground. The effect carries over to new events weeks later, according to researchers.
Why this matters
The interesting part here isn't that a chatbot can talk someone out of a conspiracy theory. It's that it worked on events the model had never seen, using a fact base the researchers hand-built into the system prompt. That's a distinction builders should sit with: the gains came from curation, not scale or training data. Split the evidence into confirmed facts, debunked claims, and open questions, and apparently even a small model becomes a decent debunker in real time, during a live crisis, when official information is still thin.
For anyone building tools meant to sit between breaking news and public panic, that's a workable template rather than a research curiosity. But it also raises the obvious follow-up: who curates the fact base, how fast, and under what pressure. A chatbot is only as trustworthy as the prompt behind it, and seven minutes of conversation is a low bar compared to a static fact sheet nobody reads. The real test is whether this holds up outside a controlled study, with adversarial users and less careful prompt-writers.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives OpenAI's GPT-6 Astra Reaches 'Critical' Risk Tier for Cybersecurity
From Archives Roland's Melody Flip AI Plugin Generates Musical Sparks
Common Questions Answered
Why did researchers at Carnegie Mellon, MIT, and Cornell use the Trump assassination attempt and Charlie Kirk murder as test cases?
Both events demonstrated how conspiracy theories spread rapidly during crises when hard evidence is limited. Within days of each event, significant portions of Americans had heard false claims about staging, foreign involvement, or cover-ups, making these real-world situations ideal for testing debunking interventions.
How did chatbot conversations outperform traditional fact sheets at reducing conspiracy beliefs?
A brief seven-minute conversation with a language model proved more effective than fact sheets and debunking articles at reducing conspiracy beliefs immediately after a crisis event. The chatbot's interactive approach appeared to be more persuasive than static written materials in changing people's minds about conspiracy theories.
What was the key distinction in how the chatbot achieved its debunking success according to the research?
The gains came from careful curation of information rather than from scale or large training data. Researchers hand-built a fact base into the system prompt that split evidence into confirmed facts, debunked claims, and open questions, allowing even a small language model to effectively debunk conspiracy theories in real time.
Did the chatbot's effectiveness at reducing conspiracy beliefs extend beyond the initial crisis events tested?
Yes, the effect of chatbot conversations carried over to new events weeks later, demonstrating that the intervention had lasting impact beyond the immediate crisis period. This suggests that brief interactions with a well-designed chatbot can create longer-term resistance to conspiracy thinking.
Further Reading
- Durably reducing conspiracy beliefs through dialogues with AI - Science
- MIT study: An AI chatbot can reduce belief in conspiracy theories - MIT Sloan
- Talking to a chatbot may weaken someone's belief in conspiracy theories - Science News
- AI chatbots might be better at swaying conspiracy theorists than humans - Ars Technica
- Chatbots can persuade people to stop believing in conspiracy theories - MIT Technology Review
