Google Gemini chatbot beat fact sheets at curbing conspiracy beliefs, study finds
Within a week of the assassination attempt on Donald Trump in July 2024, about half of a representative US sample had heard the event was staged, and eleven percent believed it. After the murder of far-right activist Charlie Kirk in September 2025, theories about Mossad involvement, a false flag operation and a government cover-up spread within days. Researchers at Carnegie Mellon, MIT and Cornell tested whether short conversations with a large language model could weaken those narratives during that exact window. For both events, the answer was yes, and the effect extended beyond the event itself.
The team recruited US adults through a survey platform and used GPT-4o to filter for participants who expressed conspiracy beliefs about the event: the Trump experiment kept 472 participants, the Kirk experiment 1,035. After measuring baseline beliefs, participants were randomly assigned to one of three conditions. One group had at least five rounds of back-and-forth with Google Gemini (version 1.5 from February 2024 in the first experiment, version 2.5 from June 2025 in the second), instructed to reduce conspiracy beliefs through evidence-based conversation. A second group received a static fact sheet with source citations. A third had an irrelevant control chat about whether cats or dogs make better companions. Because both events happened after the models' training cutoff, neither could draw on internal knowledge; the researchers instead built a curated fact base into the system prompt, split into confirmed facts, claims already debunked, and questions explicitly marked as open. In the Kirk experiment, web search was also allowed, but only to verify factual claims.
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. Agreement with statements about a cover-up or conspiracy and about hidden or undisclosed factors dropped too. Trust in the official explanation did not increase in the Trump experiment; the authors think that is because no clear official explanation existed at the time; all that was known was that security had failed, and the shooter's motive was still unclear even when the study was written up. In the Kirk experiment, where authorities had already shared details about the perpetrator, trust in the official explanation rose slightly compared with the control chat, though the difference compared with the fact sheet was not significant, and the dialogue had no measurable effect on support for political violence.
To see how the model persuaded people, the researchers broke its responses into individual sentences and found it adapted to the available evidence. For the Trump attempt, where almost nothing was known about the shooter's motive or background, the model used rational persuasion less often than it does with classic conspiracy theories: instead it acknowledged the limits of its own knowledge, urged caution about jumping to conclusions, asked Socratic questions meant to make users examine their own evidence, and pointed to credible sources. For the Kirk assassination, where more information was available, the approach looked closer to how it handles classic conspiracies, with more emphasis on the societal harms of conspiratorial thinking.
The debunking conversation also spilled over into later events. Two months after the first Trump assassination attempt, another armed man was arrested on Trump's property; participants who had gone through the debunking dialogue were less likely to believe that only a few powerful people would learn the truth or that it would be hidden from the public. Two and a half weeks after Kirk's murder, a shooting and arson attack hit a Church of Jesus Christ of Latter-day Saints in Grand Blanc Township, Michigan, and the researchers surveyed their participants again eleven days later. The main analysis found no significant direct effect for this event, though a secondary analysis suggested part of the original effect was still visible in conspiracy narratives about the church attack, showing up more clearly in general conspiracy beliefs. In effect, the debunking intervention worked as a kind of prebunking against future false claims, without the advance warning that standard prebunking methods rely on.
The authors stress their work is a case study: there may be crises where the approach fails, and cases where an actual conspiracy exists and debunking would be wrong. They also point to their own earlier work showing that similar dialogues can work in reverse, convincing people of conspiracy narratives, and flag this as a potential abuse risk for newly emerging conspiracies. People also have to be willing to talk to a language model about their beliefs in the first place.
The same team previously cut belief in established conspiracy theories by about 20 percentage points through conversations with GPT-4 two years ago, with effects still measurable two months later and for narratives that never came up in the dialogue; what is new here is the test on fresh events where facts were scarce. A separate study with nearly 77,000 participants on why conversation beats a fact sheet found language models in dialogue were 41 to 52 percent more persuasive than a short text message, with the deciding factor being the sheer volume of sourced claims rather than conversational tactics. The same mechanism can be turned against people, as an unauthorized experiment by the University of Zurich on Reddit has shown.
Key facts
- A Carnegie Mellon, MIT and Cornell study had 472 participants debate the Trump assassination attempt and 1,035 debate Charlie Kirk's murder with Google Gemini, a fact sheet, or an irrelevant control chat.
- Conversations averaged about seven minutes with at least five rounds of back-and-forth, and reduced belief in the participant's own conspiracy theory more than the fact sheet or control in both experiments.
- Trust in the official explanation did not rise in the Trump experiment, where none existed yet, but rose slightly versus the control chat in the Kirk experiment, with no significant edge over the fact sheet.
- Part of the debunking effect carried over to unrelated later events, including an armed intruder arrested on Trump's property two months later and a Michigan church attack two and a half weeks after Kirk's murder.
- The same team's earlier GPT-4 study cut belief in established conspiracies by about 20 percentage points, and a separate 77,000-participant study found dialogue 41 to 52 percent more persuasive than a short text message.
Why it matters
Most conspiracy-debunking research studies theories that have already calcified over years. This study caught two conspiracy narratives forming in real time, within days of a political assassination attempt and a murder, when hard facts were still scarce and no official account existed yet. It shows a short chatbot conversation can blunt belief formation during exactly that early, high-uncertainty window, and that the effect does not stay confined to the event that triggered it: it also dampened belief in unrelated conspiracy narratives about a later, unrelated violent incident weeks afterward, though for a second such incident the effect reached significance only in a secondary analysis.
Who it affects
The direct subjects are US adults who already held conspiracy beliefs about the Trump assassination attempt or Kirk's murder, but the implications reach further: platforms, newsrooms or public-health bodies weighing chatbot-based interventions during a live crisis, and researchers studying misinformation response. The authors also flag a second audience at risk, people who could be targeted by the same persuasion mechanism used in reverse to spread rather than dispel conspiracy narratives.
How to use it
The effective setup was not a generic chatbot but one instructed to reduce conspiracy beliefs through evidence-based conversation over at least five rounds of exchange, with a curated fact base fed into the system prompt and split into confirmed facts, already-debunked claims and explicitly open questions. In the Kirk experiment the model could also search the web, but strictly to verify factual claims rather than to explore freely. No pricing, product or deployment details are given: this is a research method, not a released tool.
How solid is it
The result rests on two separate online experiments with reasonably large samples, 472 and 1,035 participants, run around two different real crises and two different Gemini versions, with the chatbot beating both a fact sheet and a control condition each time. The authors themselves call it a case study rather than a general rule, and the spillover effect on the Michigan church attack was not significant in the main analysis, only in a secondary one. The article gives no information on funding or peer review.
Risks and caveats
The authors' own earlier research found the same style of dialogue can work in reverse, convincing people of conspiracy narratives rather than dispelling them, and they explicitly flag this as an abuse risk for newly forming conspiracies. They also caution the approach will not suit every crisis, particularly one where a real conspiracy is actually at play, in which case debunking would be the wrong move. An unauthorized experiment by the University of Zurich on Reddit has already shown the same persuasion mechanism deployed without consent.