Conspiracy theories are often treated as beliefs that facts cannot touch. Research on DebunkBot points to a more complicated picture: when people receive specific evidence that speaks directly to their own claims, many do reconsider.
What the DebunkBot study found
In research published in the journal Science, over 2,000 conspiracy believers had a text conversation with DebunkBot, a model built on OpenAI’s GPT-4 Turbo. GPT-4 Turbo was described as the most up-to-date GPT model at that time.
The setup was direct. Participants first wrote, in their own words, a conspiracy theory they believed and the evidence that made it convincing to them. The AI model was then instructed to persuade the user to stop believing that conspiracy and adopt a less conspiratorial view of the world.
The exchange was short. It consisted of a three-round back-and-forth text chat, lasting 8.4 minutes on average. After that conversation, participants’ confidence in the belief dropped by 20%.
The result was not limited to one type of claim. The effect appeared for classic conspiracies, including the JFK assassination and the moon landing hoax, as well as more contemporary politically charged ones related to the 2020 election and covid-19.
About one in four participants, all of whom believed the conspiracy theory beforehand, said after the conversation that they did not believe it.
Why tailored evidence mattered
The research challenges a familiar assumption: that facts alone cannot change minds. The findings suggest that at least some conspiracy believers are not unreachable. They may be misinformed, or they may not have encountered a clear non-conspiratorial explanation for the specific issue that matters to them.
That distinction is important. A general correction may miss the point if someone’s belief rests on a particular technical claim. A tailored chatbot can respond to the exact argument the person gives, instead of offering a broad warning about misinformation.
The source article gives the example of 9/11 deniers who focus on the claim that jet fuel does not burn hot enough to melt steel. DebunkBot’s response addresses the narrower issue: while that point is true, the American Institute of Steel Construction says jet fuel does burn hot enough to reduce the strength of steel by over 50%, which is enough to cause such towers to collapse.
That kind of answer does two things at once. It acknowledges the part of the claim that is accurate, then explains why the conclusion does not follow. The study’s broader implication is that many conspiratorial arguments can seem plausible until the missing technical context is supplied.
Durable effects and fact-checking concerns
The researchers found that even people who began the conversation absolutely certain their belief was true showed marked decreases in belief. The same was true for people who said the belief was highly important to their personal worldview.
The effect also lasted. When participants were followed up with two months later, the reduction in conspiracy belief was just as large as it had been immediately after the conversations.
The study also tested whether the effect depended on people knowing they were talking to AI. In a follow-up experiment, AI debunking was just as effective when participants were told they were talking to an expert rather than an AI. That suggests the impact was not simply caused by the chatbot format.
Another large follow-up experiment pointed to the same conclusion. The effect came specifically from the facts and evidence the model provided. Letting people know the chatbot was trying to change their minds did not reduce its effectiveness, but telling the model to persuade without facts and evidence eliminated the effect.
Accuracy remains a central concern for any AI system used this way. The researchers hired a professional fact-checker to evaluate GPT-4’s claims and found that over 99% were rated as true and not politically biased. In cases where participants named conspiracies that turned out to be true, such as MK Ultra, the chatbot confirmed the accurate belief rather than trying to talk them out of it.
What this could mean for misinformation
Generative AI is often discussed as a tool for spreading disinformation. This work shows another possible use: deploying AI chatbots to answer conspiracy claims with timely, specific, evidence-based rebuttals.
The source article describes several possible applications. Bots could engage with users who share conspiratorial content on social media platforms. Search engines could link debunking AI models to conspiracy-related queries. A person could also hand a phone to a relative and let the AI handle a difficult conversation.
None of that removes the deeper social problem. The article notes that conspiracy theories play an outsize role in today’s political landscape, and that public disagreement over basic facts remains serious. But the study gives a more hopeful interpretation of how belief can change.
The core lesson is not that AI has a unique persuasive power. It is that accurate, well-targeted evidence can matter, especially when delivered with the patience and specificity that many human conversations lack.
For public debate, that is a meaningful finding. If many people do shift their thinking when they encounter strong evidence, then the challenge is not only psychological. It is also practical: how to get accurate information to people at the moment when they are weighing a claim.
Facts are not finished
The research fits into a broader argument made in the source article: facts are not dead. Corrections, warning labels, evidence-based arguments and simple reminders to consider accuracy can all reduce belief in falsehoods or the sharing of misinformation, according to the studies discussed there.
DebunkBot adds a concrete example at the extreme end of the problem. Even conspiracy beliefs, often seen as resistant to correction, can move when people receive relevant facts and evidence.
That does not mean every person will change their mind, or that every claim can be settled in one short chat. But it does suggest that the factual common ground missing from society today may be easier to rebuild if accurate information can be delivered widely, clearly and at the right level of detail.