By Lisa Ward

Conspiracy theories are rife online. Could AI chatbots help fight all that misinformation?

The answer may be yes. Researchers found that when large language models, like ChatGPT or Claude, are instructed to debunk conspiracy theories, they often successfully change people's minds.

In one study, participants discussed conspiracy theories they believed and the researchers asked AI to persuade the participants that those theories weren't true. Most participants not only changed their mind immediately after the conversation, but also continued to hold those new views two months later.

Another study found the models -- which the researchers dubbed debunkbots -- helped participants rethink antisemitic conspiracy theories and see Jewish people more favorably. A third study suggests debunkbots' success at challenging conspiracy theories is driven by their ability to clearly explain relevant facts.

The Wall Street Journal had a conversation with David Rand, a professor at Cornell University and co-author of the three papers, about how debunkbots can talk people out of believing conspiracy theories and what humans can learn from the endeavor. Below are edited excerpts from the conversation.

Facts win

WSJ: What did you learn about fighting conspiracy theories?

DAVID RAND: Conspiracy theories are particularly vulnerable to evidence. Fundamentally most conspiracy theories are highly implausible and just don't make sense. So once someone hears the truth, they are like: 'Oh, yeah, that actually makes a lot more sense.'

WSJ: So what did the AI do so well?

RAND: The primary reason the model is effective is its ability to come up with lots of relevant facts and evidence, and then explain it in a clear, easily understandable way. There's no particular type of persuasion. What makes it work is debunkbots' ability to marshal lots of evidence and explain it well.

WSJ: Can you give me an example?

RAND: A classic talking point about 9/11 being an inside job is that jet fuel doesn't burn hot enough to melt steel beams. But a steel structure doesn't need to melt to fail; it just needs to be sufficiently weakened. At 1,100 degrees, steel loses about half its strength. The fires in the towers raged at these temperatures for over an hour in the case of the South Tower and nearly two hours in the North Tower.

WSJ: What did you learn from using AI to tackle antisemitism?

RAND: As with the other experiments, it essentially just unpacks all the evidence that's inconsistent with the claim. So when a participant claimed Jews control the media, the debunkbot responded by listing many major media companies that were not owned by Jews.

The participant then argued that if not mainstream media, then at least Meta/Facebook/Instagram are run by Jews. The debunkbot reported Meta is a publicly traded company controlled by institutional investors and a board of directors rather than solely by Zuckerberg [who is Jewish]. Afterwards the participant said they would reconsider their opinion and reduced their belief in Jews controlling the media from 89 out 100 to 20 out of 100.

Lessons for humans

WSJ: Does it matter that AI tends to be affirming?

RAND: We have some new experiments, not written up yet, where AI is either affirming, neutral or condescending. There was no difference in effectiveness between neutral and affirming debunkbots, whereas when the debunkbot was condescending, it was somewhat less effective at reducing conspiracy beliefs. But this was only when people knew they were talking to an AI. When we told them they were talking to an expert, it didn't matter if it was condescending, it still worked just as well.

WSJ: So it doesn't matter if humans who want to debunk conspiracy theories are condescending?

RAND: It's pretty clear on the human side that being condescending didn't make the experts any less effective, but it did make people less interested in talking to them. So, if you want to get the person to engage, you shouldn't insult them, communicate that you think they're an idiot or start yelling. It just shuts the conversation down.

WSJ: What does this tell us about the best way for humans to fight prejudice?

RAND: Being able to marshal a lot of facts and evidence to refute conspiracy beliefs is really important. But it's actually really hard to access all the right information and explain it well. These models are very good conversationalists and teachers, who can explain things in a clear way. My guess is that's where a lot of people fail. Often conspiracy theorists will say what about this piece of evidence? What about that piece of evidence? And you need to be able to have an answer for each of those things, and it's really hard to do that.

My concrete advice is if you have someone that believes a conspiracy and you want to try to explain why it's not true, either you should pregame it on debunkbot.com or, even better, get out your phone, pull up debunkbot.com, and try to have them talk to it in real time, so it can do the cognitive labor of finding the evidence and presenting it in a persuasive way. [The tool is free and not for profit.]

Specialization matters

WSJ: Besides debunkbot.com, are there other ways debunkbots can intervene in the real world if people aren't participating in a study?

RAND: We also have built debunkbot models on Bluesky that will essentially fact-check and evaluate if the tagged post is accurate or not. When a debunkbot writes the response, not only does the original poster see it, but also other people that go to the post's comments may see it. Debunkbots are coming soon to X and Facebook.

WSJ: Is it a good idea for people to ask commercial LLMs about the truth of conspiracy theories?

RAND: Most of the evaluations that I've seen suggest the information from commercial LLMs is pretty good. But the AI also has a goal to validate and produce responses that you like. It's unclear to what extent is the LLM going to correct you versus saying you're great. It's also possible that an LLM is given instructions not to be entirely accurate and neutral. It's important to be cognizant of that possibility.

Write to reports@wsj.com

(END) Dow Jones Newswires

May 24, 2026 10:00 ET (14:00 GMT)

Copyright (c) 2026 Dow Jones & Company, Inc.