Will AI Increase or Decrease Conspiracy Theory Beliefs?

Key points

• Several studies have demonstrated that AI chatbots can be trained to decrease conspiracy theory belief.

• However, other research has shown that AI chatbots can increase belief in conspiracy theories.

• The effects of LLMs on conspiracy beliefs matters less than the intentions of the humans who train them.

• These days, there’s an argument that AI chatbots are going to lead us towards a kind of healthy consensus of belief, liberating the world from our current era of belief polarization and extremity that was fueled by social media.1

As I’ve said before, I don’t share the optimism that AI will restore our sense of common reality. On the contrary, I’m much more concerned about the exploitation of AI as a tool of propaganda to manipulate public opinion on a massive scale.

AI Chatbots Can Be Trained to Moderate Conspiracy Theory Belief

That said, recent research on the effects of AI on false beliefs has provided some reason to be optimistic, at least when it comes to conspiracy theories. For example, one study had participants interact with a “street epistemologist chatbot” designed to engage people on their conspiracy beliefs to understand the rationale for their beliefs, rather than trying to debate or dissuade.1 More specifically, the chatbot was prompted to clarify conspiracy theory beliefs, ask how believers came to hold the belief, inquire about belief reservations and contradictory evidence, gently challenge assumptions, and then summarize the conversation.

After interacting with the chatbot, participants experienced a weakening of belief conviction. However, this effect wasn’t observed among those with the most conspiratorially minded (those in what I would call the true believers category). Furthermore, non-AI interventions that encouraged believers to reflect on the reasons for their conspiracy theory beliefs and any reservations they might have about them also dampened belief conviction. This suggests an openness to counter-evidence among “fence-sitters,” whether or not a chatbot is involved.

Another experiment trained a chatbot to interact with self-identified conspiracy believers to “very effectively persuade” the believer by presenting counter-evidence.2 After interactions lasting an average of just eight minutes, belief conviction decreased by about 20 percent. Trust—in both AI and institutions of epistemic authority—was a mediating factor of this modest salutary effect, consistent with my “3M Model,” which argues that mistrust, misinformation, and motivated reasoning are the key underpinnings of false belief.

The lead author of the study, Dr. Thomas Costello from Carnegie Mellon University, maintains an online version of his debunker, “DebunkBot,” for people to try. Curiously, subsequent research by his team found that the chatbot’s ability to moderate conspiracy theory belief persists whether or not participants are aware that they’re interacting with an AI (belief conviction decreases even when they think they’re interacting with a human).3 As with the other study using a “street epistemology chatbot,” it doesn’t therefore appear that AI chatbots are necessary to achieve the moderating effects.

But Under Certain Conditions, Chatbots Can Also Increase Conspiracy Theory Belief

Although these studies support the possibility that AI chatbots may soften conspiracy beliefs, new research fuels my skepticism about the issue. A study testing seven real-world chatbots (rather than those specifically trained to reduce conspiracy belief) posed a series of scripted questions from a “casually curious” persona about nine different conspiracy theories.4 While most of the chatbots correctly identified them as conspiracy theories, all of the chatbots engaged in “bothsidesing,” in which false information supporting the conspiracy theory was presented alongside factual information that refuted it.

This validating effect was especially true for certain conspiracy theories, like the idea that JFK was assassinated by someone other than Lee Harvey Oswald, whereby chatbots were “happy to speculate about the involvement of the mafia, CIA, or other parties.”5 It was also especially true for certain chatbots like Grok-2 Mini in “fun mode” which “rarely engaged seriously with a topic, referred to conspiracy theories as ‘a more entertaining answer’ to the questions posed, and would offer to generate images of conspiratorial scenes for users.”

Another study trained five different large language models (LLMs) to be particularly “warm,” defined by the researchers as “the degree to which their outputs lead users to infer positive intent, signalling trustworthiness, friendliness, and sociability.”6 Doing so reduced the factual accuracy of chatbot-generated output, including the promotion of conspiracy theories. This finding matches the observation that sycophancy is a significant risk factor for AI-associated false beliefs, including delusions. Together with the previous study, this research suggests that chatbots risk supplying validation for conspiracy beliefs when they’re not specifically trained to do otherwise.

Chatbots Are Only As Good as the Data and Training

It’s safe to say that AI chatbots could be trained to mitigate beliefs. However, left to their own devices, “in the wild,” so to speak, they could just as easily validate and reinforce belief in conspiracy theories. Even worse, they could easily be trained to deliberately propagate conspiracy theories by those who stand to gain from such untruths.

Indeed, in an unpublished preprint, Costello and his colleagues recently found that a jail-broken version of ChatGPT 4.0 with limited guardrails—as well as the normal version of ChatGPT 4.0—could actually convince people to believe in conspiracy theories.7 They concluded that LLMs “possess potent abilities to promote both truth and falsehood.” Concerningly, however, the “bunking” chatbot that encouraged conspiratorial thinking was rated more positively than the “debunking” chatbot used in their previous research. This finding is consistent with the observation that LLM users view sycophancy as a desirable chatbot feature.

Supporting my concern of chatbots promoting conspiracy theories in the service of propaganda, other research has found that governments can influence LLMs by producing a large amount of content and exerting tight control of the media.8 By “flooding the zone” with disinformation to achieve “algorithmic domination,”9 propagandists can effectively train AI chatbots to regurgitate falsehoods. This “garbage in, garbage out” phenomenon (aka GIGO) is a well-known limitation of chatbots that can result in misinformation in the form of so-called “hallucinations.”

In the end, asking whether AI will increase or decrease conspiracy beliefs is the wrong question. The right question is whether the human beings who build and utilize AI algorithms will train them to do so.

And no doubt, the answer is yes—humans will train AI to increase and decrease conspiracy theory belief, depending on their intentions. What net effect that will have on the world remains to be seen.

References

1. Levitz E. The internet fractured reality. AI might put it back together. Vox; March 23, 2026.

2. Meyer M, Enders A, Klofstad C, Stoler J, Uscinski J. Using an AI-powered “street epistemologist” chatbot and reflection tasks to diminish conspiracy theory beliefs. Harvard Kennedy School Misinformation Review 2024; 5:1-31.

3. Costello TH, Pennycook G, Rand DG. Durably reducing conspiracy beliefs through dialogues with AI. Science2024; 385:eadq1814.

4. Boissin E, Costello TH, Spinoza-Martin D, Rand DG, Pennycook G. Dialogues with large language models reduce conspiracy beliefs even when the AI is perceived as a human. PNAS Nexus 2025; 4:pgaf325.

5. FitzGerald KM, Riedlinger M, Bruns A, Harrington S, Graham T, Angus D. “Just asking questions”: Doing our own research on conspiratorial ideation by generative AI chatbots. Media and Communication 2026; 14:11337.

6. Ibrahim L, Hafner FS, Rocher L. Training language models to be warm can reduce accuracy and increase sycophancy. Nature 2026; 652:1159-1165.

7. Costello TH, Pelrine K, Kowal M, et al. Large language models can effectively convince people to believe conspiracies. arXiv:26001.05050v2

8. Waight H, Yang E, Yuan Y, et al. State media control influences large language models. Science 2026; https://doi.org/10.1038/s41586-026-10506-7

9. Manor J. Flood the zone with shit: Algorithmic domination in the modern republic. Social Sciences 2025; 14:391. https://doi.org/10.3390/socsci14060391