iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
In a 2024 experiment involving 2,190 conspiracy believers, personalized conversations with GPT-4 Turbo reduced belief in a participant’s chosen conspiracy by about 20% on average. The reported effect lasted two months and extended to other conspiracy beliefs. That is evidence that some people can be persuaded—not proof that AI can change every believer’s mind.
What did the 2024 AI experiment find?
Thomas H. Costello, Gordon Pennycook, and David G. Rand reported the experiment in Science in 2024. Participants who believed in a conspiracy theory discussed it with GPT-4 Turbo, in a chatbot the researchers called DebunkBot. On average, belief in the selected theory fell by approximately 20%, according to MIT Sloan’s account of the study.
The change was not limited to the immediate response: the researchers reported that it persisted for two months and generalized to conspiracy theories beyond the one a participant had discussed. The study also found reduced intentions related to conspiracy content, including intentions around social-media engagement. These are measured outcomes in the study, not evidence that the intervention changed participants’ real-world behavior over time.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →MIT Sloan reports that roughly one-quarter of participants who initially rated belief in their chosen conspiracy above 50 on a 0–100 scale later rated it below 50. That means they moved from the believing side of the scale to uncertainty; it does not mean they all became certain the theory was false.
#1 Best Overall
How did DebunkBot try to persuade people?
This was not a generic fact sheet or a one-line debunk. Participants first described a conspiracy theory they believed and provided evidence they considered supportive. DebunkBot summarized their argument and asked them to confirm the summary before responding. It was instructed to make an evidence-based case against the belief, tailored to what the participant had said.
MIT Sloan reports that the intervention involved three written exchanges and took about eight minutes on average. Participants in the control group had an AI conversation about a neutral topic. The tailored exchange gave the chatbot a chance to address a person’s specific reasons for believing a theory rather than relying on a general rebuttal.
Rank #2
How do the related studies compare?
| Study | Intervention | Reported result | What it does not establish |
|---|---|---|---|
| Costello, Pennycook, and Rand, Science (2024); described by MIT Sloan | Personalized, evidence-based GPT-4 Turbo dialogue about a participant’s chosen conspiracy; three exchanges averaging about eight minutes | About a 20% average reduction in belief strength; effect persisted for two months and generalized to other conspiracy beliefs | Universal persuasion or lasting changes in real-world behavior |
| Harvard Kennedy School Misinformation Review study | Reflection prompts and an AI chatbot acting as a “street epistemologist,” probing reasons and reservations | All interventions reduced belief strength on average; people with high generalized conspiratorial thinking showed no significant reduction, while low- and middling-conspiracism participants did | The participant count and intervention duration are not stated in the study summary described here (Harvard Kennedy School Misinformation Review) |
| PNAS Nexus study (2025) | Dialogue with a large language model, with participants believing the interlocutor was human | Reduced confidence in conspiracy theories and other epistemically unwarranted beliefs; no significant differences by perceived speaker identity or prompt type | Universal effectiveness in real-world use; participant count and duration are not stated in the study summary described here (PNAS Nexus, 2025) |
Who is more likely to change—and who may resist?
The Harvard Kennedy School Misinformation Review study adds an important qualification to the 2024 result. Its reflection-based interventions lowered belief strength on average, but participants with high levels of generalized conspiratorial thinking did not show a statistically significant reduction. Participants with low or middling levels did.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
This does not contradict the Science study’s report that its intervention worked even among people whose particular beliefs were deeply entrenched. The studies describe different interventions and distinguish, in effect, between strength of a specific belief and a broader conspiratorial worldview. A person can be strongly committed to one theory without scoring highly on a general tendency to see events through conspiratorial explanations.
Rank #3
Could an AI also undermine belief in accurate claims?
Potentially. The Harvard study found that reflection reduced belief in non-conspiratorial statements as well as conspiracy beliefs. That suggests the effect may come partly from prompting people to reconsider claims generally, not from reliably identifying falsehoods and correcting only those. The authors warn that an AI designed to encourage doubt could also weaken confidence in true claims if it is misused or poorly directed.
The 2025 PNAS Nexus finding that effects did not significantly differ when participants thought they were speaking to a human suggests that the conversational approach may not depend on users recognizing an AI. It does not resolve the safety concern: a persuasive dialogue can shift confidence without guaranteeing that the shift tracks truth.
Rank #4
What can readers reasonably conclude?
The 2024 experiment shows that short, personalized, evidence-based AI conversations can reduce measured conspiracy belief for many participants, with effects that lasted two months in that study. The related findings suggest the response varies with a person’s broader conspiratorial outlook, and that reflection is not automatically truth-selective. These results support further study of conversational interventions, not treating AI persuasion as a proven way to change every mind.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
