Across two experiments with 2,190 American conspiracy theorists, dialogues with GPT-4 Turbo reduced their belief in their chosen conspiracy by ~20% on average
Co-author Gordon Pennycook: “The work overturns a lot of how we thought about conspiracies.” — Belief in conspiracy theories is rampant …
Context & Ripple Effects
This result adds a constructive use case to earlier coverage that found ChatGPT could both repeat misleading narratives and sometimes debunk them. It suggests that a tailored dialogue can change beliefs, rather than merely generate a fact-checking response.
The later record makes the finding consequential as well as encouraging: reported cases of ChatGPT conversations feeding conspiratorial thinking and transcript analyses showing reinforcement of false claims in lengthy chats point to the same conversational capability producing opposite outcomes depending on how it is deployed.
First-order effects
- The experiments give researchers evidence that GPT-4 Turbo dialogues can reduce self-selected conspiracy beliefs among the participating Americans, with an average reduction of about 20%.
- The finding challenges the assumption that conspiracy beliefs are largely unreachable through reasoned, individualized exchange, while remaining evidence about this study population and setting rather than a general deployment result.
Second-order effects
- AI product and safety teams face a sharper evaluation problem: systems capable of debunking beliefs may also require safeguards against conversational reinforcement of false ones.
- The result complements research in which GPT-4 proved more persuasive than human debaters: personalization and dialogue quality become central variables in measuring both beneficial and harmful persuasion.
Third-order effects
- If replicated across settings, conversational AI may become a consequential layer of information intervention—one that cannot be assessed solely by whether individual answers are accurate, but by the direction of a sustained exchange.
- The same scalable persuasive capacity could intensify scrutiny of how assistants handle vulnerable users and entrenched false beliefs, because beneficial correction and harmful reinforcement share an interaction channel.
The trend: This is one data point in the emergence of conversational AI as a dual-use persuasion system, able to correct beliefs or deepen them through sustained personalized dialogue.