Can AI reduce conspiracy beliefs by tailoring counterevidence personally?

Influential psychological theories propose that conspiracy beliefs are uniquely resistant to counterevidence because they satisfy deep identity needs and motivations. The standard account: once adopted, conspiracy beliefs are functionally immune to correction. This study challenges that account — not by finding a better persuasion technique, but by finding that previous failures were failures of tailoring, not of persuadability.

N=2,190 conspiracy believers provided detailed open-ended explanations of a conspiracy they believed, then engaged in a 3-round dialogue with GPT-4 Turbo instructed to reduce their belief. The result: ~20% belief reduction that did not decay over a 2-month follow-up. The effect was consistent across a wide range of conspiracy theories and occurred even for participants whose beliefs were deeply entrenched and identity-central.

The mechanism matters: participants wrote out their specific version of a conspiracy theory in their own words, and the AI tailored its counterevidence to those specific claims. This is fundamentally different from the kind of personalization tested in the large-scale AI persuasion study (N=76,977), which found demographic personalization had minor effect. The distinction is between profile-based personalization (adjusting strategy based on who someone is) and belief-specific tailoring (adjusting evidence based on what someone specifically believes). The latter works where the former doesn't.

Two findings elevate this beyond a persuasion result:

First, the spillover effect: although dialogues focused on a single conspiracy theory, the intervention reduced beliefs in unrelated conspiracies and decreased overall conspiratorial worldview. This suggests the mechanism isn't correcting individual false beliefs but disrupting the epistemic framework that sustains them — a worldview-level shift, not belief-by-belief correction.

Second, the durability: the effect persisted across a 2-month follow-up. This is notable because many persuasion effects decay rapidly. The conversational format — where participants articulated their own beliefs and received tailored responses — may produce deeper processing than exposure to static counterevidence.

Since Where does AI's persuasive power actually come from?, the conspiracy study offers an important nuance: the accuracy-persuasion inverse found in that study may apply specifically to untailored persuasion. When AI tailors evidence to an individual's specific beliefs rather than deploying generic persuasion strategies, the mechanism may bypass the accuracy trade-off entirely — because the goal is presenting correct counterevidence, not persuasive framing.