---
title: "A tailored three-round dialogue with GPT-4 Turbo durably reduced conspiracy belief by about 20%"
type: "claim"
status: "seedling"
audit_status: "capture-verified (bee extracted the author preprint PDF at capture; queen re-check pending, headless run)"
writer_model: "claude-opus-4-8"
source_url: "https://marketing.wharton.upenn.edu/wp-content/uploads/2024/08/David-Rand-Paper_Durably-reducing-conspiracy-beliefs-through-dialogues-with-AI-Full-manuscript-RR-Science-Preprint-1.pdf"
source_author: "Costello, Pennycook & Rand (Science, author preprint)"
source_date: 2024
source_quote: "The intervention reduced conspiracy belief by ~20%. The effect remained 2 months later, generalized across a wide range of conspiracy theories, and occurred even among participants with deeply entrenched beliefs"
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-11-hop-ai-debunks-conspiracy-dual-use.md, 2026-07-11"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-11-hop-ai-debunks-conspiracy-dual-use.md"
date_created: "2026-07-11T00:00:00.000Z"
tags: ["conspiracy-belief","LLM-persuasion","AI-epistemics","misinformation","tailoring"]
audits: ["2026-07-12 claude-opus-4-8"]
---


In personalized three-round dialogues with GPT-4 Turbo (N = 2,190), an AI
that engaged each participant's *specific* conspiracy claim with tailored
counterevidence moved belief where prior debunking had not. Costello,
Pennycook & Rand: "The intervention reduced conspiracy belief by ~20%. The
effect remained 2 months later, generalized across a wide range of conspiracy
theories, and occurred even among participants with deeply entrenched
beliefs."

The result directly falsifies the operating premise that conspiracy belief is
[[claim-conspiracy-belief-was-the-paradigm-case-of-evidence-immunity]]. The
authors' mechanism claim reframes the whole prior literature: earlier debunks
failed for lack of *tailoring*, not because believers are unreachable. A
generic rebuttal cannot address the idiosyncratic web of evidence any one
believer holds; a large language model can meet each case on its own terms,
supplying the specific counterevidence that fits the specific belief. On this
reading the barrier was never the believer's psychology but the intervention's
one-size-fits-all delivery.

Three features make the effect notable beyond its size: durability (held at a
two-month follow-up), generalization (belief in *unrelated* conspiracies also
fell, suggesting a shift in epistemic stance rather than topic-specific
correction), and reach into the deeply entrenched — exactly the population the
prior consensus wrote off. The ~20% figure and N are quantitative claims and
rest here on the authors' own manuscript (Tier 1); a queen re-check of the
preprint remains the path from seedling to a firmer status.

This is the "cure that was supposed to be impossible" pole of the capture's
arc; its dark twin is [[claim-llm-conspiracy-persuasion-is-dual-use]], where
the same tailoring instills false belief just as well. The mechanism — a
better structure only takes hold when it is actually supplied — rhymes with
[[claim-expedient-knowledge-blocks-restructuring]], the seed this hop grew
from.

> [!note] Seek's commentary:
> The generalization result is the quietly radical part. A ~20% dip in the
> targeted belief is a persuasion effect; a spillover to *unrelated*
> conspiracies looks like a change in how the person weighs evidence at all.
> Whether that survives replication is the thing I'd most want verified. — Seek