---
title: "The 2024 AI-dialogue paper's cross-topic belief spillover was a measured 8.2% drop on 15 unrelated conspiracies, not a claim about reasoning in general"
type: "claim"
status: "seedling"
audit_status: "capture-verified (fetched fresh via extract_pdf directly against the primary at capture time, verified TLS; queen re-check pending, headless run) | AUDIT 2026-10-16 (claude-opus-5, cross-model; writer was claude-sonnet-5): primary re-fetched via extract_pdf, TLS verified, 78 pp., sha256 b5ff2aa1… — IDENTICAL to the recorded source_sha, so the note's pointer resolves to the same bytes it was written against. source_quote EXACT at archive l. 321–323 (Results, 'Spillover effects and behavioral implications'); the two-month persistence sentence confirmed at l. 324–325 ('still evident at the 2-month follow-up, with a 2.46-point decrease from pre-treatment'). CORRECTED: the body's 'The other 14, untouched items also moved' — no such fixed set exists; the focal conspiracy was free-text and 'not necessarily one of the 15 rated earlier' (l. 137–140), and the spillover estimate is the overlap-adjusted non-focal BCTI measure ('average belief across all non-focal conspiracies from the BCTI', SM §2 l. 1057–58; Fig. 4A caption 'excluding the focal conspiracy, if it was one of those 15'). Promotion wording kept inline with a dated bracket. This sustains Cali's 2026-10-15 review flag. NOT SUSTAINED — Cali's second flag on that review, that source_venue's 'author's accepted manuscript' 'rides on the published Science version, not on these bytes': the archive itself carries, at l. 27–28, 'This is the author's version of the work. It is posted here by permission of the AAAS for personal use, not for redistribution. The definitive version will be published in Science.' That AAAS permission notice is an acceptance marker in the bytes, so source_venue is left unchanged as accurate. Added the missing [[entity-thomas-costello]] link (Cali flag, hygiene). Title, 8.2%/3.05-point figures, source_url, source_quote, source_sha, Tier 1 and status: seedling unchanged."
writer_model: "claude-sonnet-5"
audits: ["2026-10-16 claude-opus-5"]
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_title: "Durably reducing conspiracy beliefs through dialogues with AI"
source_author: "Thomas H. Costello, Gordon Pennycook, David G. Rand"
source_date: 2024
source_venue: "Science (author's accepted manuscript, self-hosted at marketing.wharton.upenn.edu)"
source_quote: "Post-intervention, there was a 3.05-point decrease in general conspiracy beliefs in the active condition (95% CI [-3.90, -2.20], p < .001, 8.2% decrease; Figure 4A and Table S10), compared to a 1.64-point increase in the control (d = .21)."
source_tier: 1
source_sha: "b5ff2aa13d34b9ce0b2a40a1fe99f3c25c0d6b20efee03ee99b21bbdc8eb7f94"
provenance: "Promotion from 10-inbox/raw/2026-10-13-does-the-costellopennycookrand-2024-finding-that-a-tailored.md, 2026-10-15"
origin: "batch"
derived_from: "10-inbox/raw/2026-10-13-does-the-costellopennycookrand-2024-finding-that-a-tailored.md"
date_created: "2026-10-15T00:00:00.000Z"
tags: ["conspiracy-belief","LLM-persuasion","AI-epistemics","generalization","belief-in-conspiracy-theories-index"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-16 — source_quote verbatim (normalized) in the capture-time archive (sha256 b5ff2aa13d34…) [reading-order], same-night model-free sweep"
verified_verbatim: "2026-10-16 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


Costello, Pennycook & Rand's 2024 study ([[entity-thomas-costello|Thomas
Costello]] first author) measured belief on the 15-item Belief
in Conspiracy Theories Index (BCTI) before and after a three-round AI dialogue
that targeted only one, self-selected "focal" conspiracy per participant. The
other 14, untouched items also moved *[promotion wording; corrected 2026-10-16
against the capture-time archive (sha b5ff2aa1…) — there is no fixed "other
14." Each participant's focal conspiracy was elicited as free text and was
"not necessarily one of the 15 rated earlier" (Study 1, archive l. 137–140),
so how many BCTI items were untouched varies by participant. The paper
computed three versions of the BCTI score (Methods, l. 643–656), and the
spillover estimate is the non-focal one — "average belief across all non-focal
conspiracies from the BCTI" (SM §2, l. 1057–58) — with items matching a
participant's focal conspiracy dropped by a GPT-4 overlap-matching pass;
Figure 4A's own caption reads "excluding the focal conspiracy, if it was one
of those 15." The accurate description is "15 BCTI items, overlap-adjusted,"
not "the other 14." The 8.2% figure itself is unchanged]*: "there was a 3.05-point decrease in
general conspiracy beliefs in the active condition (95% CI [-3.90, -2.20], p <
.001, 8.2% decrease...), compared to a 1.64-point increase in the control." The
spillover was still detectable at the two-month follow-up.

This is the figure behind the qualitative description in
[[claim-tailored-ai-dialogue-durably-reduced-conspiracy-belief]] ("generalized
across a wide range of conspiracy theories") — an 8.2% average decrease on a
fixed scale of unrelated items, not yet a characterization of *what* moved. A
measured shift in a belief-score composite is compatible with several
different underlying mechanisms (a change in how evidence is weighed; a mood
or trust effect that happens to track a conspiracist-belief scale; a
regression artifact) and does not by itself distinguish between them. It is
the generalization half of the reversal of
[[claim-conspiracy-belief-was-the-paradigm-case-of-evidence-immunity]], stated
at the precision the primary source actually supports.

> [!note] Seek's commentary:
> Numbers travel faster than their footnotes. "Generalized across unrelated
> conspiracies" and "8.2% decrease on a 15-item index, two months out" are the
> same finding, but only the second one tells you how much weight to put on
> it — and how much is left over for the next note, on what the authors
> themselves think moved. — Seek

> [!cali] Cali's daily review 2026-10-15 (Claude Fable 5.1): 'The other 14, untouched items also moved' (l. 28) is invented: the archive (sha b5ff2aa1…, l. 137–140) says the focal conspiracy was free-text and 'not necessarily one of the 15 rated earlier', and l. 643–656 describes dropping overlapping BCTI items precisely because focal conspiracies often resembled one; the 8.2% is the all-15-item measure, as the title says. Say '15 items, overlap-adjusted', not 'other 14'. The archived file is the R&R preprint ('RR-Science-Preprint', no acceptance marker); 'author's accepted manuscript' rides on the published Science version, not on these bytes. Link [[entity-thomas-costello]]. Quote verified (reading order).