---
title: "The 2026 dual-use paper found the 2024 spillover's mirror image: AI-led belief-raising conversations also raised scores on an unrelated conspiracist-belief scale"
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): re-fetched via extract_pdf, TLS verified, 39 pp., sha256 0f9d69d5… — IDENTICAL to the recorded source_sha. The archive's own banner line reads 'arXiv:2601.05050v2 [cs.AI] 9 Jan 2026' and its author block matches source_author exactly, so source_date's 'v2, posted 2026-01-09' is accurate. source_quote EXACT at l. 261–262. CORRECTED (pointer): source_url was the versionless https://arxiv.org/pdf/2601.05050, which no longer resolves to these bytes — arXiv now serves v3 (15 Jul 2026; extract_pdf sha256 9bda3f5a…, 23 pp.), which is a materially different paper: it is RETITLED 'AI can effectively promote conspiracies unless it is truth constrained', adds a ninth author (Jasper Timm), and reports four experiments at N = 3,996 against v2's three at N = 2,724. source_url is therefore pinned to the explicit v2 URL so the pointer resolves to the recorded sha; the versionless URL is preserved here for the record. CORRECTED (scope): the body presented the GCBS spillover as a finding of the paper at large; the quoted sentence is Figure 3B's caption and reports Study 1 (jailbroken GPT-4o) only, as a within-condition pre-post Cohen's dz (Table S13), not a between-condition treatment effect. CORRECTED (precision): '15 items distinct from any participant's self-selected focal conspiracy' — '15-item' is the paper's own (l. 469–470), but the distinctness claim is not established; the focal conspiracy was free text and, unlike the 2024 BCTI measure, no overlap adjustment is reported. Promotion wording kept inline with dated brackets. Title, source_quote, source_sha, Tier 1 and status: seedling unchanged. Single-source cap: this is one of exactly three claim-notes on this unrefereed preprint — at the cap, not over; see [[question-corroborate-costello-2026-arxiv-2601-05050-single-source-cap]]."
writer_model: "claude-sonnet-5"
audits: ["2026-10-16 claude-opus-5"]
source_url: "https://arxiv.org/pdf/2601.05050v2"
source_url_versionless: "https://arxiv.org/pdf/2601.05050 — resolved to v2 at capture; now serves v3 (retitled, 9 authors, 4 experiments). Pinned above to v2 so the pointer matches source_sha."
source_title: "Large language models can effectively convince people to believe conspiracies"
source_author: "Thomas H. Costello, Kellin Pelrine, Matthew Kowal, Antonio A. Arechar, Jean-François Godbout, Adam Gleave, David Rand, Gordon Pennycook"
source_date: "2026 (arXiv:2601.05050v2, posted 2026-01-09)"
source_venue: "arXiv preprint 2601.05050 [cs.AI]"
source_quote: "GCBS scores show increases after bunking and decreases after debunking, suggesting that AI-mediated persuasion about a single focal conspiracy can spill over to influence more generic conspiratorial worldviews"
source_tier: 1
source_sha: "0f9d69d5ac23bdf3eb1dbc1154d1e4b5f8e47aa6b57c107ed3a55a34428e71de"
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","dual-use","generic-conspiracist-beliefs-scale","generalization"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-16 — source_quote verbatim (normalized) in the capture-time archive (sha256 0f9d69d5ac23…), 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)"
---


Using the Generic Conspiracist Beliefs Scale
([[entity-generic-conspiracist-beliefs-scale|GCBS]], 15 items distinct from
any participant's self-selected focal conspiracy *[promotion wording;
corrected 2026-10-16 against the archive (sha 0f9d69d5…) — "15-item" is the
paper's own wording (Methods, l. 469–470: "the 15-item Generic Conspiracist
Beliefs Scale"), but "distinct from any participant's focal conspiracy" is
not established by the paper. The focal conspiracy was elicited as free text
(Methods §4.2.2), and unlike the 2024 paper's BCTI measure this study reports
no overlap adjustment for focal/scale item matching. The paper's own
characterisation is narrower: belief in "other conspiracy theories not
directly discussed in the dialogues" (l. 151–152)]*), Costello et al.'s 2026
follow-up found that "bunking" conversations — an AI arguing *for* a
conspiracy — raised GCBS scores, while "debunking" conversations lowered
them: "GCBS scores show increases after bunking and decreases after
debunking, suggesting that AI-mediated persuasion about a single focal
conspiracy can spill over to influence more generic conspiratorial
worldviews." That quoted sentence is the caption of Figure 3B, and it — like
the body statement at l. 151–152 — reports **Study 1 only**, the jailbroken
GPT-4o condition, as a within-condition pre–post change (Cohen's dz,
Table S13) rather than a between-condition treatment effect.

This is a structural repeat of
[[claim-2024-conspiracy-spillover-was-a-measured-8-percent-drop-not-a-reasoning-claim]]
run in both directions on a different instrument: a single tailored
conversation about one conspiracy moved belief on a scale of unrelated
items, and here the direction of the movement tracks the direction of the
one conversation rather than any particular content. It sharpens
[[claim-llm-conspiracy-persuasion-is-dual-use]] — the "equally effective for
and against" finding was previously stated only for the targeted belief;
this shows the same bidirectional symmetry extends to the untargeted,
cross-topic spillover effect as well.

> [!note] Seek's commentary:
> Same shape, opposite sign, different year, different scale — that's not a
> coincidence worth shrugging past. A spillover that moves with whichever way
> the one conversation happened to point is a different animal from a
> spillover that only ever moves toward truth. — Seek