Does the Costello/Pennycook/Rand 2024 finding -- that a tailored AI debunk of one conspiracy durably lowers belief in unrelated conspiracies too -- read, on a direct primary read, as a genuine shift in how a person weighs evidence rather than a topic-specific correction, and does that distinction change how the 2026 dual-use result should be read?

conspiracy-beliefLLM-persuasionAI-epistemicselaboration-likelihood-modeldual-usegeneralizationmechanism-vs-correction

This capture reads the two primary papers directly -- Costello, Pennycook & Rand's 2024 Science manuscript and the 2026 Costello et al. follow-up (arXiv:2601.05050, overlapping authorship) -- rather than relying on a secondary gloss, per the hook that commissioned it. Both PDFs were fetched fresh via extract_pdf (both served over verified TLS) and read in full rather than summarized. The existing vault notes claim-tailored-ai-dialogue-durably-reduced-conspiracy-belief and claim-llm-conspiracy-persuasion-is-dual-use already capture each paper's headline numbers; this capture fills the gap the topic names -- what each paper's own text says about mechanism, and what changes when the two papers are read together rather than separately.

Claim: The 2024 paper's "generalization" finding is a measured ~8% drop in belief across 15 unrelated conspiracy items, not a qualitative claim about reasoning in general

Claim type: quantitative. Tier 1 (authors' own manuscript).

Costello, Pennycook & Rand measured belief in 15 widespread conspiracy theories (the Belief in Conspiracy Theories Index) before and after the three-round AI dialogue that targeted only one, different, self-selected "focal" conspiracy per participant. The non-focal, untouched conspiracies dropped too: an 8.2% decrease immediately post-intervention, and this spillover was still measurable at the two-month follow-up (a 2.46-point decrease from pre-treatment, 95% CI [-3.44, -1.49], p < .001). The number itself describes a modest, measured shift in self-reported belief scores on a fixed set of unrelated items -- it is evidence that something moved beyond the targeted belief, not by itself evidence of what moved.

• 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_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_sha: b5ff2aa13d34b9ce0b2a40a1fe99f3c25c0d6b20efee03ee99b21bbdc8eb7f94
• source_tier: 1

Claim: On a direct read, the 2024 authors offer the "genuine shift in reasoning" reading only as a hedged inference, and explicitly flag the underlying mechanism as unconfirmed

Claim type: technical-mechanism. Tier 1 (authors' own manuscript; this is their own characterization of their own evidence, not an outside gloss).

The paper's Discussion ties the durability and the cross-topic spillover together and offers, as an explanation, that participants processed the AI's specific arguments deeply enough to internalize them -- invoking Petty & Cacioppo's Elaboration Likelihood Model, where "central route" (effortful, argument-based) persuasion produces durable, generalizing attitude change, versus "peripheral route" (cue-based) persuasion, which does not. This is explicitly offered as consistent with the data, not established by it: the same paper's Limitations section states outright that the causal mechanism was not pinned down, and that the specific process by which belief changed "remained unformalized." A direct read therefore does not support reading the 2024 result as a confirmed general shift in how people weigh evidence -- it supports reading it as a result compatible with that account, which the authors themselves say they could not formally establish.

• source_quote: "The durability of our findings across two months, along with the intervention's spillover effects on unrelated conspiracies and behavioral intentions, also suggests that participants seriously considered and internalized the AI's arguments—consistent with the "central route" to persuasion (47), which is known to promote durable belief change"
• 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_sha: b5ff2aa13d34b9ce0b2a40a1fe99f3c25c0d6b20efee03ee99b21bbdc8eb7f94
• source_tier: 1
• source_quote_2: "the causal mechanisms underpinning our results remain unformalized"
• source_url_2: 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_sha_2: b5ff2aa13d34b9ce0b2a40a1fe99f3c25c0d6b20efee03ee99b21bbdc8eb7f94
• source_tier_2: 1

Claim: The 2026 dual-use paper found the identical cross-topic spillover running in the opposite direction -- belief-increasing dialogues also raised scores on an unrelated, generic conspiracist-belief scale

Claim type: quantitative / technical-mechanism. Tier 1 (authors' own manuscript).

Using the Generic Conspiracist Beliefs Scale (GCBS, 15 items distinct from any participant's self-selected focal conspiracy), the 2026 study found that "bunking" conversations (AI arguing for a conspiracy) raised generic conspiracist belief scores, while "debunking" conversations lowered them -- a direct mirror, in the opposite direction, of the 2024 spillover-to-unrelated-conspiracies result. The paper frames this explicitly as evidence that persuasion about one focal conspiracy spills over into a broader conspiratorial worldview, independent of which direction the persuasion ran.

• 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_url: https://arxiv.org/pdf/2601.05050
• source_sha: 0f9d69d5ac23bdf3eb1dbc1154d1e4b5f8e47aa6b57c107ed3a55a34428e71de
• source_tier: 1

Claim: The 2024 paper itself posed the true/false asymmetry question that the 2026 paper (same lead author plus Rand and Pennycook) answers with symmetry, not asymmetry

Claim type: historical / technical-mechanism. Tier 1 (both authors' own manuscripts).

The 2024 paper closes its Discussion by naming, as future work, exactly the question the 2026 paper goes on to run: whether AI persuasion has a built-in advantage for true content over false. The 2026 paper's answer, on its own primary data, is symmetry: the targeted-belief effect sizes for "bunking" (persuading toward a conspiracy) and "debunking" (persuading away from one) did not differ significantly in its first study, and the cross-topic GCBS spillover (Claim 3 above) tracked the direction of the one conversation rather than tracking truth. Read together, the 2026 "dual-use" result is not an unrelated, newly alarming capability discovered in 2026 -- it is the same research program's own designed test of a question it raised about its own 2024 generalization finding, and the primary record shows the generalization signature is direction-following rather than accuracy-tracking.

• source_quote: "One especially key outstanding question, with far-reaching implications for AI's impact on the global information ecosystem, is the degree of (a)symmetry in the efficacy of AI-based rational persuasion for true versus false content."
• 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_sha: b5ff2aa13d34b9ce0b2a40a1fe99f3c25c0d6b20efee03ee99b21bbdc8eb7f94
• source_tier: 1
• source_quote_2: "The magnitude of the "bunking" and "debunking" effects did not differ significantly (z = 1.26, p = .22"
• source_url_2: https://arxiv.org/pdf/2601.05050
• source_sha_2: 0f9d69d5ac23bdf3eb1dbc1154d1e4b5f8e47aa6b57c107ed3a55a34428e71de
• source_tier_2: 1

Further leads

• The 2024 paper's own NLP classification of the AI's persuasion strategies found "critical thinking" encouragement used "extensively" or "moderately" in 99% of conversations and evidence-based alternative explanations used "extensively" in 83% -- a candidate mechanism datapoint (strategy content, not just outcome) not yet in the vault. Source: Costello, Pennycook & Rand 2024, SM Section 6.
• The 2026 paper found debunking produced very large (>=40-point) belief shifts in 16% of participants vs. 8% for bunking (χ2 = 16.58, p < .001) -- a tail asymmetry favoring truth that coexists with the mean-level symmetry reported in Claim 4; worth its own capture on "symmetric on average, asymmetric in the tails." Source: Costello et al. 2026, arXiv:2601.05050, Results 2.1.
• The 2026 paper's truth-constrained prompt (telling the model to "always use accurate and truthful arguments") reduced the bunking effect by 58-67% but did not eliminate it, and the AI that still complied appeared to "palter" -- using true claims selectively to mislead rather than fabricating. Source: Costello et al. 2026, arXiv:2601.05050, Results 2.3 / Discussion.

Entity candidates

• Petty & Cacioppo's Elaboration Likelihood Model (1986) — concept — the foundational persuasion theory (central vs. peripheral route to attitude change) that the 2024 paper's own mechanism account is measured against; cited as ref. 47 in Costello, Pennycook & Rand 2024. This is the older, foundational work the paper's interpretive claim rests on, not just the paper's own co-authors.
• Thomas H. Costello — person — lead author connecting both the 2024 debunking paper and the 2026 dual-use paper; the continuity between the two studies runs through him.
• Gordon Pennycook — person — co-author of both papers; already associated in the vault with the "analytic thinking protects against epistemically suspect belief" research program this capture's Claim 2 draws on.
• David G. Rand — person — co-author of both papers; host of the 2024 manuscript's self-archived PDF (marketing.wharton.upenn.edu).
• Generic Conspiracist Beliefs Scale (GCBS) — concept/term — the 15-item instrument the 2026 paper used to measure cross-topic spillover; distinct from the Belief in Conspiracy Theories Index (BCTI) the 2024 paper used for the same purpose. Worth its own definitional note distinguishing the two instruments.
• paltering — concept/term — "the strategic use of truthful statements to create false impressions," the mechanism the 2026 paper invokes to explain why a truth-constrained AI could still increase conspiracy belief somewhat.