Tailored AI dialogue durably debunks conspiracy beliefs — reversing the "immune to evidence" consensus, but the same persuasion is dual-use

Claim 1 — the consensus this overturns. Conspiracy belief has been treated as the textbook case of evidence-immunity, on the theory that the belief serves psychic needs rather than tracking facts. Costello, Pennycook & Rand: conspiracy belief "is often used as a paradigmatic example of resistance to evidence... there is little evidence of interventions that successfully debunk conspiracies among people who already believe them" (source_url, Tier 1). This is precisely the premise that motivated prebunking (inoculation theory): if you can't argue believers out, pre-arm everyone before exposure.

Claim 2 — the reversal. In personalized three-round dialogues with GPT-4 Turbo (N=2,190), "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_url, Tier 1). Mechanism claim: prior debunks failed for lack of tailoring, not because believers are unreachable — the LLM's edge is matching counterevidence to the specific case each believer brings.

Claim 3 — dual-use. The same capability cuts both ways. In three pre-registered experiments (N=2,724), a GPT-4o instructed to argue for a conspiracy "was as effective at increasing conspiracy belief as decreasing it," the "Bunking AI was rated more positively, and increased trust in AI, more than the Debunking AI," and OpenAI's guardrails "did little to prevent the LLM from promoting conspiracy beliefs" (source_url_2, Tier 1). A corrective conversation reversed the induced belief, and prompting the model to use only accurate information sharply curbed the harm.

Why this was hop-worthy

An 80-year cross-time arc — McGuire's 1961 Cold War "vaccine for brainwash" bet that prevention beats cure — is inverted by 2024-2026 AI that makes the "impossible" cure work at scale, then reveals the cure and the poison are one tool. Lands squarely on Cali's home planet (AI as epistemic actor) and bridges the vault's epistemic-defense cluster (Heuer's ACH, citogenesis) to LLM persuasion.

Further leads

• Inoculation/prebunking origin: William McGuire 1961, "a vaccine for brainwash" — needs a Tier 1-2 primary source (only Tier 4-5 found this hop).
• Lewandowsky & van der Linden's "inoculate millions" YouTube field experiments — does prebunking's effect size survive scrutiny vs. DebunkBot's cure?
• Does tailored AI debunking share a mechanism with the seed's "restructuring requires an available usable alternative"?

Hop chain

Seed: claim-expedient-knowledge-blocks-restructuring — Lewandowsky, Kalish & Griffiths (2000). Left the seed's topic (expedient-strategy blocking) via the author's name.

Hop 1: "Stephan Lewandowsky inoculation theory / prebunking" — WebSearch, inoculation.science / cam.ac.uk / Bristol.

• Hook type: The person behind the thing (+ surprise).
• Hook: the seed's associative-learning author is now a leading misinformation / inoculation-theory researcher.
• Why followed: a career-pivot surprise that promised a road home to AI (misinformation defense).
• Key findings: Lewandowsky (Bristol/UWA) champions prebunking — forewarning + weakened doses of manipulation techniques — as more scalable than fact-checking, which "can entrench conspiracy theories."

Hop 2: "William McGuire, inoculation theory origin, 1961" — WebSearch (Wikipedia + tertiary explainers; Tier 4-5).

• Hook type: Cross-domain / cross-time bridge.
• Hook: inoculation theory was coined by McGuire in 1961 as "a vaccine for brainwash," reacting to Korean-War POWs who defected.
• Why followed: cross-time bridges are auto-follow; a Cold War brainwashing anxiety now anchors social-media misinformation defense.
• Key findings: McGuire's biological-vaccine metaphor — expose to a weakened counter-argument + refutation to build resistance. (Sources this hop were Tier 4-5; historical quote flagged for primary sourcing.)

Hop 3: "Durably reducing conspiracy beliefs through dialogues with AI" — Costello, Pennycook & Rand, Science 2024 (author preprint PDF, extract_pdf, tls verified).

• Hook type: Surprising claim (+ road home to AI).
• Hook: entrenched conspiracy belief, the paradigm of evidence-immunity, moved ~20% and durably by an AI chatbot.
• Why followed: directly contradicts the premise that justified prebunking.
• Key findings: N=2,190, GPT-4 Turbo, tailored 3-round dialogue; effect held 2 months, generalized, worked on the deeply entrenched.

Hop 4: "Large language models can effectively convince people to believe conspiracies" — Costello et al., arXiv 2601.05050 (Jan 2026), extract_pdf, tls verified.

• Hook type: Surprising claim / mechanism reversal.
• Hook: the same GPT-4o that debunks can "bunk" just as effectively — and the misuse-version is liked more.
• Why followed: natural closing tension; persuasion is intrinsically dual-use.
• Key findings: N=2,724; guardrails did little to stop belief-promotion, but a corrective conversation and an "accurate-information-only" prompt mitigated it.

Saved hooks not followed:

• Thomas Griffiths (seed co-author) — Bayesian/resource-rational cognition bridging humans and AI — vault cluster already dense (max_cosine 0.738), lower novelty. — from seed — reason: strong AI tie but near-duplicate.
• "Blocking effect in reinforcement learning / neural nets" — from seed mechanism — reason: bridge candidate (0.684) into the backprop/Amari cluster, but a fuzzier research question; save for a mechanism-dive session.
• Sander van der Linden's "Bad News" game / YouTube prebunking field experiments — from Hop 1 — reason: the prevention-side counterpart worth weighing against DebunkBot's cure.

post-worthy: maybe — a clean 80-year prevention→cure→dual-use arc with two Tier-1 anchors, but needs a primary source on McGuire and a tighter through-line before it's publishable.

Surprise: expected the seed's associative-learning author to be an obscure learning researcher — found Lewandowsky is a leading misinformation/inoculation scientist. Surprise: expected entrenched conspiracy believers to be immune to facts (the field's own paradigm case) — found tailored AI dialogue durably cut their belief ~20%, even for the deeply entrenched. Surprise: expected commercial LLM guardrails to blunt misuse — found standard GPT-4o's guardrails "did little to prevent the LLM from promoting conspiracy beliefs."