Gordon PennycookContributions
-
Two large experiments (N₁=1,290, N₂=903) testing which features of AI dialogue are necessary for belief reduction. Study 1 manipulates content (Just-the-Facts vs. No Facts), framing (Overt vs. Adversarial), and presentation (Concise, Abstracted Summary, Belief-First Rating). Study 2 tests whether dialogue itself is necessary (Single Message vs. Baseline vs. Control).
-
Mechanism evidence: Facts & Evidence is the primary load-bearing component (b=4.85, p<.001); Rapport, storytelling, and non-adversarial framing contribute negligibly. Removing factual content (No Facts condition) reduces the effect by 7.2 points relative to Baseline (95% CrI [−11.4, −2.8]).
-
Durability and boundary conditions: Treatment effect generalizes across seven manipulations and holds even for high-conviction believers. Trust in AI moderates slightly (BF₁₀ > 30); social embeddedness and prior exposure to counterarguments do not significantly bound the effect.
-
Content analysis: The AI employs five persuasion-strategy components—Facts & Evidence, Sources & Experts, Report/Common Ground, Harms & Empathy, Stories/Metaphors—with measured efficacy. Facts & Evidence and Sources & Experts drive the bulk of the belief reduction; Harms & Empathy and Rapport are near-zero or negative.
Method
Participants: U.S. adults recruited via CloudResearch in two studies. Study 1: two waves (Sample 1a: n=806 after screening; Sample 1b: n=484 after weighted randomization) for a total analytic N=1,290 across eight conditions. Study 2: single wave with target of 400 conspiracy believers per arm and 200 control; final N=903 (Baseline n=336, Single Message n=364, Control n=203).
Design: Both studies pre-register an open-ended elicitation of one conspiracy belief, then randomize to treatment. Participants rate the statement on a 0–100 confidence scale pre and post. Exclusion criteria include CAPTCHA failures, attention-check failures, mobile-device use, and AI-generated or low-quality responses.
Treatment: A multi-turn dialogue in which an LLM (Study 1: GPT-4; Study 2: GPT-5.2) engages with the participant's stated conspiracy belief, providing counterevidence, alternative explanations, and reasoned arguments. Variations systematically remove or modify content (Just-the-Facts, No Facts, Concise) or framing (Overt persuasive intent, Adversarial debate format, Belief-First measurement).
Outcome: Belief in the focal conspiracy statement, measured by confidence rating immediately post-treatment (primary). Secondary measures include post-hoc belief importance, trust in generative AI, social embeddedness of the belief, and—in Study 2—12-day follow-up on open-minded thinking.
Results
Study 1 (N=1,290): Baseline treatment reduces belief by 10.9 points (95% CI [8.65, 13.1], p<.001, d=0.54). All evidence-based conditions (Just-the-Facts, Concise, Adversarial, Overt) maintain near-maximal effects (all posterior mean differences > 0.55, BF₁₀ > 9.12). The No Facts condition is dramatically weaker: −7.2 points (95% CrI [−11.4, −2.8], BF₁₀=11.43), supporting the hypothesis that facts are load-bearing. Among the highest-conviction believers (6–8 on belief importance), the effect is robust: 8.03 points (95% CI [0.35, 0.55], d=0.45).
Study 2 (N=903): Baseline treatment reduces belief by 14.6 points (80.5 → 65.9, 95% CI [−16.5, −12.7], p<.001). Single Message (one tailored AI response without back-and-forth) reduces belief by 2.8 points relative to Control (posterior mean difference=−10.28, 95% CrI [6.93, 13.65], BF₁₀ > 1,000), indicating dialogue is advantageous but not essential for a sizable effect.
Content analysis: The AI's dialogue incorporates five strategy families (Facts & Evidence, Sources & Experts, Report/Common Ground, Harms & Empathy, Stories/Metaphors). Facts & Evidence shows the strongest association with belief reduction (b=4.85 per SD, p<.001); Sources & Experts is second (b=3.24); Harms & Empathy and Rapport are near-zero or negative. When the AI is instructed to avoid evidence (No Facts), it compensates with anecdotes and appeals to empathy; these do not drive belief change.
Moderators: Trust in generative AI moderates weakly (posterior mean=2.13 per trust point, BF₁₀>30); even those with low trust show substantial effects (10.7 points at −1 SD below mean AI trust). Social embeddedness, prior counterargument exposure, pre-treatment belief importance, religiosity, and political identity do not significantly moderate the treatment effect.
Notes
Strengths: Large, pre-registered experiments with careful attention to mechanism. The systematic ablation (removing facts, shortening responses, reframing as debate) clearly demonstrates that diagnostic evidence is the critical component, while popular mechanisms (rapport, non-adversarial dialogue, persuasive appeals) are not necessary. The finding contradicts decades of folklore suggesting conspiracy beliefs are "sticky" and resistant to evidence. The work highlights a scalability advantage of AI: a human debunker might marshal 13–26 specific factual claims per conversation, but can't do so for every belief holder; an AI can.
Limitations: U.S.-only sample; beliefs about a single focal conspiracy per participant (though generalization across topics is shown). Follow-up beyond 12 days is limited. The paper does not address whether belief reduction translates to downstream behavioral change (e.g., sharing, voting) or whether effects persist beyond 2+ months (though follow-up data from Costello et al. 2024 suggests yes). Cost-benefit analysis of deploying such a system at scale (moderation, privacy, misuse prevention) is outside the paper's scope.
Why it matters: This is empirical proof that conspiracy beliefs are not immutable. The traditional view—that conspiracy theories are psychologically sticky, driven by identity-protective cognition, and resistant to evidence—implied that interventions must focus on identity, emotion, or social norms. This work shows that when evidence is delivered with sufficient quality and breadth, minds do change, even for high-conviction believers. For the misinformation-intervention literature, it is a significant counter to decades of null results and backfire effects by showing a mechanism (diagnostic factual counterevidence) that does move beliefs durably.