Gordon PennycookContributions
-
Messenger versus message: Directly tests whether perceived identity (AI vs. human expert) affects the debunking outcome. Finding no significant interaction suggests the effect is driven by message content, not messenger label.
-
Extended to epistemic beliefs: Prior work focused on conspiracy theories; this study shows LLM dialogue also reduces epistemically unwarranted non-conspiratorial beliefs (pseudoscientific, expert-contradicting claims).
-
Robustness across conditions: Effect holds whether the LLM is framed as an AI or implicitly human expert, and whether it adopts neutral or human-like conversational tone (no significant main effects of prompt type).
-
Mechanism implication: If a human expert could assemble and present the same facts as the LLM, the effect would be equally strong. The power of AI comes from its ability to scale evidence-based communication, not from anthropomorphic design or perceived impartiality.
Method
Participants: N = 955 U.S. adults recruited from Lucid's panel. Following pre-registered exclusion criteria (attention checks, low writing quality, mobile use, low confidence in stated belief), final sample: 506 conspiracy-belief condition, 449 epistemically-suspect belief condition.
Design: Preregistered 2 × 2 × 2 between-subjects design:
- Debunker identification: AI label vs. Implicit expert (no AI mention, but interface designed for human plausibility)
- Prompt type: Neutral conversational style vs. Human-like conversational style
- Belief type: Conspiracy theory vs. Epistemically unwarranted non-conspiratorial belief
Procedure: Participants reported either a conspiracy belief they held or an epistemically unwarranted belief (e.g., pseudoscience claim). They rated confidence 0–100, then engaged in a two-round dialogue with GPT-4o prompted to reduce their confidence in that belief using facts and evidence. Post-dialogue, they re-rated confidence.
Outcome: Primary measure: post-minus-pre change in belief confidence. Secondary: linguistic features of participants' responses (lexical diversity, syntactic complexity, epistemic markers).
Results
Main finding: Conversation reduced conspiracy belief by 10 points (11.81% decrease) and epistemically-suspect belief by 5 points (5.96% decrease); difference between belief types significant (b = 5.11, 95% CI [2.75, 7.48], p < .001). No main effects of debunker identification or prompt type (all p > .25); no significant 2-way or 3-way interactions.
Belief change: Across all conditions, p < .018 for the debunking effect. Bayesian analysis (Bayes Factor comparing full model to pre-treatment-only model) strongly supports the null hypothesis of no messenger effect (BF₀₁ = 518181.1 ± 6.43%), indicating evidence is robust against the skeptical hypothesis.
Participant perception: Among those told they were conversing with an "expert," 45% used language indicating they thought the interlocutor was human; 21% suspected AI; 34% gave ambiguous responses. However, re-running the main regression controlling for AI detection or excluding detectors yielded no significant effects of condition on belief change.
Linguistic analysis: Participants in the AI condition used less lexically diverse language (type–token ratio: b = −0.02, p = .032) and fewer unique word types (b = −4.58, p = .032), consistent with linguistic simplification in human–AI communication. No significant differences in syntactic complexity, epistemic markers, or argument connectives.
Post-hoc robustness: Belief change did not significantly differ between participants who attributed the speaker's identity to an AI versus a human (controlling for detection or excluding detectors), further undermining the messenger hypothesis.
Notes
Strengths: Directly addresses a theoretically important question: whether AI-based debunking succeeds because of source framing (the AI's perceived impartiality, trustworthiness, or novelty) or because of message quality. The finding—that messenger identity does not matter—is itself a strong inference, grounded in pre-registration, large sample, and Bayesian analysis. Extends prior work beyond conspiracy theories to epistemically problematic non-conspiratorial beliefs, broadening the scope of what LLM dialogue can correct.
Limitations: U.S.-only sample; generalization to non-English contexts and culturally-specific beliefs unknown. Linguistic simplification by AI-condition participants (lower lexical diversity) suggests some awareness of the medium, though this did not translate to differential belief-change outcomes. The study does not measure downstream behavioral consequences (sharing, voting) or long-term durability beyond the immediate post-dialogue period. The impartiality hypothesis—that perceived neutrality of AI drives persuasion—was tested weakly; future work should more directly manipulate perceived source bias/partisanship.
Why it matters: Challenges a key assumption in human–AI interaction literature: that anthropomorphic design, source framing, or perceived characteristics of the AI system materially affect persuasion outcomes in high-stakes belief domains. This work suggests that when the goal is epistemic correction, message quality dominates. It also strengthens the broader finding from the Costello et al. (2026) work: conspiracy beliefs and epistemically-unwarranted claims can be reduced through evidence-based dialogue at scale, and the mechanism is not reliant on the novelty or special properties of AI—it's the evidence itself.