The turn from static messages to interactive dialogue
The papers filed here trace a clear historical arc: from crafting the right message to building the right conversation. Voelkel2026-lc represents the endpoint of the static-message paradigm taken to its most rigorous extreme—a registered-report megastudy testing the ten most-cited climate messaging strategies on 13,544 Americans. Its conclusion is sobering: several strategies reliably nudge attitudes by a few percentage points, effects are strikingly similar across parties (undermining the assumption that audience-targeted framing is necessary), and none moves costly behavior like actual donations. This attitude-behavior gap, and the modesty of one-shot message effects generally, sets up the motivating problem for much of the rest of the collection: if brief, static messages have a low ceiling, what else might work?
One answer is interactivity. Costello2024-bg demonstrates that personalized, multi-round dialogue with an LLM can durably (≥2 months) cut conspiracy belief by ~20%, generalizing even to unrelated conspiracies—a result the authors read as vindicating analytic-reasoning accounts of belief revision over motivational-immunity accounts. Kotz2026-lk extends this logic across contested domains (climate, vaccination, inequality), showing that GPT-5 dialogues shift both beliefs and policy support, work best on the most skeptical participants, and are amplified by trust in science—while also showing that belief change does not automatically transfer to policy support, which must be targeted directly. Lin2025-xp pushes the same mechanism into the electoral arena, finding that brief AI dialogues shift candidate preference and ballot-measure support more than traditional political ads, driven mainly by evidence and facts rather than manipulation—though with a troubling asymmetry in which right-leaning AI personas made more inaccurate claims. Together these three papers make a strong cumulative case that the “compelling, personalized evidence” delivered conversationally, rather than the single clever framing, is where persuasive traction now lies.
Chatbots as inoculation and trust-brokers
A parallel strand asks what happens when this conversational capacity is folded into inoculation theory itself. Szabo2026-rd operationalizes “Conversational Inoculation” via the MindFort chatbot, finding that dialogue-based prebunking outperforms traditional reading/writing refutation once individual baseline susceptibility is controlled—success depending as much on interactional factors (trust-building, scaffolding independent thought, minimizing friction) as on argument content. Dubey2026-bl complicates the trust story from another angle: a “balanced” climate chatbot presenting mainstream and alternative views is, counterintuitively, trusted more by high-conspiracy-belief users than by skeptics of conspiracy, challenging assumptions that distrust is a fixed property of conspiratorial cognition rather than something contingent on how information is framed and delivered. Read alongside Costello’s and Kotz’s findings that skeptics are often the most movable audience, these chatbot studies converge on a provocative claim: distrustful audiences are not simply unreachable, and design choices around perceived neutrality, adaptivity, and dialogic partnership can matter more than the raw informational content.
Scaling and generalizing inoculation beyond the lab
Where the chatbot studies test depth of interaction, van-der-Linden2026-jt tests breadth: a 19-second prebunking video deployed as an Instagram ad to over 375,000 users, showing a 21-point gain in identifying fearmongering that persists five months and triples click-through rates—evidence that classic single-technique inoculation can be scaled cheaply into real social-media feeds, not just demonstrated in lab RCTs. This sits in productive tension with Lieu2025-nl, which finds that the specific rhetorical fallacy underlying a piece of climate misinformation (FLICC categories) makes little difference to perceived veracity, whereas the content category does—misinformation attacking climate solutions is rated most credible and most politically polarizing. Taken together, these two papers suggest inoculation and debunking efforts may get more traction from targeting emotionally resonant content domains (fearmongering, solution-attacking claims) than from cataloguing logical fallacies in the abstract.
Emotion, modality, and the psychology of the discerning reader
Two experimental papers turn attention to the psychological substrate that any intervention must work through. Xue2025-bp shows that fact-checks themselves are saturated with emotional language despite fact-checking’s rhetoric of objectivity, and that this emotionality shapes engagement and sentiment toward fact-checked targets—a finding that sits awkwardly beside Lieu2025-nl’s null effect for fallacy type but resonance with content/emotion, suggesting affect may be a more potent lever than logical structure across both misinformation and correction. Choi2026-bz identifies a more subtle mechanism operating independent of message content altogether: difficulty authenticating one deepfake carries over to shape confidence and susceptibility to the next piece of fake content, but only within the same modality (text-to-text, video-to-video). This “modality-congruent carryover effect” implies that interventions calibrated to one media format may not transfer their protective effects to another, a caution relevant to designers of prebunking campaigns like van-der-Linden2026-jt who must consider whether video-based inoculation protects against text-based manipulation and vice versa.
Synthesis and self-critique
Spampatti2026-kx functions as the collection’s methodological conscience, arguing that misinformation research should borrow from climate psychology’s longer struggle against organized denial. Its core critiques land squarely on the papers above: fewer than 1% of misinformation studies measure real-world behavioral outcomes (a charge Voelkel2026-lc and van-der-Linden2026-jt partially answer by testing donations and click-throughs respectively, with the former’s null donation effects illustrating exactly the attitude-behavior gap Spampatti warns about); interventions should engage motivational and social drivers, not just cognitive ones (echoed in Szabo’s finding that trust and interactional design matter as much as argument logic); and generalizability requires attention to cross-country and cross-domain heterogeneity, a gap that Kotz2026-lk and Lin2025-xp begin to close by testing multiple issues and multiple national elections. Read as a whole, the arc across these eleven papers moves from the limits of one-shot persuasive messaging, through the discovery that dialogic, personalized, AI-mediated interaction breaks through where static messages plateau, toward an emerging recognition—voiced most explicitly by Spampatti—that durable, generalizable, real-world impact will require treating misinformation resistance as a systems-level, trust-mediated, and behaviorally measured problem rather than a search for the single best message or fallacy to debunk.