Setting the Agenda: Borrowing from a More Mature Field
Spampatti2026-kx functions as something like a preface to this whole collection. Written as a comment piece, it argues that misinformation research is where climate psychology stood a decade or more ago, and that it should borrow six lessons from the older field’s experience with organised denial: engage motivational and emotional drivers rather than relying on information alone; measure real-world, high-impact behaviour instead of proxy attitudes; attend to cross-country and cross-domain heterogeneity; integrate individual- and system-level analysis; treat industry collaborations warily; and organise researchers into resilient networks. Read against the rest of the papers filed here, this piece works less as a stand-alone contribution than as a diagnostic lens — several of its critiques (especially the over-reliance on cognitive/attitudinal outcomes and the neglect of real-world behaviour) are precisely what the empirical papers below either confirm or attempt to remedy.
Message Content versus Message Form
Two papers interrogate what, exactly, makes a climate message persuasive or a piece of misinformation credible. Voelkel2026-lc is the most ambitious attempt in the set to settle this at scale: a registered-report megastudy of the ten most-cited climate messaging strategies finds that six shift attitudinal outcomes, but only by a few percentage points, with almost no partisan heterogeneity and — crucially — no effect on actual pro-environmental donations. This is close to a direct empirical instantiation of Spampatti et al.’s worry that the field chases attitudinal movement while behaviour stays untouched, and it undercuts the widespread assumption that audience-tailored (e.g., Republican-specific) framing is necessary. Lieu2025-nl approaches the mirror-image question from the misinformation side, crossing the CARDS taxonomy of contrarian content against the FLICC taxonomy of rhetorical fallacies. Its central finding — that logical form (the fallacy) barely matters while topical content does, with claims that attack climate solutions rated most credible and most polarising — suggests that debunking and prebunking efforts organised purely around “spotting the fallacy” may be misdirected; audiences respond to what is being claimed, not how it is logically malformed. Together these two papers push toward a content-first, behaviour-skeptical view of message effects that resonates with Spampatti’s call for closer attention to what actually moves people.
Scaling Inoculation: From Static Content to Adaptive Dialogue
A second cluster follows inoculation theory as it moves from lab demonstrations toward ecologically valid and increasingly adaptive delivery formats. van-der-Linden2026-jt is the field-scale proof of concept: a 19-second prebunking ad shown to nearly 400,000 Instagram users produces a 21-point gain in fearmongering detection that persists five months later and drives measurably more information-seeking — a rare case in this literature of an intervention tested at real platform scale with a durable, non-laboratory outcome, directly answering Spampatti’s call for real-world behavioural measurement. Szabo2026-rd pushes the same inoculation logic toward personalisation, comparing traditional Reading and Writing inoculation formats against a novel “Conversational Inoculation” delivered by an LLM chatbot; when individual baseline susceptibility is accounted for, the chatbot format outperforms both static formats, and qualitative analysis attributes this to the agent’s capacity to adapt to the user’s attitudes, build trust, and scaffold independent reasoning rather than simply supplying stronger arguments. Read together, these two papers trace an arc from broadcast-style prebunking (effective, scalable, but one-size-fits-all) to conversational, adaptive inoculation (more effective per-individual, but harder to deploy at Instagram-level scale) — a tension the field will need to resolve.
Conversational AI as a General-Purpose Belief-Change Engine
The most striking convergence in this set concerns dialogue-based AI interventions applied directly to belief change rather than pre-exposure inoculation. Costello2024-bg is the foundational result: personalised, evidence-based three-round conversations with GPT-4 reduce conspiracy belief by roughly 20%, durably, generalising even to unrelated conspiracies — directly challenging motivational accounts that treat conspiracy belief as impervious to counter-evidence. Kotz2026-lk extends this logic beyond conspiracy theories to structurally different contested domains (climate, vaccination, inequality), finding that brief AI dialogues shift both beliefs and concrete policy support, with the largest gains among initially skeptical participants and trust in science as the most consistent moderator — suggesting the mechanism is less about a peculiar psychology of conspiracists and more about the general power of tailored, high-quality evidence delivered conversationally. Dubey2026-bl complicates this picture from the chatbot-design side: testing “balanced” chatbots that mix mainstream and conspiratorial perspectives on climate, it finds — against the authors’ own hypothesis — that high-conspiracy-belief users trust and accept such chatbots more than low-conspiracy users, implying that the reactance-and-distrust story often assumed in this literature may not hold once the source presents itself as balanced rather than corrective. Read as a sequence, these three papers suggest AI dialogue systems are converging as a genuinely general-purpose tool for engaging skeptical audiences, but that the design choice between “corrective” (Costello, Kotz) and “balanced” (Dubey) framing may matter for who is willing to engage at all.
The Vulnerability Landscape: Deepfakes, Emotion, and Social Meaning-Making
A final cluster examines how misinformation’s psychological impact is shaped by modality, context, and affect rather than content alone. Choi2026-bz shows that difficulty authenticating one piece of (deep)fake content carries over to affect confidence and susceptibility toward later fake content in the same modality — a finding suggesting that repeated deepfake exposure has cumulative, format-specific effects on audience epistemic vigilance. Wack2026-bt extends the deepfake question outward from the individual to the social environment, showing in a large Kenyan field experiment that the political impact of a deepfake is substantially determined by the surrounding comments — credulous reactions amplify harm, skeptical ones partially neutralise it, and partisanship governs which direction commenters push — reframing deepfake impact as a matter of “collective sensemaking” rather than fixed, individual-level credibility judgment. Xue2025-bp rounds out this cluster by turning the emotion question back onto the correctives themselves, finding that fact-checking posts — ostensibly the genre of dispassionate correction — are often highly emotional, and that this emotionality shapes engagement and sentiment toward fact-checked targets, undercutting the normative assumption that debunking succeeds through neutral information provision alone.
Methodological Infrastructure
Arminio2025-tw sits slightly apart as a methods contribution rather than an intervention study, but it is directly relevant infrastructure for this whole topic: by using vision-language models to generate connotative textual descriptions of climate-related social media images (rather than relying on CNN-based denotative clustering), it offers a scalable way to characterise the visual and memetic content — including “flexibly deceptive” or symbolically loaded imagery — that several papers above (notably Wack’s deepfakes and Xue’s emotionally charged fact-checks) treat as central to how climate and conspiracy content actually circulates online.