Cognitive Foundations: Why Corrections Fail and What Discernment Requires
The structure note’s deepest roots lie in cognitive psychology’s account of why false beliefs persist. Lewandowsky2012-vn establishes the foundational puzzle: retractions rarely erase misinformation and can even backfire, because belief relies on fluency, coherence, and worldview-congruence rather than simple logical updating. This sets an agenda that much of the later work answers directly — only pre-exposure warnings, repetition of the correction, and alternative causal narratives reliably help. Pennycook2021-jq complicates the partisan-motivated-reasoning narrative that might otherwise seem to follow from worldview-driven resistance, arguing instead that poor truth discernment stems mainly from lazy, non-analytic thinking and inattention rather than directional motivation — a reframing with direct consequences for intervention design, since it implies that prompting reflection (not just providing more facts) should reduce susceptibility. Choi2026-bz extends this cognitive lens into the deepfake era, showing that the experience of struggling to authenticate one piece of fake content carries over to shape confidence and vulnerability when encountering subsequent fake content in the same modality — a finding that suggests metacognitive calibration, not just factual knowledge, is itself manipulable and fragile.
Testing Message Content: What Climate Framings Actually Move
A cluster of papers interrogates specific message content rather than general cognitive mechanisms. Voelkel2026-lc delivers a sobering corrective to the messaging literature: a large registered-report megastudy of the ten most-cited climate communication strategies finds only small, largely non-partisan effects on attitudes and essentially no effect on costly behavior (donations), directly challenging assumptions — shared by much prior work — that audience-tailored moral framing is necessary or that attitude shifts translate into behavior change. Lieu2025-bp… rather, Lieu2025-nl shows that for misinformation content specifically, the logical fallacy used matters far less than the substantive claim category, with attacks on climate solutions perceived as most credible and most polarizing — suggesting debunking resources should be triaged by content type rather than by rhetorical technique. Xue2025-bp pushes on the presentation side of correction itself, documenting that fact-checking posts are saturated with emotional language despite fact-checking’s normative claim to objectivity, and that this emotionality shapes engagement and sentiment — raising the uncomfortable possibility that corrective communication is already borrowing from the persuasive toolkit it is meant to counteract. Bravo2026-ue zooms out to the visual register, surveying how climate change is depicted in media imagery and finding the field itself skewed toward Western, traditional-media case studies, flagging a gap in understanding how visual framing — as opposed to verbal framing — shapes public engagement.
Inoculation and Prebunking: From Lab to Feed
A second arc traces inoculation theory’s migration from controlled lab paradigms toward real-world deployment. van-der-Linden2026-jt provides a landmark field demonstration: a 19-second prebunking ad deployed at scale on Instagram improved recognition of emotional manipulation by 21 points, with effects persisting five months and driving increased information-seeking — strong evidence that inoculation can work ecologically, cheaply, and at scale, not merely in survey experiments. Szabo2026-rd pushes inoculation into a more interactive register, introducing “Conversational Inoculation” via an LLM chatbot and finding that dialogic refutation outperforms static reading/writing inoculation once individual baseline susceptibility is controlled, with qualitative analysis pointing to adaptability, trust-building, and minimized friction as the mechanisms that make conversation superior to static exposure. Together these two papers mark a shift from “deliver the inoculating message” toward “calibrate the inoculating interaction,” a theme picked up more fully in the conversational-AI cluster below.
Conversational AI as a Persuasive and Corrective Technology
The most concentrated empirical arc concerns dialogue-based AI interventions. Costello2024-bg is the pivotal result: brief, personalized, evidence-based GPT-4 conversations durably cut conspiracy belief by roughly 20%, generalizing to unrelated conspiracies and persisting two months — directly contradicting motivational accounts that treat conspiracy belief as impervious to counter-evidence, and instead supporting an analytic-reasoning account consonant with Pennycook2021-jq’s broader claims about reflection. Kotz2026-lk extends this logic across domains (climate, vaccination, inequality) with GPT-5, showing that AI dialogue shifts both beliefs and policy support, is most effective among skeptics rather than the already-convinced, and is moderated by trust in science rather than political identity — a finding that, together with Voelkel2026-lc’s null partisan-heterogeneity result, suggests contested climate attitudes may be less rigidly identity-locked than audience-segmentation theories assume. Dubey2026-bl complicates the trust story, however: when the chatbot is framed as deliberately “balanced” between mainstream and conspiratorial perspectives, it is higher-conspiracy users who trust and accept it more, not less — an unexpected reversal that cautions against assuming conspiracy believers are uniformly defensive toward chatbot-mediated correction, and that sits in productive tension with Costello and Kotz’s more straightforwardly corrective designs. Read together, Szabo2026-rd, Costello2024-bg, Kotz2026-lk, and Dubey2026-bl trace a maturing but unsettled picture of conversational AI: powerful for shifting belief and policy attitudes, but highly sensitive to framing, perceived neutrality, and the user’s prior trust architecture.
Beyond Claim-Level Correction: Narratives and Arguments
Ramos2026-qo argues that the entire claim-correction paradigm underlying fact-checking is structurally insufficient for crisis disinformation, because false narratives mutate and regroup faster than individual claims can be debunked. Drawing on the 2024 Valencia flood conspiracy, the authors propose “argument-checking” — targeting the fallacious premises and warrants that hold a narrative together — as a complement to fact-checking, echoing Lieu2025-nl’s content-over-fallacy finding but inverting its implication: where Lieu suggests fallacy type barely matters for audience judgments, Ramos argues that exposing fallacious argument structure is precisely what gives corrections transferability across a narrative’s many claim-level mutations.
Synthesis and Methodological Self-Reflection
Two papers step back from individual interventions to assess the field itself. Spampatti2026-kx imports lessons from the more mature climate-psychology literature, arguing that misinformation research over-relies on cognitive proxies, underuses real-world behavioral outcomes, and needs to integrate individual- and system-level analysis — a critique that retrospectively frames Voelkel2026-lc’s behavioral null results and van-der-Linden2026-jt’s field-based design as exemplars of the rigor the field should aspire to. Bravo2026-ue’s review performs an analogous diagnostic function for visual climate communication, flagging geographic and methodological imbalances. Taken as a whole, the collection traces an arc from explaining why misinformation sticks (Lewandowsky2012-vn, Pennycook2021-jq), through testing which messages and formats move attitudes (Voelkel2026-lc, Lieu2025-nl, Xue2025-bp, Choi2026-bz), to scaling and personalizing interventions via inoculation and AI dialogue (van-der-Linden2026-jt, Szabo2026-rd, Costello2024-bg, Kotz2026-lk, Dubey2026-bl), and finally toward structural critiques of both the interventions themselves and the field’s evidentiary standards (Ramos2026-qo, Spampatti2026-kx, Bravo2026-ue).