The evidentiary pushback against alarm
The starting provocation for this cluster is Budak2024-ef, which names three misperceptions that have become conventional wisdom in journalism and policy: that problematic content is pervasive, that algorithms are its principal engine, and that social media is a primary cause of societal harm. Everything else filed here can be read as either supplying the evidence for that corrective, complicating it, or extending it into new domains (AI-generated content, health information, visual media, cross-national contexts). Lazer2018-mm is the field’s own earlier admission that it did not yet have this evidence — a 2018 call for a “science of fake news” whose gaps the subsequent seven years of large-scale data partnerships have partially filled. Frischlich2025-vn adds a meta-level argument: single analogies (virus, weapon, pollutant) each capture one level of a multilevel system and none is adequate on its own, which is itself a caution against the totalizing alarm that motivates this whole literature.
Concentration, not saturation
The most consistent empirical regularity across the corpus is that exposure to and sharing of misinformation is rare in the aggregate but extremely concentrated in a thin slice of the population. Allcott2017-yz and Grinberg2019-ua establish this for the 2016 US election — the average voter saw roughly one fake article, and 1% of Twitter users accounted for 80% of exposures. Guess2019-ym and Guess2021-ym extend the pattern to Facebook sharing, isolating age (more than ideology) as the most robust predictor. Guess2020-rr shows the “deep but narrow” echo-chamber structure explicitly: 62% of untrustworthy-site visits came from the most conservative fifth of the population. The same architecture reappears in newer, methodologically heavier studies: Eady2023-xg finds Russian IRA exposure on Twitter dwarfed by domestic news and concentrated among strong Republicans already predisposed to Trump; Appel2026-qr finds three of forty-nine deceptive networks driving 80% of Facebook reach, with most of that reach arriving indirectly via reshares rather than direct posting; Bergeron-Boutin2026-yh finds low average, highly concentrated exposure to untrustworthy Facebook/Instagram sources; and Lyons2026-ca finds the same skew in an entirely different domain — low-credibility health content — where the top 10% of users account for 77% of exposure, concentrated among older adults with conspiracist worldviews. Rossi2023-847d5a9f replicates the concentration-by-age finding for France, Germany, and Italy while showing that a rising share of untrustworthy URLs does not translate into a rising share of views, a dissociation that recurs throughout the literature. Allcott2019-gn documents an actual temporal decline in the relative weight of Facebook misinformation after 2016, even as Twitter’s did not fall — evidence that platform-specific interventions can matter without requiring wholesale re-narration of the “infodemic.”
What large field experiments do not find
A parallel strand, mostly drawn from the Meta-academic 2020 US Election Studies collaboration, uses randomized manipulation of real feeds rather than correlational exposure measures, and consistently returns null or minimal effects on downstream attitudes. Guess2023-ai shows that stripping reshares from feeds sharply cuts exposure to untrustworthy content and partisan news but leaves polarization and most attitudes untouched. Guess2023-ur finds the same for replacing the algorithmic feed with reverse chronological ordering: large experiential effects, no detectable political effects. Nyhan2023-gb shows that like-minded content is prevalent but reducing it by a third does nothing to affective polarization or belief in falsehoods. Allcott2025-jb finds that removing political ads entirely for six weeks before the 2020 election changed nothing measurable. Allen2024-av complicates the “misinformation causes vaccine hesitancy” narrative not by finding no effect, but by showing the effect is real yet swamped: fact-checked misinformation is more persuasive per view but reaches so few people that unflagged, technically-true “vaccine-skeptical” content from mainstream outlets does forty-six times more aggregate damage — a finding that simultaneously vindicates and redirects concern about harm. Ventura2025-sw pushes the experimental logic into a new venue, a WhatsApp deactivation study in Brazil, testing whether private messaging (largely invisible to the feed-centric literature) behaves differently — an important complement given how much of this evidence base is Facebook/Twitter-specific. Berlinski2023-rj is a partial exception, showing that elite fraud claims do dent election confidence among receptive partisans and that fact-checks fail to repair it — a reminder that “limited effects” is not the same as “no effects,” particularly for elite-sourced, identity-congruent claims rather than diffuse platform-level exposure.
Measurement artifacts: why the alarming numbers can be wrong, and why the reassuring ones can be too
Several papers turn the recalibration inward, arguing that both alarmist and reassuring estimates depend heavily on measurement choices. Allen2021-ai shows that Facebook’s own public URL-sharing dataset overstates fake-news prevalence roughly fourfold because privacy-preserving thresholds censor content that isn’t shared enough — a caution that big data can still be biased data. Goel2025-iq and Green2025-ap both attack the field’s default operationalization of misinformation as domain-level untrustworthiness, showing instead that mainstream, factually accurate articles are routinely repurposed into misleading narratives, and that individual stories from “moderate” domains reach highly divergent partisan audiences that domain-level scores erase entirely — meaning some domain-based reassurance may itself be an artifact of the wrong unit of analysis. Pante2025-pq applies the same corrective logic to influence-operations research, arguing that earlier claims of coordinated inter-state collaboration do not survive rigorous coordination-detection methods and proper control datasets. Renault2025-uh uses the crowdsourced Community Notes system as a bias-resistant instrument and finds Republicans are flagged 2.3 times more often than Democrats — evidence that at least one asymmetry survives even when the measurement method is redesigned specifically to rule out fact-checker bias.
Echo chambers and the production–consumption illusion
A companion line of research recalibrates not exposure to false content per se but the broader diagnosis that social media algorithmically imprisons people in ideological bubbles. Bakshy2015-rn and Flaxman2016-lm are the foundational demonstrations that algorithms and social/search channels modestly increase segregation relative to friend networks while also increasing exposure to opposing views, with the bulk of consumption occurring through direct, mainstream browsing rather than curated feeds. Messing2014-jc shows experimentally that social endorsement cues wash out partisan source selectivity almost entirely. Oswald2025-km and Schulz2026-ts extend the recalibration to the perception layer itself: what looks like a polarized public sphere is largely an artifact of a small, vocal, ideologically extreme minority who post, while the silent majority who merely consume are invisible in exactly the data researchers and citizens use to infer public opinion — a point that reframes many “misinformation is everywhere” claims as claims about the visible tip of an unrepresentative iceberg.
Broadening what “misinformation” actually looks like
Several papers argue the alarmist frame is not just exaggerated in scale but wrongly specified in kind. Hourigan2026-oc finds that everyday, user-flagged misinformation in Australia is overwhelmingly textual, sourced from mainstream and alternative news (not fringe actors), and concentrated in mundane domains like business and economics rather than the “hot button” topics the literature fixates on — and that only a small fraction of flagged content is independently verifiable as false at all, suggesting perceived and actual misinformation diverge. Tsfati2020-uo makes the structurally similar point that mainstream media, in the act of debunking, may be the largest actual disseminator of fake news narratives, since direct traffic to fake sites is so small. Nenno2025-xa shows that the news-values signature of flagged misinformation, while real, is small in magnitude and inconsistent across WEIRD and non-WEIRD countries, cautioning against universalizing US/Brazil-derived intuitions about what “fake news” looks like. Two papers complicate rather than confirm the reassurance: Yang2023-cg argues that link-based measures have drastically undercounted misinformation by ignoring images, finding 20%+ of political image posts on Facebook misleading (with an 8:1 partisan skew), and Vincent_undated-re reports rising, substantial misinformation prevalence on TikTok in Europe — both a reminder that “limited and concentrated” findings from Twitter/Facebook text-and-URL studies may not generalize to newer platforms and modalities, and that recalibration must keep pace with where audiences actually move. Mosleh2024-op similarly warns that engagement patterns diverge across seven platforms, so single-platform (usually Twitter) conclusions are an unstable basis for ecosystem-level claims. Zooming out further, Allen2020-nj reframes the entire debate by showing that news of any kind is a sliver of Americans’ media diet and fake news is roughly 0.15% of it — suggesting that if misinformedness is a real problem, its causes more plausibly lie in ordinary news bias, framing, and avoidance than in engineered falsehood, a claim Vosoughi2018-at partially complicates by showing that, within the population of news that does circulate on Twitter, verified falsehoods diffuse markedly faster and farther than truths — a finding about differential virality that coexists uneasily with claims about low absolute prevalence.
Cognition over conspiracy, and the limits of correction
Two psychology-oriented papers relocate the explanatory burden from malicious design to ordinary cognitive limits. Pennycook2021-jq argues that inattention and weak analytic reasoning, not motivated partisan reasoning, best explain both belief in and sharing of false news, and that the sharing–belief gap is a bigger problem than deliberate deception. Lewandowsky2012-vn documents why corrections so often fail or backfire, tempering optimism about fact-checking as a fix while reinforcing that the psychological mechanisms sustaining misinformation are general features of memory and worldview rather than social-media-specific pathologies — a point consonant with Berlinski2023-rj’s finding that fact-checks fail to blunt partisan-congenial fraud claims.
New frontiers: does the recalibration hold for generative AI?
The newest papers test whether the reassuring pattern extends to AI-generated content, where alarm is currently peaking. Hackenburg2025-dj finds that conversational AI’s persuasive power is real but driven by mundane levers (information density, post-training) rather than exotic personalization/microtargeting, and crucially that persuasiveness trades off against accuracy — a nuanced, neither-alarmist-nor-dismissive result. Voelkel2026-lc shows, in the adjacent domain of climate persuasion, that even the most-cited messaging strategies move attitudes only marginally and not behavior at all, undercutting assumptions about the potency of strategic communication generally. Gilardi2026-hw finds AI-generated news is judged no differently from human-written news in blind quality ratings, with disclosure producing only curiosity-driven, short-lived engagement rather than durable acceptance or rejection. Hameleers2026-mc directly challenges deepfake alarmism by showing AI-generated still images are not more credible or engaging than text, and that even video disinformation’s edge is context-bound, contingent on the availability of real footage to recontextualize — suggesting low-tech decontextualization, not generative AI per se, remains the more potent threat. Wack2026-bt relocates deepfake impact away from the artifact itself toward the surrounding social commentary, arguing effects are “collectively negotiated” rather than fixed at the point of AI production. Together these suggest the same recalibrating logic that applied to text-based misinformation — modest, mechanism-specific, context-dependent effects rather than uniquely terrifying new technology — is beginning to apply to generative AI as well, though the evidence base is far thinner.
Context: trust, not just falsehood
Finally, Fletcher2026-lv supplies the longue durée backdrop against which all of this should be read: trust in news has been declining globally since the 1980s, predating the internet and accelerating only around 2000, and the decline is neither universal nor primarily digital in origin — even authoritarian contexts show trust rising. This chapter is a useful closing corrective to the entire cluster’s implicit periodization: much of what gets blamed on “the current misinformation crisis” is continuous with much longer-run dynamics in institutional trust that a purely platform-centric or exposure-centric literature risks mistaking for a novel, acute pathology.