Recalibrating misinformation exposure and harms narratives

The founding move of this literature is empirical deflation. A cluster of large-scale, trace-data studies of the 2016 and 2020 U.S. elections established the now-canonical pattern: exposure to “fake news” and untrustworthy sources is real but small in aggregate and extremely concentrated among a narrow slice of highly engaged, disproportionately older and more conservative users. Grinberg2019-ua and Guess2020-rr showed that fake-news exposure and sharing on Twitter and the open web were dominated by a tiny fraction of “supersharers,” with most people’s news diets still overwhelmingly mainstream; Allcott2017-yz used an economic framing to argue the average voter saw and remembered roughly one fake story, far too little to be decisive; Guess2019-ym and Guess2021-ym extended this to Facebook sharing and viewing, finding fake-news sharing rare overall but robustly concentrated among older users; and Allen2020-nj situated all of this within the full media ecosystem, showing that news itself—let alone fake news—is a vanishingly small share of Americans’ total media diet dominated by television. Vosoughi2018-at is the partial counterweight here, documenting that falsehood does diffuse farther and faster than truth on Twitter, a finding this literature has had to reconcile with the concentration results rather than simply dismiss. Allcott2019-gn tracks the trajectory after 2016, showing Facebook engagement with fake news falling sharply relative to Twitter, suggesting platform-specific mitigation rather than an ever-worsening crisis.

Extending concentration findings to new domains and platforms. This recalibration is not frozen in 2016–2020; subsequent work generalizes the concentration logic to new content types, populations, and geographies. Lyons2026-ca finds the same skewed distribution for low-credibility health content, concentrated among older, more conspiracist-minded users reached via cross-domain referral rather than search or social media. Eady2023-xg applies the same lens to the Russian IRA’s 2016 Twitter campaign, finding exposure overshadowed by domestic media by an order of magnitude and concentrated among already-committed partisans least in need of persuasion. Appel2026-qr extends the method to 49 Meta-identified deceptive networks around 2020, again finding extreme concentration (three networks drove most reach) and showing that correlational links to downstream beliefs evaporate once pre-exposure characteristics are controlled. Bergeron-Boutin2026-yh and Rossi2023-847d5a9f replicate the concentration-with-limited-exposure pattern at platform scale for Facebook/Instagram in 2020 and across Germany, France, and Italy, respectively, while Gonzalez-Bailon2024-rq and Mosleh2024-op push the descriptive project further into diffusion-tree analysis and seven-platform comparison, underscoring that single-platform (especially Twitter-centric) generalizations are unreliable. Oswald2025-km supplies the structural explanation for why visible discourse so badly misrepresents underlying exposure and opinion: a small minority of highly active users produce almost all observable content, a “production-consumption gap” that Schulz2026-ts shows manufactures false polarization through selective posting by ideological extremes, a dynamic further complicated by Green2025-ap’s demonstration that story-level “curation bubbles” make individual articles far more partisan-coded than their source domain’s aggregate rating suggests.

Causal tests confirm limited, not absent, effects. A second wave of studies moves from exposure measurement to causal identification, generally reinforcing the “limited harms” conclusion while refining its boundaries. Large field experiments on Facebook—Guess2023-ai (reshares), Nyhan2023-gb (like-minded sources), and Allcott2025-jb (political advertising)—each find that substantially altering users’ information diets produces no detectable shift in polarization or most political attitudes, despite real changes in content mix. Bakshy2015-rn and Flaxman2016-lm anticipate this with evidence that algorithmic curation plays a smaller role in ideological segregation than individual choice and direct browsing, respectively, complicating simple “filter bubble” stories. Bail2018-fk is the pointed counter-case, showing that deliberate cross-partisan exposure can backfire into greater polarization—an important caution against assuming all interventions net to zero. On health and civic harms specifically, Allen2024-av delivers perhaps the most consequential recalibration: fact-check-flagged misinformation was more persuasive per view but reached so few people that unflagged, mainstream “vaccine-skeptical” content had a far larger aggregate effect on hesitancy—relocating the locus of harm from fringe falsehoods to ambiguous mainstream framing. Bollenbacher2026-vz offers a rare causal estimate running the other direction, modeling antivaccine tweets’ epidemiological downstream effects, while Berlinski2023-rj shows that even minimal exposure to elite fraud claims measurably erodes confidence in elections among receptive partisans—evidence that some narrow, elite-driven effects remain real even as diffuse platform-level effects do not. Ventura2025-sw and Brown2026-br extend causal designs to under-studied terrain (WhatsApp deactivation in Brazil; YouTube’s recommendation algorithm), with Brown’s audit finding strong “rabbit hole” dynamics but no support for algorithmic radicalization per se—again isolating user choice as the more potent driver than algorithmic push.

Methodological corrections underpinning the recalibration. Much of this empirical reassessment rests on exposing flaws in how misinformation has been measured. Allen2021-ai shows that Facebook’s own privacy-preserving URL dataset systematically overstates fake-news prevalence by 4x due to a public-sharing censoring threshold, a cautionary methodological note that other scale claims in the field should heed. Luhring2025-od audits NewsGuard, the workhorse source-trustworthiness database underlying much of this research, finding its continuous scores reasonably stable but binary classifications prone to threshold-driven distortion. Pante2025-pq similarly re-examines claims of coordinated inter-state influence operations, arguing prior findings used weak coordination-detection methods lacking proper control comparisons. Goel2025-iq identifies a blind spot in domain-level measurement itself: users repurposing factually accurate mainstream content into misleading narratives, a phenomenon invisible to any approach that classifies misinformation by source alone. Holt2026-zq turns the critical lens on crowdsourced alternatives to fact-checking, finding that Facebook users’ “False News” flags mostly target polarizing-but-true content they disagree with rather than falsehoods—a sobering diagnostic for the post-fact-checking era—while Renault2025-uh shows that even under X’s Community Notes bridging algorithm, Republican-authored posts are flagged as misleading at over twice the rate of Democratic ones, undercutting claims that crowdsourcing alone resolves partisan asymmetry concerns.

Broadening the definition: misinformation beyond “fake news.” A parallel thread argues the dominant “fake news domain” paradigm is not just imprecisely measured but conceptually too narrow. Hourigan2026-oc’s diary study finds everyday perceived misinformation is overwhelmingly textual, mainstream-sourced, and concerned with mundane topics like economics rather than exotic conspiracy content—misinformation-as-clickbait rather than misinformation-as-fabrication. Yang2023-cg makes an analogous corrective for images, arguing text/link-based studies dramatically undercount visual misinformation, which shows stark partisan asymmetry but little engagement advantage. Suau_Martinez2026-lv and Prochaska2026-pr reconceptualize disinformation as recurring narratives and rumor “grammars” that circulate across the hybrid media ecosystem rather than discrete false claims, while De_Leon2025-qn proposes “networked exposure” to capture how ordinary, true content becomes weaponized within conspiracy-aligned networks. Together these push against both over-narrow operationalizations and over-broad alarm, insisting the phenomenon is at once more mundane and more structurally embedded than “fake news” implies.

Explicit syntheses of the recalibration thesis. Several papers state the corrective argument directly rather than through new data. Budak2024-ef is the clearest programmatic statement: public discourse on misinformation has diverged from the evidence on three counts—prevalence, algorithmic causation, and societal harm—echoing Lazer2018-mm’s earlier call for a rigorous “science of fake news” built on data the field still largely lacked. Tsfati2020-uo complicates the picture by relocating the amplification problem onto mainstream media’s own paradoxical role in disseminating and legitimizing fake news through corrective coverage. Pennycook2021-jq supplies the underlying psychological mechanism for much of the concentration/limited-effects story: poor truth discernment is driven by inattention and low analytic reasoning rather than motivated partisanship, and the sharing-belief gap is itself largely inattentive rather than deliberate. Messing2014-jc offers an early structural explanation for why online environments needn’t default to partisan silos: social endorsement cues can override source-partisan cues in news selection, increasing cross-cutting exposure. Fletcher2026-lv zooms out furthest, showing that the decline in trust in news is real but predates the internet, is far from universal, and inverts in unfree-press contexts—cautioning against attributing a long-run, globally uneven trend solely to social media or misinformation.

Interventions and what little still moves the needle. A smaller set of studies asks what, if anything, reliably improves outcomes, generally finding modest but real effects for carefully targeted interventions. van-der-Linden2026-jt demonstrates that brief inoculation ads scaled on Instagram durably improve recognition of fearmongering; Voelkel2026-lc’s climate-messaging megastudy finds several widely-cited messages do move attitudes, but only by a few percentage points and not behavior; Lieu2025-nl shows content category, not rhetorical fallacy type, drives perceived veracity of climate misinformation; and Lin2025-xp, more surprisingly, finds LLM dialogue persuasion on voting substantially exceeds traditional ad effects, achieved mainly through evidence rather than manipulation—though with a troubling partisan asymmetry in AI-generated factual accuracy.

New frontiers: does generative AI reset the calculus? The newest papers ask whether AI-generated content threatens to upend the empirically-grounded, de-escalated picture built over the preceding decade. Emilio2026-ik and Dierickx2026-tw argue conceptually that synthetic media and “emergent facts” require new epistemic categories beyond prior fact-checking frameworks, while Triedman2025-uy’s empirical audit of Grokipedia finds concrete evidence of ideologically skewed, low-quality sourcing diverging from Wikipedia on controversial topics—an early empirical test case for whether AI-mediated knowledge infrastructures reproduce or amplify the asymmetries documented elsewhere in this literature. Rogers2026-zz bridges old and new by tracing how Meta’s shrinking data-access regimes and shifting terminology (“fake news” → “borderline content”) have made the underlying problem harder to measure even as its right-leaning skew persists. Scalco2026-bd and Vincent_undated-re represent the measurement infrastructure this whole research program depends on going forward: quantifying “information voids” and tracking prevalence across platforms, ensuring that as the information environment is reshaped by AI, synthetic media, and new platforms, claims about exposure and harm remain anchored in data rather than alarm.