The arc of inquiry: from echo chambers to algorithmic curation to institutional capture
The papers filed here trace a shift in how partisan media effects have been studied: from early concerns about selective exposure and “echo chambers” in user choice, through a wave of large-scale platform-collaborative experiments that complicated that story, to a current generation of work locating polarization in algorithmic governance, elite strategy, and cross-platform institutional dynamics. Read together, they form something like a thesis-antithesis-synthesis arc.
The echo chamber thesis and its early complications
The foundational claim—that homophilous networks and selective exposure segregate partisans into mutually isolated information worlds—is established by Barbera2015-je’s ideal-point estimation method and Del-Vicario2016-uj’s percolation model of homogeneous cascades. But even within this founding generation, the picture was never simple. Bakshy2015-rn showed that friend networks on Facebook contain substantial cross-cutting ties, with individual click choices suppressing diverse exposure more than the News Feed algorithm itself. Flaxman2016-lm found modest, bidirectional effects of search and social media on segregation. Barbera2015-fw demonstrated that polarization is topic- and time-dependent, with nonpolitical events producing genuine “national conversation,” and crucially identified an ideological asymmetry: conservatives engage in less cross-cutting communication than liberals. Messing2014-jc offered a countervailing mechanism—social endorsement cues can override partisan source cues in selection decisions. This early literature set up two recurring tensions that the rest of the corpus keeps returning to: asymmetry (is polarization symmetric across left/right, or driven disproportionately by one side?) and attribution (is the fragmentation caused by algorithms, by user choice, or by the entangled interaction of both?).
The null-results turn: platform experiments and their limits
A pivotal cluster of papers—largely emerging from the 2020 Facebook/Instagram election studies—pushes back hard against folk theories of algorithmic radicalization. Guess2023-ur found that replacing the algorithmic feed with a chronological one altered engagement and content mix dramatically but produced no detectable change in polarization, knowledge, or participation. Guess2023-ai found the same null pattern for removing reshares, despite reshares driving much of the political-news and misinformation exposure. Nyhan2023-gb showed like-minded content is prevalent but reducing it by a third moved no attitudinal needle. Gonzalez-Bailon2023-uy complicates this reassurance from within the same data ecosystem, finding that segregation increases sharply along the “funnel of engagement” and is markedly asymmetric—conservative audiences occupy a uniquely homogeneous corner where most fact-checked misinformation concentrates, echoing Benkler2018-lw’s earlier argument that the crisis is institutional and asymmetric rather than a generic technological pathology, and Yang2023-cg’s finding of a 5-8x partisan asymmetry in image-based misinformation. Bail2018-fk’s field experiment adds a further wrinkle: exposure to opposing views can backfire, polarizing Republicans further—undermining any simple remedy based on cross-cutting exposure, a worry later formalized by van-Eck2026-xg’s review showing depolarization recommendations are rarely evidence-based.
More recent work reopens the algorithmic question with sharper causal designs. Gauthier2025—properly Gauthier2026-iq—finds that X’s “For You” algorithm does shift attitudes rightward, resolving the apparent contradiction with the Meta null results by identifying a persistence mechanism: the algorithm reshapes who users follow, an effect that survives even after the algorithm is switched off. Efstdoulis—Efstratiou2025-gs—shows that apparent right-leaning algorithmic amplification on Twitter is substantially explained by agitating content and proximity to Musk rather than ideology per se, while documenting a new centralizing force: the platform owner himself. Brady2026-ln reframes the mechanism again, proposing that engagement-based feeds distort perceived social norms about political discourse rather than attitudes directly, and shows a “diversified extremity” algorithm can mitigate this. Together these papers suggest the null results of 2023 may reflect limitations of the chronological-feed manipulation specifically, not of algorithmic curation in general.
Measurement itself as a site of contestation
A significant methodological strand interrogates how partisan slant and polarization are operationalized, arguing that measurement choices are not neutral. Hartmann2025-px’s systematic review of 129 echo-chamber studies argues that disagreement in the field stems largely from heterogeneous operationalizations rather than divergent realities—a diagnosis Schemer2026-mh substantiates empirically via specification-curve analysis, showing that partisan-media-polarization associations vary enormously in magnitude (though not direction) depending on how media slant is scored. Green2025-ap extends this critique to the object of measurement itself: domain-level partisan audience scores mask substantial story-level heterogeneity, especially for “moderate” outlets, because networked curation unbundles and rebundles individual stories rather than whole sources—a point Gaisbauer2025-px (filed as Gaisbauer2025-px)… rather, Gaisbauer2025-px is cited elsewhere; the relevant methodological companion here is Gaisbauer2025-px. Eady2025-vm offers a unifying measurement advance, placing media, politicians, and ordinary users on a common ideological scale via URL-sharing behavior, while Lai2024-to extends latent ideology estimation to video content via Reddit sharing patterns. Giglietto2026-632ef967 shows that the very sharing-to-viewership relationship is nonlinear and shaped by active, time-varying platform governance (the 2020 “break the glass” interventions)—meaning Facebook is not a neutral conduit whose curation can be treated as a stable background condition, a point reinforced longitudinally by McNally2025-dn’s detection of algorithmic ranking effects via time-series anomaly analysis, and by Giglietto2025-1765bb4f’s finding that Meta’s political-content-reduction policy suppressed mainstream politicians’ reach while extremist accounts compensated through sheer volume—exposing both asymmetric effects and transparency deficits in platform self-reporting.
Elite strategy, institutional capture, and the “logic of connective faction”
Running parallel to the platform-centric literature is an institutionalist strand asking how elites, parties, and legacy media interact with partisan digital ecosystems. Chadwick2013-ns and Chadwick2015-ii furnish the foundational “hybrid media system” framework—old and new media logics interacting through adaptation rather than displacement—which Bennett2025-xs updates explicitly to explain democratic backsliding, arguing that “digital surrogate organizations” pull conservative parties toward illiberalism in ways that purely cognitive/disinformation accounts miss. Knupfer2025-vt operationalizes this via the “logic of connective faction,” showing how Republican adoption of Critical Race Theory as a talking point tracked ideological extremity and reward from right-wing digital networks. Tornberg2025-ir offers comparative confirmation across 26 countries: neither populism nor right-wing ideology alone predicts elite misinformation sharing, but their conjunction—radical-right populism—does, while Darius2026-xl’s cross-platform EU study shows PRR parties enjoy disproportionate audience support (especially on TikTok) independent of posting volume, suggesting resonance rather than mere supply. Balluff2026-bv documents a concrete mechanism of capture—government-advertising-for-coverage exchange in Austria—while Balluff2026-ev complicates assumptions about alternative media’s independence by showing substantial entity overlap with legacy outlets. Copland2025-em and Heft2021-ky both trace transnational right-wing media infrastructure (Sky News Australia, European/US hyperlink networks), with Heft’s counterintuitive finding that these networks’ shared reference points are dominated by legacy media, undercutting the “insulated sphere” thesis. Iris2026-pg’s cross-national study of the 2024 EP elections shows structural, platform-general overrepresentation of Radical Right actors in news coverage itself—suggesting normalization operates upstream of social media curation, in traditional journalistic visibility.
Affect, rhetoric, and the discursive construction of polarization
A further cluster turns from exposure/structure to the affective and rhetorical textures of partisan media. Iyengar2019-jj anchors this with the foundational distinction between affective and ideological polarization, showing partisan animus functions as social identity independent of issue extremity—a distinction van-Eck2026-xg shows is underexploited in depolarization scholarship. Osmundsen2021-et extends this to misinformation sharing specifically, finding partisan (especially out-party) animus, not ignorance, drives fake-news sharing on Twitter, a finding resonant with Vosoughi2018-at’s landmark demonstration that falsehood diffuses farther and faster than truth, especially in the political domain, driven by human novelty-seeking rather than bots. Several papers examine the performative and weaponized use of polarization-adjacent vocabulary: Yoo2026-ev and Mosca2026-yh both show “fake news” accusations function as rhetorical weapons of delegitimization rather than descriptions of actual misinformation, while Elfes2026-jb introduces “narrative polarisation” to show that contestation over actor roles (perpetrator/victim) persists even where surface opinion appears to converge. The affective-affordance literature—Kalsnes2025-zb, Anwar2024-34dba628, Marino2026-he—shows how platform-specific emotional reactions (Facebook’s Love/Angry buttons) are differentially mobilized by populist actors and fluctuate unpredictably over time rather than reflecting stable echo-chamber dynamics, directly complicating Dehghan2026-sy’s Reddit finding of “sedimented” polarization as a structural precondition independent of actual discursive similarity.
Partisan asymmetry as the recurring empirical crux
Across methods and platforms, one question recurs insistently: is polarization symmetric? The evidence assembled here leans toward documenting right-leaning asymmetries in misinformation and visibility—Renault2025-uh (Republicans flagged 2.3x more often even under Community Notes’ cross-partisan bridging design), Bouchaud2026-np (showing this bridging design structurally undermoderates polarizing content), Yang2023-cg, Guess2021-ym, Gonzalez-Bailon2023-uy, and Ventura2026-yc’s striking finding that Brazil’s X ban produced a durable rightward “sorting ratchet.” Yet Van_Erkel2026-mk’s “hostile misinformation effect” and Ghezzi2023-8bebc91f’s study of UK scientific sourcing show that perceived and constructed asymmetries can be driven by identity-protective cognition and selective media sourcing rather than objective differences in content—and Much2026-gu’s podcast study introduces an orthogonal axis of sorting (gendered media spaces) that produces ideological exposure without ideological selection as such.
Open frontiers: LLMs, methodology, and the call for globalization
The newest layer of this literature turns reflexively on method and scope. Lee2026-je and Paci2025-ag probe what large language models can and cannot infer or interpret in partisan discourse—one showing surprising success at covert ideology inference, the other showing persistent failure at pragmatic/implicit meaning, bookending the promise and limits of LLM-based computational political communication exemplified methodologically by Arora2025-tx and Sarmiento2025-as’s unsupervised framing pipelines. Meanwhile Falkenberg2026-ka, Oswald2025-km, and Schulz2026-ts converge on a structural critique: polarization research is geographically skewed, built on unrepresentative vocal minorities, and subject to “false polarization” via self-selected posting—concerns given comparative texture by Rossi2023-847d5a9f and Zhu2026-tn’s European analyses and Lukito2026-il’s and Holland_Levin2026-qx’s cross-platform right-wing curation studies. Esau2025-tf closes the normative loop opened by van-Eck2026-xg, proposing deliberative reciprocity as an antidote—though the corpus as a whole, from Giglietto2012-32b6b551’s early skepticism about Facebook Likes as electoral signal through Giglietto2020-6278a4aa’s diagnosis of big data’s disruption of political communication scholarship, suggests the central lesson is methodological humility: what counts as “polarization” depends enormously on where, how, and at what level of the information funnel one chooses to look.