Mapping the terrain: why polarization research disagrees with itself
A cluster of papers in this collection functions as meta-commentary on the field itself, and it is worth starting there because it frames everything else. van-Eck2026-xg finds that depolarization recommendations in the scholarly literature are mostly disconnected from the evidence researchers themselves produce, resting instead on unexamined commitments to deliberative democracy. Hartmann2025-px makes a parallel diagnosis for echo chamber research specifically: much of the field’s apparent disagreement about whether echo chambers exist is an artifact of inconsistent conceptualization and measurement rather than a real empirical dispute. Schemer2026-mh demonstrates this concretely — running 504 specification combinations on the same panel data, they show that how you measure “partisan media slant” and “polarization” swings coefficient magnitudes sevenfold, though never flips their sign. Falkenberg2026-ka adds a geographic version of the same critique: the evidence base is disproportionately American and disciplinarily siloed. Together these pieces counsel caution before treating any single finding below — including the headline null results from Meta — as settled.
The null-results consensus and its limits
The best-resourced empirical interventions here converge on a strikingly consistent finding: algorithmic curation shapes what people see, but not obviously what they believe. Bakshy2015-rn first showed that individual clicking, not Facebook’s algorithm, is the main brake on cross-cutting exposure. The 2020 Meta-academic collaboration triad — Guess2023-ur, Nyhan2023-gb, Bergeron-Boutin2026-yh — extends this at massive scale and under experimental control: switching users to chronological feeds, cutting like-minded exposure by a third, or reducing untrustworthy-source exposure all reshape feed composition dramatically but leave affective polarization, belief change, and political behavior essentially untouched. Gonzalez-Bailon2023-uy complicates the picture from inside the same dataset family, showing segregation is real, asymmetric, and concentrated in a homogeneously conservative corner of the ecosystem saturated with misinformation — segregation in exposure, in other words, coexists with null attitudinal effects downstream.
This consensus is now being directly challenged. Gauthier2026-iq is the pointed counter-case: an X field experiment finding that algorithmic feeds do shift attitudes rightward, and — crucially — offers a mechanism explaining why Meta’s studies might have found nothing: the algorithm’s effect operates by reshaping who users follow, a change that persists even after the algorithm is switched off. Because Meta’s studies measured within-experiment attitude change without this following-based channel fully in view, the asymmetry Gauthier documents (on-effects persist, off-effects don’t reverse them) offers a plausible reconciliation of seemingly contradictory findings rather than a refutation. Brady2026-ln’s Bluesky field experiment pushes the null-effects story further still, finding that engagement-based algorithms amplify moralized/toxic content and distort perceived social norms even while failing to move users’ own behavior in an 8-week window — suggesting the causal action, if any, runs through perception of the discourse climate rather than direct persuasion. Efstratiou2025-gs meanwhile shows that apparent “right-leaning” algorithmic bias on Twitter/X dissolves once you control for agitating content and proximity to Elon Musk — a reminder that algorithmic effects are often better explained by behavioral and ownership variables than by partisanship per se.
From exposure to engagement: the Facebook/Instagram visibility pipeline
A second strand shifts attention from who sees what to how seeing becomes visible, viral, and governed. Giglietto2026-632ef967 is the anchor piece here: analyzing 130,448 viral Facebook URLs, it shows that highly partisan content systematically underperforms its share count in views while journalistically rigorous content overperforms — and that these penalties/rewards fluctuate in lockstep with known Meta governance interventions like the 2020 “break the glass” measures. This directly challenges the framing (implicit in Nyhan2023-gb and Guess2023-ur) of curation as a roughly stable background condition; Facebook is an active, time-varying curator, not a neutral pipe. McNally2025-dn corroborates this with a decade of Guardian engagement data, finding detectable ~19-24 day lagged effects of ranking changes specifically on hard news, puncturing the “black box” framing and making a case for systemic DSA-style auditing. Gillespie2022-jx supplies the conceptual vocabulary for this whole phenomenon: “reduction” (algorithmic demotion short of removal) as an underscrutinized, largely invisible form of content moderation that concentrates platform power precisely by avoiding the visibility of removal decisions.
Downstream of visibility is affect. Green2025-ap shows that source-level partisanship scores systematically mask story-level heterogeneity — the real unit of partisan curation is the individual story unbundled and re-bundled by networked sharers, not the domain. This matters for reading papers like Eady2025-vm, which maps ideology via URL-sharing and finds that ordinary politically engaged users, not politicians, generate most of the extreme content in elite information ecosystems. On the affective-engagement side, Kalsnes2025-zb, Larsson2026-nl, and Anwar2024-34dba628 (a systematic review of Facebook Reactions research) converge on a simple but consequential pattern: anger/negative affect mobilizes sharing more reliably than love/positive affect, and populist and challenger parties are disproportionately good at eliciting it — though Larsson2026-nl’s longitudinal Norwegian data suggests this negativity advantage may be eroding over time, perhaps through audience fatigue. Marino2026-he’s three-year study of pro-Bolsonaro Facebook networks is the crucial corrective to any static reading of these dynamics: affective engagement within even a hyperpartisan community is highly unstable, event-driven, and penetrated by cross-partisan “brigading,” undermining the assumption that hyperpartisan communities are simply fixed echo chambers.
Insularity and the architecture of partisan communities
A distinct line examines partisan community structure directly, often via network and discourse methods rather than attitude surveys. Giglietto2019-882f1900, on the 2018 Italian election, is a key precedent for Marino2026-he: it shows that insularity (via Twitter-measured source concentration) correlates with amplification-via-sharing on Facebook, while cross-partisan exposure correlates with contestation-via-commenting — and that M5S’s populist “cyber party” structure let it exploit both mechanisms strategically. Dehghan2026-sy extends the insularity argument to Reddit across 16 years, finding negligible cross-ideological overlap even between discursively similar subreddits and arguing that polarization is “sedimented” as a precondition of platform architecture rather than an emergent outcome of discourse. Barbera2015-fw and Barbera2015-je supply the methodological backbone much of this literature depends on — ideal-point estimation from network-following structure — while showing that echo-chamber intensity is topic- and time-contingent rather than uniform, with liberals engaging in more cross-ideological retweeting than conservatives, an asymmetry Bail2018-fk’s famous backfire experiment inverts at the individual level (conservative exposure subjects moved right; liberal subjects did not move significantly).
Grusauskaite2026’s study of Great Reset discourse supplies a non-network account of the same phenomenon: insular communities can be constituted relationally and affectively — through boundary work, stigma reclamation, and shared “deep stories” — rather than primarily by platform structure or ideology content, a claim that resonates with Prochaska2025-ef’s account of deep-story collaboration around 2020/2022 election denial and complicates purely architectural readings like Dehghan’s.
Asymmetry, radicalization, and the right-leaning media ecosystem
Running through many of these papers is a recurring empirical asymmetry: conservative/right-wing actors and audiences appear disproportionately insular, disproportionately rewarded by algorithmic attention, and disproportionately linked to low-quality or false content. Benkler2018-lw is the theoretical keystone, arguing the “epistemic crisis” is institutionally and ideologically asymmetric — a radicalized right-wing media ecosystem insulated from corrective journalistic norms — rather than a symmetric technological pathology. Gonzalez-Bailon2023-uy’s finding that misinformation clusters almost entirely in the conservative corner of Facebook, Guess2021-ym’s finding that older/conservative users disproportionately share low-credibility news, and Gauthier2026-iq’s finding that X’s algorithm promotes conservative activist content all triangulate on the same asymmetry. Knupfer2025-vt offers a mechanism at the elite level — a “logic of connective faction” whereby ideologically extreme Republicans build ties to hyperpartisan media and are rewarded with engagement, illustrated through the CRT case. Ventura2026-yc’s Brazil ban study shows the asymmetry can also be behavioral/compliance-based: conservative users circumvented an X ban while liberals complied, durably shifting the platform’s partisan composition rightward. Much2026-gu’s “gendered media spaces” argument on podcasts and the 2024 youth gender gap extends this asymmetry into incidental, non-political exposure — masculine-coded entertainment content correlates with conservative political content, giving Trump a persuasion channel Harris’s more ideologically-sorted strategy lacked. Against this grain, Osmundsen2021-et locates the asymmetry not in psychology but in content supply: Republicans share more fake news not because of greater ignorance or cognitive deficit but because out-party hatred is the dominant driver of sharing for both sides, and only pro-Republican fake news consistently vilifies the opposing party.
Platform transformation and the future of “social”
A final set of papers steps back to ask whether the categories “social media,” “partisan engagement,” and “news” still mean what polarization research assumes they mean. Boyd2026-op and Gerbaudo2026-fo both argue that the platforms under study have morphed away from the “networked publics” model on which most of this literature (including Bakshy, Barberá, Bail) is built: toward “parasocial” consumption of influencer content and toward algorithmically “clustered publics” organized by behavioral inference rather than explicit ties. This matters directly for the topic: if TikTok-style clustering and parasocial engagement are displacing the friend-network/News-Feed model that Bakshy2015-rn through Gonzalez-Bailon2023-uy all analyze, then findings about Facebook/Instagram insularity may describe a historically bounded regime rather than a timeless property of social media polarization. Divon2026-wx’s work on influencers as “brokers of the state” on TikTok/Douyin, Nangle2026-yo’s analysis of far-right affective softening via Instagram Reels, and Cabbuag2024-me’s account of TikTok “dogshows” all document this newer, influencer- and affect-centered mode of partisan content circulation that sits awkwardly within the domain-sharing and feed-exposure paradigms dominant elsewhere in this collection — suggesting that the next generation of partisan-engagement research will need new units of analysis (narrative motifs, as in Elfes2026-jb; informativeness perceptions, as in Wedel2026-ix; LLM-inferred alignment, as in Lee2026-je) to keep pace with platforms that no longer organize engagement primarily through visible, reciprocal social ties.