Framing the Field: From Causal Models to Processual Ecologies

The broadest orienting text in this cluster is Rothut2026-wt, a scoping review of twenty-five years of online-radicalization scholarship that explicitly argues against treating “online” and “offline” as separable domains and against single-cause models of radicalization. It proposes a multi-level (macro/meso/micro) and Lasswellian (communicator/content/medium/recipient) scaffolding that implicitly organizes much of what the other papers in this topic do empirically. Read against this review, Marwick2025-vx supplies the micro-level corrective the review calls for: its qualitative analysis of “redpilling” narratives shows that most Far-Right adherents describe belief adoption as a gradual socialization process—driven by cumulative “evidence,” community participation, and a self-image of rational skepticism—rather than the sudden-exposure conversion implied by naive “exposure equals belief” models. Bailard2024-pj extends this anti-reductionist stance to collective mobilization: using Granger causality on Proud Boys Telegram data, it shows that diagnostic (grievance) and motivational (solidarity) framing, not explicit calls to action, predict offline violence, and that online and offline activity form a reciprocal four-week cycle. Together these three papers establish a throughline for the topic: radicalization and mobilization are emergent, temporally extended, multi-actor processes embedded in platform ecosystems, not discrete acts of exposure or incitement that content moderation alone can intercept.

Mainstreaming Mechanisms: Protests, Affect, and Networked Legitimation

A second cluster investigates how extremist ideas migrate from fringe to mainstream without requiring individual-level conversion, in effect operationalizing “mainstreaming” as a meso-level, discursive phenomenon. Rothut2026-or coins “protest-facilitated mainstreaming” to describe how the German Querdenken movement’s anti-elite, conspiratorial Telegram discourse structurally embedded ordinary pandemic-skeptic sympathizers within international far-right and QAnon-adjacent networks—crucially, through shared anti-establishment affect rather than explicit ideological commitment. Nangle2026-yo identifies an analogous affective logic on Instagram, where humor, positive affect, and memetic “softening” render far-right content palatable within an “affective-discursive ecology” shaped by algorithmic proximity. Askanius2026-de shows the same hybridization occurring through influencer economics: Swedish far-right YouTubers fuse microcelebrity branding, monetization, and political propaganda, using pandemic-era alliances with wellness and anti-vaccine communities to professionalize and diversify their reach, with YouTube functioning as a “sanitized” gateway to more extreme alt-tech platforms. Groebner2026-pc scales this mainstreaming logic transnationally, finding substantial commenter overlap among English-language hyperpartisan news channels (Fox News, Sky News Australia, Rebel News, GB News) and identifying a US-centered, Anglophone far-right information space bound together by shared enemies (notably trans people) rather than national particulars. Klinger2023-qu provides perhaps the sharpest mechanism for this whole cluster: its network analysis of Germany’s “Migrationspakt Stoppen” campaign shows how an extremist Identitarian Movement initiative was “source-laundered” through superspreader accounts and far-right media before being adopted, almost verbatim, by the AfD party leadership—demonstrating that mainstreaming can proceed through networked rather than individual or even directly coordinated transmission, evading usual legal and moderation categories built around attributable coordination.

Platform Affordances and Algorithmic Architectures

A third strand interrogates the technical substrate that makes such mainstreaming and evasion possible. Kakavand2026-kt offers a comparative anchor: analyzing German far-right actors across five platforms, it shows that identical seed accounts produce starkly different network topologies—Twitter/X as sprawling broadcast scalability, Telegram as inward-facing fringe cohesion, Instagram as fragmented but tightly reciprocal echo chambers, Facebook as a hierarchical AfD-run operation—demonstrating that “affordance” is not a fixed platform property but a relational outcome of design and practice. Karo2026-dn narrows into a single platform’s vulnerabilities, cataloguing five tactics (audio camouflage, meme infiltration, blurred intent, emoji codes, bait-and-switch) by which jihadist supporters embed propaganda in TikTok’s vernacular culture, arguing this constitutes “everyday extremism” that collapses spectacle into ritual and exposes TikTok’s moderation as structurally, not just operationally, flawed. Rieder2026-pp documents a parallel structural outpacing on YouTube: even after Andrew Tate’s official deplatforming, a diffuse “Tate-space” of interviews, remixes, and algorithmically amplified Shorts sustains and even increases his visibility, because recommendation-driven circulation moves faster than removal. Against these accounts of algorithmic complicity, Brown2026-br offers an important counterweight using an experimental audit design: it finds no evidence that YouTube’s recommender independently drives users down ideologically extremizing “radicalization pathways,” locating observed partisan differences instead in user choice, while still identifying a platform-wide algorithmic tilt toward moderately conservative content and strong content “rabbit holes.” Read together, these four papers refine rather than resolve the “does the algorithm radicalize” debate, suggesting the more tractable empirical object is not algorithmic causation of belief change but algorithmically amplified circulation and differential platform affordance.

Moderation, Deplatforming, and Their Limits

A closely related set of papers asks what happens when platforms intervene. Simeone2025-vo offers a rare in-situ, networked account of deplatforming’s efficacy: Twitter’s 2021 suspension of eight “stop the steal” accounts collapsed the Arizona Election Review community’s hub/authority structure and shifted its discourse from mobilization to abstract grievance about censorship—evidence that targeted removal of a few central nodes can produce disproportionate network-wide effects, a more optimistic finding than the ambient persistence documented for Tate on YouTube in Rieder2026-pp or the structural evasion tactics shown for TikTok in Karo2026-dn. The tension between these cases—deplatforming “works” for a geographically and temporally bounded mobilization network but fails against a diffuse, memeified ideological presence—suggests moderation efficacy may depend on whether extremist communication is organized around identifiable hubs or distributed across vernacular, algorithmically renewable content.

Organized and Synthetic Propaganda: State Actors and Generative AI

A final pair of threads extends the topic beyond grassroots and influencer-driven extremism into organized and increasingly automated propaganda production. Poliakoff2026-fa reframes Russia’s Internet Research Agency not as a covert intelligence unit but as an “outsourced troll” operation integrated into Russia’s ordinary media labor market, recruited heavily around the 2014 Crimea annexation rather than solely the 2016 US election—a corrective to content-centric disinformation scholarship that neglects organizational structure. Bastos2025-ol complements this with a computational-vision approach to the same genre of operation, showing that foreign influence campaigns leave detectable visual “signatures” (gender, race, camera angle, sensuality) in fake profile imagery that could support attribution by country of origin. Looking forward, Perez-Curiel2026-ld and Crichton2026-kn register the emergent generative-AI frontier of this topic: the former documents far-right parties’ adoption of AI-generated multimedia in the 2024 European election campaigns, while the latter’s media experiment argues that commercial generative models do not neutralize extremist imagery but sanitize its explicit markers while structurally preserving or displacing its ideological content—reframing the danger of AI-generated extremist media away from deception/deepfakes and toward ideological transmission embedded in the models’ own training-data biases.

Synthesis

Taken as a whole, the arc of this topic moves from conceptual correctives to simplistic exposure/causation models (Rothut2026-wt, Marwick2025-vx, Bailard2024-pj), through empirical demonstrations of how mainstreaming operates as a networked, affective, and economically incentivized process across platforms (Rothut2026-or, Nangle2026-yo, Askanius2026-de, Groebner2026-pc, Klinger2023-qu), into granular accounts of the platform-specific affordances and algorithmic dynamics that enable both propaganda circulation and its partial containment (Kakavand2026-kt, Karo2026-dn, Rieder2026-pp, Brown2026-br, Simeone2025-vo), and finally outward to the organized and increasingly synthetic infrastructures—state-linked troll farms and generative AI—that now co-produce extremist content at scale (Poliakoff2026-fa, Bastos2025-ol, Perez-Curiel2026-ld, Crichton2026-kn). The collective lesson is that extremist mobilization online is neither a simple function of exposure nor purely an algorithmic artifact, but an emergent property of affordances, labor, affect, and increasingly automated content production interacting across a fragmented, multi-platform ecosystem—one that existing content-moderation paradigms, built around discrete speech acts, are structurally ill-equipped to govern.