Mapping a fragmented field
Any synthesis of this literature must begin with Rothut2026-wt, whose scoping review of 382 publications spanning 25 years supplies the conceptual vocabulary the rest of the corpus implicitly assumes: radicalization as an incremental, multi-factor process rather than a single event, the internet as catalyst rather than cause, and the increasing untenability of a clean online/offline divide. Its call to map mechanisms across macro (discursive), meso (group), and micro (individual) levels, and along Lasswell’s communicator–content–medium–recipient chain, offers a useful lattice onto which the empirical studies filed here can be hung. Notably, the review flags a persistent evidentiary gap—longitudinal, causal work remains rare—that several of the papers below try to address, and its warning against treating “online radicalization” as a unitary phenomenon anticipates the field’s more recent turn toward differentiated, platform-specific, and ideology-specific accounts.
Algorithmic architectures and adversarial creativity
A cluster of papers examines how extremist actors read and exploit the technical grain of specific platforms. Karo2026-dn catalogues five concrete tactics—audio camouflage, meme infiltration, blurred visual intent, emoji codes, bait-and-switch—by which jihadist supporters embed propaganda in TikTok’s vernacular culture and defeat keyword- and image-based moderation, arguing for a shift from content removal to “infrastructural critique.” Kakavand2026-kt generalizes the affordance argument across five platforms simultaneously, showing that German far-right actors do not reproduce a single network but adapt distinctively to each platform’s connectivity, replicability, and scalability logics—Twitter as sprawling broadcast arena, Telegram as inward-facing fringe hub, Facebook as a disciplined AfD hierarchy. Rieder2026-pp extends this to deplatforming’s aftermath: even after Andrew Tate’s removal from YouTube, a diffuse “Tate-space” of interviews, remixes, and memeified content sustains his ambient influence because recommendation-driven circulation structurally outpaces moderation. Yet Brown2026-br’s carefully designed algorithmic audit complicates the emerging consensus that algorithms drive extremity: disentangling supply-side from demand-side effects, it finds strong “rabbit hole” dynamics (heavy dependence on the currently watched video) but no evidence of a distinct radicalization pathway, and only a mild, largely user-driven echo-chamber effect—alongside a platform-wide algorithmic tilt toward moderate conservatism for everyone. Read together, these papers suggest algorithmic complicity is real but uneven and format-dependent, cautioning against a single “the algorithm radicalizes” narrative even as they document concrete adversarial exploitation of platform design.
Softening, humor, and everyday normalization
A second thread foregrounds affect and vernacular style as normalization mechanisms rather than overt ideological argument. Nangle2026-yo frames far-right Instagram as an “affective-discursive ecology” in which humour, positive affect, and memetic Reels—rather than hostility—soften extremist worldviews and gain algorithmic proximity to mainstream content. Cabbuag2024-me documents an analogous, culturally specific dynamic in the Philippines, where TikTok “dogshows” normalize legal-but-harmful humor that blurs into cyberbullying and hate speech among Gen Z influencer audiences. Both resonate with Karo2026-dn’s notion of “everyday extremism,” in which propaganda collapses into platform vernaculars rather than remaining a discrete, spectacular speech act. Marwick2025-vx supplies the individual-level counterpart to this normalization story: qualitative analysis of “redpilling” narratives shows that most Far-Right adherents describe gradual socialization—accumulating “evidence,” books, and community affect—rather than sudden conversion, and that self-presentation as rational, scientific evaluators legitimizes hateful belief. This finding directly challenges simple “exposure equals belief” models and dovetails with Brown2026-br’s null finding on algorithmic radicalization pathways: across very different methods, both papers push back against reflexive assumptions that content exposure alone explains extremist uptake.
Mainstreaming through protest, influence networks, and transnational media
A third cluster examines mainstreaming as a meso-level, infrastructural process. Rothut2026-or coins “protest-facilitated mainstreaming” to describe how Germany’s Querdenken movement, while not itself extremist, became structurally embedded—via outlinking and network position—in far-right and conspiracist Telegram communities, with anti-elite framing (rather than explicit ideology) serving as the bridge. Askanius2026-de documents a parallel but more commercially inflected process on Swedish YouTube, where an “alternative influence network” fuses influencer microcelebrity, monetization, and political propaganda, using YouTube as a “sanitised” gateway that funnels audiences toward more extreme alt-tech platforms. Groebner2026-pc scales this argument transnationally, showing sustained commenter overlap across four English-language hyperpartisan YouTube channels (Fox News, Sky News Australia, Rebel News, GB News), with shared themes like transphobia and US politics acting as connective tissue for a genuinely transnational far-right audience. De_Leon2025-qn offers a conceptual bridge for this whole cluster, proposing “networked exposure” to capture how even ordinary content (news stories, Wikipedia articles) becomes politically weaponized simply through its circulation within extremist-aligned networks—a notion that fits neatly with Rothut’s content/connective mechanism distinction and Askanius’s account of hyperlinking as infrastructure.
From online discourse to offline mobilization
Where the mainstreaming literature traces discursive diffusion, Bailard2024-pj tackles the harder causal question of whether online talk predicts offline violence. Using Granger causality on Proud Boys Telegram data merged with ACLED event records, it finds diagnostic (grievance) and motivational (solidarity) framing—but not explicit calls-to-action—predict subsequent violent events, and identifies a reciprocal four-week online–offline mobilization cycle. This is one of the few papers in the set offering genuinely longitudinal, causally-oriented evidence of the kind Rothut2026-wt identifies as scarce in the field, and its explicit argument against a content-moderation lens focused on discrete posts echoes Karo2026-dn and Rieder2026-pp’s calls to attend to ecosystems and circulation dynamics rather than isolated pieces of content.
Coordinated inauthentic behavior, digital authoritarianism, and computational propaganda
A final cluster shifts from grassroots extremism to state-linked and AI-assisted manipulation, broadening “extremism” toward adjacent phenomena of manufactured consent. Kulichkina2026-zk and Jovanovic-Harrington2026-ze both examine authoritarian-context coordination—Twitter activity around China’s 2022 COVID protests and Serbia’s SNS-linked networks respectively—with Jovanovic-Harrington’s key contribution being that sophisticated CIB now mimics authentic grassroots support well enough to evade standard bot-detection frameworks, a subtler “digital authoritarianism” than centralized bot farms. Kim2026-br brings large-scale explainable ML to two decades of Korean troll comments, showing that morally condemning rhetoric—rather than direct pro-foreign praise—both dominates and is disproportionately amplified by engagement-based ranking, a finding with implications for how extremist and state-propaganda content alike exploits platform amplification logics documented elsewhere in this set. Beacken2026-zb and Perez-Curiel2026-ld extend the picture into generative AI: the former shows GenAI propaganda adoption varies by structural context across six democratically weakened states rather than following a uniform technological trajectory, while the latter examines how far-right actors deployed AI-generated multimedia in 2024 European Parliament campaigns on Instagram and X, suggesting the affordance-exploitation dynamics central to this whole topic (per Kakavand2026-kt and Nangle2026-yo) are already being extended to synthetic media.
Moderation and its limits
Threading through nearly every cluster is a shared skepticism about content moderation’s efficacy, though the papers disagree on remedies. Simeone2025-vo offers the most optimistic case: targeted deplatforming of a handful of central Twitter accounts collapsed the Arizona Election Review network’s hub/authority structure and ended coordinated mobilization, functioning as an effective “network intervention” distinct from the visibility-reduction approaches now favored post-acquisition. Against this, Rieder2026-pp and Karo2026-dn show moderation structurally outpaced—by recommendation-driven circulation on YouTube and by adversarial vernacular camouflage on TikTok, respectively—while Askanius2026-de documents actors reframing moderation itself as censorship to fuel persecution narratives. Taken together, this final thread suggests that moderation’s effectiveness may hinge less on removing content than on disrupting network structure and infrastructural affordances—precisely the shift in analytical focus that Bailard2024-pj, Rothut2026-wt, and Karo2026-dn all separately call for.