Coordinated Inauthentic Behavior

From CLSB to a field: methodological genealogy

The structure note’s methodological spine runs directly from Giglietto2020-9d8acdd7, which first operationalized “coordinated link sharing behavior” (CLSB) on Facebook during the 2018–2019 Italian elections, through the operational hardening of that approach in Giglietto2022-0e951ac5 (the MINE-FACTS fact-checking prototype) and Giglietto2023-fa71a001 (a continuously-updating workflow for the 2022 Italian election). These papers establish the signature move of this line: shift attention from content veracity or actor identity to collective behavioral signatures—near-simultaneous, repeated link-sharing across accounts—as a content-agnostic, evasion-resistant proxy for inauthentic coordination. This lineage culminates in the vera.ai-funded present: Giglietto2026-9b6a992d (gambling promotion and generative-AI visual persuasion) and Giada2026-fc9a3833 (the “Zombie Internet” synthesis of pro-Putin propaganda, gambling, and adult-content networks), both explicitly framed as outputs of the same CooRnet-descended detection pipeline feeding the VERA-AI Alert monitoring system. Rogers2026-cy, also a vera.ai product, extends the lineage laterally into “post-truth spaces,” combining network mapping with source-provenance investigation to serve fact-checkers directly—showing the applied, interventionist orientation that distinguishes this tradition from purely academic detection work.

Theoretical foundations: what “coordination” even means

Running parallel to the applied CLSB line is a more conceptual strand asking what coordination is and whether “coordinated inauthentic behavior” (Meta’s own term) is theoretically coherent. Starbird2019-qv reframes disinformation as participatory, collaborative “work,” insisting that the hard problem is distinguishing orchestration from organic crowd behavior—a provocation that haunts nearly everything downstream. Nizzoli2020-cf operationalizes coordination as a continuum rather than a binary, showing coordination and automation are orthogonal. Keller2019-nk, using rare ground-truth data from the Korean NIS astroturfing scandal, shows coordination traces (not bot-like features) are the most robust signal, while also demonstrating that even well-resourced campaigns can have strikingly limited discourse-level impact. This skepticism about impact resurfaces powerfully in Di-Marco2025-aa and its sequel Di_Marco2026-xu, which formally quantify (rather than merely detect) how much influence coordinated accounts actually exert in cascades—finding real CIB accounts perform close to random placement, far below optimal influence benchmarks. Thiele2025-ol pushes the conceptual critique furthest, arguing CIB conflates inauthenticity with intent and proposing an attribution-focused typology grounded in principal-agent theory. Mannocci2026-kc, a PRISMA survey of 122 papers, is the keystone synthesis here, proposing actors/actions/intent as the general definition and authenticity/harmfulness/orchestration/time-variance as orthogonal characterizing dimensions—effectively canonizing the field this structure note maps.

Detection methods: multiplex, temporal, statistical, and foundation-model advances

A dense cluster of 2025–2026 papers pushes detection methodology beyond single-modality, time-window heuristics. Mannocci2025-ig systematically compares five multimodal operationalizations on the 2019 UK election dataset, finding multiplex community detection the most balanced. Iannucci2025-eg independently converges on multiplex, time-decayed representations, validated against the 26-dataset benchmark later used by Iannucci2026-my, whose hypothesis-testing reframing—coordination detection as a planted-submatrix problem—delivers the field’s most rigorous detectability boundaries and a sobering lesson: dense recovered clusters are not evidence of coordination unless benchmarked against explicit null models. Yang2025-iv attacks the timing-manipulability weakness directly, proposing speed/frequency regularities as more evasion-resistant signals. Minici2024-tf (IOHunter) and Pante2025-pq extend detection into graph foundation models and cross-state control-dataset rigor, respectively, the latter explicitly reassessing—and complicating—prior claims of inter-state IO collaboration. Platform-specific adaptation is its own emerging front: Luceri2025-tr on TikTok’s video-first ecosystem, Rodriguez_Farres2025-sg on real-time Bluesky bot detection, and Kansaon2025-id extending Twitter-derived “Rapid Spread Network” logic into WhatsApp’s encrypted ecosystem.

Empirical case studies: platforms, polities, and actors

A broad empirical layer tests these methods against concrete campaigns. On Facebook, Song2025-yh uses CooRnet-style CLSB to contrast pro- and anti-vaccine coordination in the UK and US, and Oprea2025-lf documents hyperactive-user amplification in Romanian EP election pages. On Twitter/X, Simeone2025-vo shows deplatforming’s “network ripple effects” independent of CIB, while Appel2026-qr offers an unusually large-scale, platform-collaborative analysis of Meta-identified deceptive networks’ real exposure during the 2020 US election—crucially finding that resharing by unaffiliated accounts, not direct network posting, drives most reach, and that correlational “effects” on beliefs evaporate under controls. State-linked operations receive sustained attention: Poliakoff2026-fa reconstructs the IRA’s organizational and labor-market structure via OSINT; Bastos2025-ol probes visual self-presentation signatures across the Twitter Moderation Research Consortium dataset; Kuznetsova2025-nu and Kulichkina2026-zk examine Telegram/Twitter coordination in Russian-Belarusian and Chinese authoritarian contexts respectively; Jovanovic-Harrington2026-ze argues Serbia’s SNS exemplifies a subtler, harder-to-detect mimetic authoritarianism; and Kim2026-br applies explainable ML across two decades of Korean comment data to find condemning rhetoric, not direct praise, is the dominant and most-amplified troll strategy. Beyond elections, Gaw2025-ru reframes Philippine influence operations as political brokerage infrastructure, and Copland2025-em traces Sky News Australia’s propaganda role via network and topic analysis.

Science communication, platform governance, and generative AI as new frontier

Several papers push the topic’s boundaries outward. Efstratiou2026-ij shows coordinated amplification of contrarian science communicators intersecting with news-media pathways, while Zhao2025-ny and Arceneaux2026-xk examine bots as agenda-builders via hashtag hijacking and campaign messaging. Governance-adjacent work—Jurg2025-ur auditing YouTube’s moderation during EU elections—complements the detection literature by interrogating platform accountability itself. Most consequentially for the field’s near future, a cluster of 2025–2026 papers anticipates generative AI reshaping coordination: Orlando2025-ul shows LLM agents spontaneously reproduce IO coordination signatures even without central control; Schroeder2026-im extrapolates to “malicious AI swarms” as a democratic threat requiring new governance; Tang2026-gu documents AI-driven “matrix account” virality engineering in China; and FitzGerald2025-nv notes manipulation campaigns’ persistence through event-appropriation. Together with Giglietto2026-9b6a992d’s empirical documentation of ChatGPT-triggered volume explosions in gambling networks, these papers mark the line’s newest arc: from detecting human-coordinated inauthenticity to anticipating synthetic, agentic coordination at a scale and adaptivity existing detection frameworks were not built for—precisely the challenge the VERA-AI Alert system is now built to confront.