Attributing coordinated social media manipulation: A theoretical model and typology

Summary

This conceptual paper tackles the “attribution problem” in research on coordinated social media manipulation (CSMM): how to infer the covert principals behind campaigns when only downstream traces are visible. The authors argue that platform-derived concepts like Meta’s “coordinated inauthentic behavior” conflate inauthenticity with manipulative intent and are poorly suited to comparative scholarly inquiry. They propose a consolidated definition of CSMM centered on manipulative intent, synthesize a literature-based catalogue of observable campaign features, and adapt Esser’s rational choice framework to model principals’ commissioning decisions. The output is a 2×2×2 typology of eight campaign types that maps observable structural features (scale, elaborateness, disguise) onto hidden principal attributes (resources, stakes, influence strategy).

Key Contributions

• A consolidated definition of CSMM organized around three dimensions—coordination, social media, manipulation (influence + deception)—that decouples manipulative intent from inauthenticity.
• A review-derived catalogue of observable (scale, elaborateness, disguise) and hidden (principal type, agent organization, audience, strategy) CSMM characteristics.
• A rational choice expected-utility model formalizing the principal’s decision: EU_i = p_Wi·U_W − (1−p_Wi)·C_F − p_Di·p_Pi·C_P − C_i.
• An eight-cell typology with worked empirical exemplars and bridging hypotheses linking campaign structure to principal attributes.
• An interdisciplinary research agenda emphasizing triangulation, expanded platform data access (e.g., via the EU DSA), and the integration of computational and social-scientific methods.

Methods

The paper is theory-building rather than empirical. The authors conduct a systematic literature review across Web of Science and EBSCOhost (515 articles screened by five coders, 62 empirical studies retained) to inductively extract recurring campaign features. They then adapt Esser’s (1999) rational choice framework to derive a three-stage decision model (principal → agents → audience) and dichotomize three observable dimensions to construct a 2×2×2 typology. The typology is illustrated through documented cases including the IRA, the Chinese “50c party,” the South Korean NIS, Saudi Twitter operations, GRU activity in Syria, and the Adani coal mine lobbying campaign.

Findings

• Three observable characteristics recur across the empirical literature: scale (accounts, posts, platforms, duration), elaborateness (content sophistication), and disguise (camouflage techniques).
• Disguise sophistication is a particularly diagnostic signal of what is at stake for the principal, while scale and elaborateness index available resources and strategy.
• Existing strategy typologies (boosting, pollution, bullying, polarization, distraction) are non-mutually-exclusive and conceptually inconsistent.
• Seven of the eight typology cells have recognizable empirical exemplars; “low-stakes targeting” has no clear documented case, suggesting either a real rarity or a systematic detection gap.
• Attribution-relevant ground-truth data is heavily Twitter-skewed; Meta and Reddit disclosures lag far behind, constraining comparative research.
• Structural features should be prioritized over content for attribution, since content is often deliberately misleading.

Connections

This paper provides a conceptual scaffold for much of the empirical CIB/CSMM literature in this register, including work on troll-farm operations and strategic actor identification such as Kulichkina2026-zk, DeVerna2025-dl, Luceri2025-tr, Kuznetsova2025-nu, and Bollenbacher2026-vz. Its critique of platform-derived definitions and call for analytic clarity speaks directly to debates on the conceptual status of disinformation and CIB raised in Farkas2026-lr, Starbird2025-jj, and Marwick2025-ov. The principal–agent and rational choice framing also connects to research on astroturfing detection and account-coordination methods like Minici2024-tf and Graham2025-gp.

Podcast

A research-radio episode discusses this paper: Listen