fg-zettelkasten

How deceptive online networks reached millions in the US 2020 elections

4 min read

How deceptive online networks reached millions in the US 2020 elections

Summary

This paper provides the first large-scale exposure-based measurement of “deceptive online networks” — coordinated operations using identity deception, whether politically or financially motivated — that targeted US users on Facebook and Instagram around the 2020 elections. Combining platform-level data on 49 networks identified and removed by Meta (13 Coordinated Inauthentic Behavior networks and 36 Financially Motivated Operations) with individual-level survey and behavioral data from ~73,000 consenting users in the US 2020 Facebook and Instagram Election Study (FIES), the authors show that reach was both substantial (37M Facebook users) and extremely concentrated (3 networks ≈ 80% of reach), and that most exposure flowed through reshares by ordinary unaffiliated users rather than direct posts from network accounts. After balancing on pre-exposure user characteristics, naive correlations between exposure and outcomes like factual discernment or election-legitimacy beliefs largely vanish, cautioning against causal interpretations of exposure-effect estimates.

Key Contributions

  • Proposes the broader category of deceptive online networks, encompassing both politically motivated CIB and financially motivated operations (FMOs) that produce political content.
  • First platform-wide measurement of actual exposure (not engagement proxies) to coordinated deceptive networks during a major US election.
  • Documents that non-network reshares by ordinary users drive most reach, shifting the analytical frame from producers to amplifiers.
  • Demonstrates that FMOs can rival or exceed CIB networks in political reach, arguing they merit equivalent scholarly and platform scrutiny.
  • Releases a de-identified dataset via the Social Media Archive (SOMAR/ICPSR).
  • Methodological cautionary tale: shows how user-characteristic confounding produces spurious exposure–outcome associations in observational designs.

Methods

Pre-registered observational study covering 26 June 2020 – 15 February 2021. Platform-level analyses use Meta-supplied aggregated data on 49 deceptive networks, decomposing direct vs. indirect (reshare-driven) exposure and computing cascade size, depth, breadth, and structural virality. Content is topic-classified using Meta’s Topic and Civic classifiers. Individual-level analyses use five waves of FIES survey data (~73,000 consenting users) linked to behavioral logs, with entropy balancing to reweight non-exposed users across nine progressively richer covariate specifications, plus pre-registered sensitivity analyses for unobserved confounding. Post-hoc exploratory analyses examine the effect of FIES experimental feed interventions (no-reshare, chronological, reduced like-minded content) on exposure.

Findings

  • Networks reached ~37M Facebook users (14.85%) and ~3M Instagram users (1.86%), generating 175M and 70M views respectively.
  • Extreme concentration: 3 networks (Rally Forge/CIB9, FMO27 Kosovo, FMO34) ≈ 80% of unique FB viewers; CIB10 ≈ 95% of IG viewers.
  • Geographic origins: CIB networks chiefly from Russia, US, and Iran; FMO networks concentrated in the Balkans (16) and South Asia (11).
  • Politics/social issues made up 65% of CIB and 32% of FMO direct content on Facebook, despite the latter’s ostensibly financial motives.
  • Heavy-tailed exposure: 1% of users absorbed 55% of network content views on FB and 96% on IG.
  • For Rally Forge, 13M of 13.4M viewers were reached indirectly via non-network reshares; only 1.3M directly.
  • Just 5.67% (FB) and 0.34% (IG) of exposed users reshared network content, yet these resharers drove most of the amplification.
  • Network content averaged only 0.3% of pre-election political content views for exposed users.
  • Exposed users skewed older, more conservative, heavier platform users, and consumed more untrustworthy sources — regardless of network type.
  • After entropy balancing, exposure–outcome associations (discernment, election legitimacy, partisan clicks) largely disappear; sensitivity analyses show vulnerability to unobserved confounding.
  • No-reshare and chronological feed interventions reduced exposure but also overall engagement; effects underpowered.

Connections

This paper is a flagship output of the FIES collaboration and connects directly to other FIES-adjacent work on platform-scale exposure and curation effects such as Gonzalez-Bailon2024-rq and Budak2024-ef. Its emphasis on indirect amplification by ordinary users echoes findings on small-share-of-diet untrustworthy content and superspreader dynamics in Mosleh2024-op and DeVerna2025-dl, while its conceptual move to bundle financially motivated operations with CIB resonates with detection-and-typology work like Minici2024-tf, Luceri2025-tr, and Kulichkina2026-zk. The methodological caution about confounded exposure–effect estimates speaks to debates raised in Marwick2025-ov and Starbird2025-jj about how the field conceptualizes “harm” from influence operations.

Podcast

A research-radio episode discusses this paper: Listen

Graph View

Backlinks