Appel, R. E., Kim, Y. M., Pan, J., Xu, Y., Nimmo, B., Thomas, D. R., Allcott, H., Barberá, P., Brown, T., Crespo-Tenorio, A., Dimmery, D., Freelon, D., Gentzkow, M., González-Bailón, S., Guess, A. M., Iyengar, S., Lazer, D., Malhotra, N., Moehler, D., Nyhan, B., Settle, J., Thorson, E., Tromble, R., Rivera, C. V., Wilkins, A., Wojcieszak, M., Xiong, B., de Jonge, C. K., Franco, A., Mason, W., Stroud, N. J., & Tucker, J. A. (2026). How deceptive online networks reached millions in the US 2020 elections. Nature Human Behaviour, 1–15. https://doi.org/10.1038/s41562-026-02435-2
Summary
This study offers the first large-scale, platform-wide measurement of actual user exposure to deceptive online networks during the 2020 US elections, drawing on 49 networks that Meta identified and removed from Facebook and Instagram. By combining Meta’s platform-level exposure data with individual-level survey and behavioral data from the US 2020 Facebook and Instagram Election Study (FIES), the authors reframe the object of study: rather than restricting analysis to politically motivated foreign influence operations, they introduce the broader category of “deceptive online networks” — coordinated efforts using identity deception, whether their motives are political or financial. Their central findings are that reach was highly concentrated in a handful of networks, that most exposure occurred indirectly through reshares by ordinary unaffiliated users, and that the people exposed were systematically distinctive (older, more conservative, heavier platform users). Crucially, once these pre-exposure characteristics are controlled for, apparent associations between exposure and outcomes like factual discernment or election legitimacy beliefs largely vanish, cautioning against causal readings of observational exposure data.
Key Contributions
- Introduces the concept of “deceptive online networks” to encompass both politically motivated influence operations and financially motivated operations (FMOs) that produce political content.
- Provides the first platform-wide measurement of actual exposure (rather than engagement proxies) to deceptive networks across an entire major platform user base during a US election.
- Documents the central role of unaffiliated non-network reshares in amplifying deceptive content, shifting analytic focus from network actors to user amplification dynamics.
- Shows that financially motivated networks can out-reach politically motivated ones, arguing FMOs deserve equal scholarly and platform scrutiny.
- Offers a methodological caution about causal inference from observational exposure data, showing how user-characteristic confounding produces spurious associations.
- Releases a de-identified dataset via the Social Media Archive (SOMAR/ICPSR).
Methods
The pre-registered observational study covers 49 deceptive networks (13 CIB and 36 FMO) targeting US users from June 2020 to February 2021. Platform-level aggregated data from Meta enabled analysis of direct versus indirect exposure, including cascade size, depth, breadth, and structural virality; content was topic-classified using Meta’s Topic and Civic classifiers. This was linked to individual-level survey and behavioral data from roughly 73,000 consenting Facebook and Instagram users across five FIES waves. Entropy balancing reweighted non-exposed participants to achieve pre-exposure covariate balance across nine sequential model specifications, supplemented by pre-registered sensitivity analyses for unobserved confounding and a post hoc analysis of FIES feed interventions (no-reshare, chronological, reduced like-minded content).
Findings
- Networks reached ~37 million unique Facebook users (14.85%) and ~3 million Instagram users (1.86%), generating 175 million Facebook and 70 million Instagram views.
- Three networks (Rally Forge, FMO27 from Kosovo, and FMO34) accounted for nearly 80% of unique Facebook viewers; a single network drove ~95% of Instagram viewers.
- Politics and social issues made up 65% of CIB and 32% of FMO direct Facebook content, despite FMOs being financially motivated.
- For Rally Forge (the highest-reach network), 13 million of 13.4 million viewers were reached indirectly via reshares, versus only 1.3 million directly.
- Only 5.67% of exposed Facebook users (and 0.34% on Instagram) reshared network content, yet these users disproportionately amplified reach; 1% of users accounted for 55% of Facebook exposures.
- Network content averaged just 0.3% of exposed participants’ political content views in the 41 days before the election.
- After entropy balancing, associations between exposure and outcomes (factual discernment, election legitimacy, partisan news clicks) largely disappeared and proved vulnerable to unobserved confounding.
- Feed interventions reduced exposure but also reduced overall engagement; effects were underpowered.
Connections
This paper is a flagship output of the FIES Meta–academic collaboration and sits alongside other work from that partnership on platform effects and information exposure, notably Allcott2025-jb and Gonzalez-Bailon2024-rq. Its reconceptualization of coordinated identity deception as “deceptive online networks” and its emphasis on amplification by ordinary users connect it to the broader coordinated-inauthentic-behavior literature, including Luceri2025-tr and Starbird2025-jj. Its methodological caution about spurious exposure–outcome associations resonates with debates about measuring the real reach and impact of influence operations rather than relying on engagement proxies.
Podcast
A research-radio episode discusses this paper: 🎧 MP3 · Spotify · Apple Podcasts