For nearly a decade, Washington has fought foreign influence operations by chasing content. Analysts track narratives, count reach, measure belief change, and flag falsehoods. A preprint posted to arXiv on August 31, 2026 — not yet peer-reviewed — suggests that framework has been aimed at the wrong target.
Researchers from the London School of Economics, Oxford, Corvinus University, and Central European University analyzed roughly 2.5 million tweets around the 2020 Black Lives Matter protest peak and found that heavier exposure to bots predicted a sharper subsequent collapse in human-to-human network cohesion — with the damage concentrated among the movement’s own core supporters.
The mechanism is not persuasion. It is the quiet dissolution of the human ties through which any movement, of any ideology, actually functions. If the finding holds, every U.S. counter-disinformation framework in operation today is measuring the wrong thing.
What The Study Actually Found
Linda Li, Orsolya Vásárhelyi, and Balázs Vedres built longitudinal communication networks at three points in 2020: a May baseline, the June 7–8 protest peak following George Floyd’s death, and a September post-peak snapshot. The peak core contained 92,078 users — 27,221 human, 50,982 bot — and 372,932 edges. They measured cohesion two ways: whether a user’s human contacts remained connected to each other (triadic closure), and whether human-to-human links held up inside detected communities (edge density).
Both eroded after bot exposure. The community-level effect was strong and statistically significant (β = −0.7052, p < 0.001), and the erosion was asymmetric — it hit pro-movement supporters hardest. As the authors put it, “Bots may weaken activism less by changing what people believe than by dissolving the ties through which collective action is sustained.
Why BLM Is A Useful Test Case, Not The Point
BLM is a useful case because it is already the most documented target of foreign influence operations in recent U.S. history. The bipartisan Senate Intelligence Committee concluded in 2019 that “no single group of Americans was targeted by IRA information operatives more than African Americans,” and Kremlin operators ran both pro-BLM and Blue Lives Matter accounts to inflame both sides.
The mechanism the new paper documents is ideology-agnostic. The same technique would work against a church coalition organizing on persecution, a campus pro-Israel group, a Tea Party network, or a Christian conservative movement. Automated accounts do not need to know what a community believes to fragment it.
The authors are explicit about the study’s limits. The analysis is observational. It does not identify who ran the bots, their nationality, or their intent. Roughly 55 percent of accounts in the peak core were classified as automated — a mix that plausibly includes commercial spam, activist automation, and state-linked actors. What the paper claims is that the mechanism is real. Attribution is a separate investigation.
The Blind Spot In U.S. Doctrine
The State Department’s Framework to Counter Foreign State Information Manipulation is organized around five action areas focused on narratives, exposure, and public resilience. Carnegie’s 2024 evidence-based guide to counter-disinformation evaluates ten interventions — every one content- or belief-oriented. None systematically measures whether the human network inside a targeted community is intact
.That blind spot is widening. Congress terminated the Global Engagement Center on December 23, 2024, after declining to reauthorize it in the NDAA. Its successor, the Counter Foreign Information Manipulation and Interference Office, was substantially dismantled in 2025. Meanwhile, the Justice Department disrupted a Russian AI-enhanced bot farm operated by RT and FSB affiliates targeting U.S. audiences — evidence that adversary automation is scaling, not receding.
The Bots Within
The new paper builds on prior work by University of Washington researcher Kate Starbird, who warned in 2018 that Russian information operations were “targeting us within our online communities, the places we go to have our voices heard.” The same team behind the new preprint showed in 2024 that bot interactions soured activist sentiment without eroding engagement — users kept posting but felt worse. The 2026 paper closes the loop: engagement holds, but the network beneath it decays.
The authors’ bottom line is stark. “The primary threat of highly sophisticated automation,” they write, “may not lie in its power of ideological persuasion, but in its capacity to occupy online social spaces in ways that fragment human participation to the extent that movements will disappear.”
What Comes Next
The policy implication is direct. Platforms and government analysts should track bot-to-human relational density and monitor the decay of human ties inside targeted communities — not just posts, narratives, and beliefs.
A movement can look loud on a dashboard while its organizing capacity quietly hollows out. Content-based counter-disinformation, on its own, will keep missing it. Adversaries who understand this mechanism would not need to convince Americans of anything. They would only need to keep them from finding each other.