A new warning about AI and democracy is not focused on a single fake video or one misleading post. It is about scale, coordination, and the possibility that one person with advanced AI tools could direct thousands of social media accounts that look and behave like real people.
A paper published in Science argues that disinformation campaigns may be approaching a new phase. The concern is that AI-controlled agents could maintain identities, remember past interactions, coordinate with one another, and adjust their behavior as online conversations unfold.
From troll farms to AI swarms
The article begins with a comparison to 2016, when hundreds of Russians worked from 55 Savushkina Street in St. Petersburg as part of the Internet Research Agency. Those employees manually commented on news articles, posted on Facebook and Twitter, and tried to provoke Americans during the presidential election.
That operation became infamous after it was uncovered. It drew media coverage, Senate hearings, and changes from social media platforms around user verification. Yet the source notes that its impact was minimal compared with another Russia-linked campaign involving Hilary Clinton’s emails being leaked just before the election.
The new paper’s argument is that the next model could be very different. Instead of a large office filled with workers, the researchers describe a scenario in which a single operator uses AI agents to run “swarms” of thousands of accounts. These accounts would not merely repeat canned messages. They could generate unique posts, respond to human users, and evolve without constant supervision.
Why humanlike agents are harder to fight
The central concern is not only that AI can produce content quickly. It is that AI agents could simulate believable online personas over time. The systems described by the researchers would have persistent identities and memory, allowing them to appear more consistent and credible than older bot accounts.
They could also coordinate around shared goals while producing individual messages that do not look identical. That matters because many detection systems are built around finding coordinated inauthentic behavior. If every account appears to have its own voice, history, and pattern of interaction, the activity becomes harder to separate from normal online conversation.
The report puts the stakes directly: “Advances in artificial intelligence offer the prospect of manipulating beliefs and behaviors on a population-wide level,” the report says. “By adaptively mimicking human social dynamics, they threaten democracy.”
The paper was authored by 22 experts from fields including computer science, artificial intelligence, cybersecurity, psychology, computational social science, journalism, and government policy. Other specialists who reviewed the paper also described the threat as serious.
Targeting communities, not just platforms
The source article emphasizes that the threat is not limited to accounts that look human. The researchers also point to the ability to map social networks at scale. That could allow those running disinformation campaigns to place AI agents in specific communities where they may have the greatest effect.
According to the researchers, swarms could tailor messages to the beliefs and cultural cues of each community. That would make the targeting more precise than earlier botnets. Instead of broadcasting the same message broadly, an AI-driven system could test what works inside different groups and refine its approach.
The paper also describes the possibility of self-improvement. If posts generate enough feedback, the systems could use responses as signals, run “millions of microA/B tests,” spread the strongest variants, and keep iterating at machine speed.
That makes the issue more than a content moderation problem. A platform may be dealing with accounts that learn from the environment, adapt to human reactions, and shift tactics faster than traditional responses can keep up.
What experts say about the risk
Lukasz Olejnik, a visiting senior research fellow at King’s College London’s Department of War Studies and the author of Propaganda: From Disinformation and Influence to Operations and Information Warfare, said targeting individuals or communities will become easier and more powerful. “This is an extremely challenging environment for a democratic society. We're in big trouble.”
Barry O’Sullivan, a professor at the School of Computer Science and IT at University College Cork, said AI-enabled influence campaigns are already within the current state of the technology. He also said the paper shows how difficult governance and defensive responses may become.
Jonas Kunst, a professor of communication at BI Norwegian Business School and one of the report’s coauthors, said the classic bot approach is becoming outdated. He also said it is unclear whether this tactic is already being used, because current systems for tracking coordinated inauthentic behavior may not be able to detect it.
Kunst added that these systems are likely to still have some human oversight as they are being developed. He predicted they may not have a massive impact on the 2026 US midterms in November, but are very likely to be deployed to disrupt the 2028 presidential election.
The proposed response
To address the threat, the researchers propose an “AI Influence Observatory.” The idea is to bring together academic groups and nongovernmental organizations to “standardize evidence, improve situational awareness, and enable faster collective response rather than impose top-down reputational penalties.”
Notably, the proposal does not include executives from social media platforms. The researchers argue that these companies prioritize engagement, which could limit their incentive to identify AI swarms if the activity boosts usage, ad views, and apparent platform vitality.
That creates a difficult tension. If swarms become so common that users stop trusting anyone and leave, the platforms would face a direct threat. But if the activity increases engagement in the short term, the researchers suggest platforms may have less reason to reveal it.
The paper’s warning is therefore not just that AI can generate more disinformation. It is that AI could make influence operations more adaptive, more targeted, and more difficult to see. The central question is whether democratic societies can build detection and response systems before these swarms become a normal part of online politics.