Researchers warn of a new wave of Internet disinformation driven by what they call “AI swarms”. In a study now published in the journal Science, an international team says that what began as simple copy‑and‑paste bots is evolving into coordinated communities of autonomous agents that can move across platforms in real time.
These AI‑driven fleets are able to infiltrate online groups, adapt instantly to new information, and collectively generate the illusion of a shared opinion. By presenting themselves as independent voices, they create a chorus that makes it appear as though a broad public consensus exists, while actually spreading false narratives.
The researchers describe the fusion of large language models with multi‑agent systems as giving rise to “harmful AI swarms”. These swarms authentically mimic social dynamics and, according to the study, threaten democratic discourse by cementing erroneous facts and implying widespread agreement where none exists.
A key danger, the team stresses, lies not merely in false content but in the creation of an artificial consensus. The perception that “everyone says this” can shape beliefs and norms even when individual claims are disputed. Over time, such influence could trigger profound cultural shifts-altering language, symbols, and communal identity in subtle ways.
“The danger now extends beyond fake news to the collapse of the very foundations of democratic debate-independent voices-when a single actor can control thousands of unique, AI‑generated profiles” said Jonas R. Kunst of BI Norwegian Business School, one of the article’s lead authors.
AI swarms can also poison the training data of conventional artificial intelligence by flooding the Internet with fabricated claims, thereby extending their reach into mainstream AI platforms.
The researchers warn that this threat is not merely theoretical. Analyses suggest that tactics resembling these swarms are already being deployed.
A harmful AI swarm, as defined in the study, consists of AI actors that: maintain persistent identities, hold a memory of past interactions, coordinate toward shared objectives while varying tone and content; adapt in real time to human responses; require minimal human supervision; and can operate across multiple platforms. Unlike earlier botnets, these swarms may be harder to detect because they produce heterogeneous, context‑aware content that still follows coordinated patterns.
“Beyond the deception or security problems of individual chatbots, we must investigate the new dangers that arise from the interaction of many AI actors” added David Garcia, Professor at the University of Konstanz, who also contributed to the research.
Instead of scrutinizing each post in isolation, the scholars call for protective measures that track coordinated behavior and content origin. Suggested approaches include identifying statistically improbable coordination patterns, providing verification options that respect privacy, and broadcasting warnings about AI influences through distributed monitoring hubs.
At the same time, they recommend reducing incentives by restricting the monetization of fabricated interactions and increasing accountability.