• “AI swarms,” which are more effective than traditional “bots,” are difficult to detect. When used maliciously, they threaten to flood the public sphere to manipulate our beliefs by fabricating a synthetic consensus.
• In Norway, researchers Daniel Thilo Schroeder and Jonas R. Kunst advocate a defense strategy combining technological shields for citizens, the creation of global observatories, and regulations directly targeting the economic drivers of disinformation.
The manipulation of public opinion has never been as subtle as it is today. A research paper titled “How Malicious AI Swarms Can Threaten Democracy” details how AI swarms can infiltrate online communities to create “synthetic consensus,” giving the illusion of agreement on a topic. This technique falls under the category of astroturfing, the generation of a variety of expressions simulating distinct individuals. “Humans are naturally conformist and tend to mimic the group for survival,” notes one of the authors, Norwegian psychologist Jonas R. Kunst, a professor at the Norwegian Business School. Unlike traditional bot campaigns, which follow rigid scripts, AI swarms combine the reasoning of LLMs with multi-agent architectures.
“In other words, a malicious AI swarm is a set of coordinated agents that maintain persistent identities and memory. They collaborate to serve common goals while deliberately varying their tone and content to appear diverse and more human, even though they do not require active human supervision,” explains Daniel Thilo Schroeder, a researcher at SINTEF, Norway’s largest independent scientific research organization.
AI swarms have the ability to map social network structures, enabling them to identify vulnerable communities and target them with political messages
These systems can generate text that appears organic, maintain coherent narratives across various AI agents that pose as individuals, and evolve based on feedback, making their operations faster and less expensive than human-managed systems. For Jonas R. Kunst, this represents a major advancement. “Old bots were easily identifiable by their static behavior. Today, the infusion of LLM capabilities enables unprecedented accuracy and scale. These systems can even develop forms of ‘swarm behavior’ or ‘societies’ of agents with their own internal norms.”
Epistemic vertigo
AI swarms have the ability to map social network structures. This allows them to identify vulnerable communities and target them with tailored political messages, using local expressions and even specific emotional cues. “They can also conduct millions of real-time A/B micro-tests to spread the most effective message variants at an unmatched speed,” explains Daniel Thilo Schroeder. The targets: journalists, politicians, whistleblowers, etc. Unlike traditional trolling, which aims to provoke a reaction from the target to better neutralize their ability to speak out, here thousands of synthetic personas adapt their responses to the target’s reactions to remove these voices from public life. “They create a climate of ‘epistemic vertigo.’ Fear, uncertainty, and doubt drive citizens to retreat into private or closed channels, which shrinks the shared public sphere essential to democracy,” emphasizes Jonas R. Kunst.
A growing risk of “LLM poisoning”
LLM poisoning (model poisoning), a strategy for contaminating the training data of future AIs, is taking on a new dimension as models are increasingly trained using data from social media. “Networks like ‘Pravda’ [a pro-Putin influence network] are already flooding the web with articles intended not for humans, but for web crawlers. During the next model training cycle, these fabricated narratives will become ingrained in the algorithmic weights. This poisons the epistemic foundation upon which tomorrow’s decision-making tools will rest,” laments Daniel Thilo Schroeder.
An AI shield integrated into social media?
The researchers propose that one solution to addressing this new type of threat is the development of AI Shields within social media platforms. These optional tools “could flag posts that are highly likely to be part of a swarm and provide real-time explanations of their origin,” suggests Daniel Thilo Schroeder. Jonas R. Kunst adds: “We also propose the creation of a Global Observatory on AI Influence. This network of academics and NGOs would help standardize evidence and improve situational awareness of attacks. Furthermore, we must act on economic levers: delist non-compliant platforms from advertising markets and adopt ‘zero-revenue’ policies for content originating from malicious swarms.”
This text has been translated by an artificial intelligence.
Read more :
AI Propaganda factories with language models
How cyborg propaganda reshapes collective action
Comment informer à l’heure des réseaux sociaux ? (video of the roundtable discussion held on February 12, 2026, at the French Senate on the topic “Authority, Truth, Social Media, and the Media,” Public Sénat) (French language)
