AI “swarms” could fake public consensus and quietly distort democracy, Science Policy Forum warns

Published today, a new Science Policy Forum article warns that the next generation of influence operations – coordinated campaigns designed to manipulate perceptions of consensus, credibility, and normality – may not look like obvious “copy-paste bots,” but like coordinated communities: fleets of artificial intelligence (AI) -driven personas that can adapt in real time, infiltrate groups, and manufacture the appearance of public agreement at scale.

In the world-renowned journal, the authors, including Andrea Baronchelli, Professor of Complexity Science at the School of Science & Technology, City St George’s, University of London, and from twenty further academic institutions, describe how the fusion of large language models (LLMs) with multi-agent systems could enable “malicious AI swarms” that imitate authentic social dynamics – and threaten democratic discourse by counterfeiting social proof and consensus.

The article argues that the central risk is not only false content, but synthetic consensus: the illusion that “everyone is saying this,” which can influence beliefs and norms even when individual claims are contested. According to the authors, this risk compounds existing vulnerabilities in online information ecosystems shaped by engagement-driven platform incentives, fragmented audiences, and declining trust.

The authors define a malicious AI swarm as a set of AI-controlled agents that can maintain persistent identities and memory; coordinate toward shared objectives while varying tone and content; adapt to engagement and human responses; operate with minimal oversight; and deploy across platforms. Compared with earlier botnets, such swarms could be harder to detect because they can generate heterogeneous, context-aware content while still moving in coordinated patterns.

Instead of moderating posts one by one, the authors argue for defenses that track coordinated behavior and content provenance: detect statistically unlikely coordination (with transparent audits), stress-test social media platforms via simulations, offer privacy-preserving verification options, and share evidence through a distributed AI Influence Observatory—while also reducing incentives by limiting monetization of inauthentic engagement and increasing accountability.

Reflecting on the article, Professor Baronchelli said:

This Science Policy Forum focuses on a shift in democratic risk, from persuasion to the manipulation of perceived social norms. Our earlier work showed that AI agents can spontaneously develop shared conventions without central control. In information ecosystems, the same collective dynamics can be exploited at scale, which is why governance must move beyond single-model safety and address multi-agent AI systems as a policy priority.