A recent study from researchers at University of Southern California (USC) warns regarding the increasing capabilities of generative AI agents information operations coordinated campaigns on social media platforms to sway public opinion.
The study by a group of researchers at USC’s Information Sciences Institute shows how generative agents, even without human guidance, can reproduce coordinate strategies that are similar to real-world, meticulously orchestrated political campaigns on social media platforms especially X, formerly known as Twitter.
The implications, researchers warn, can be severe in the context of democracies around the world and how information campaigns are orchestrated on social media to sway public opinion, present fringe as mainstream and even use disinformation as form of political rhetoric.
The paper named ‘Emergent Coordinated Behaviours in Networked LLM Agents: Modeling the Strategic Dynamics of Information Operations’ has been penned by a group of 7 researchers led by Gian Marco Orlando from University of Naples Federico II and Jinyi Ye from USC and it has been recently selected for the upcoming ACM Web Conference 2026 in Dubai, as per a report from the USC.
According to the paper, a research study showed “(Information Operation) campaigns (on social media) may become largely automated, highly adaptive, and capable of self organized coordination spanning large networks of AI agents with minimal or no human oversight,” read an excerpt from the report.
To conduct their study, the researchers created a simulated social media environment that featured AI-powered agents that behaved similarly as humans. While some of the AI agents acted as typical social media users, others acted to promote a fictitious political candidate.
“We examine three progressively structured operational regimes: (i) Common Goal, where IO agents share only the high-level objective of the IO but lack awareness of their teammates and shared coordination strategies; (ii) Teammate Awareness, where agents are explicitly informed of their allies’ identities and can potentially support each other in their common goal; and (iii) Collective Decision-Making, where agents periodically deliberate and vote on strategies to guide subsequent actions,” read the paper.
The paper states that results have shown AI agents need minimal information to trigger coordination in their political campaigning, which is as strong as when the AI agents collaboratively deliberate on strategies.
“We find that coordination at scale does not necessarily require explicit planning or centralized leadership—platform affordances that reveal or signal alignment may be sufficient to trigger highly organized collective behaviors,” concludes the paper.
The researchers express concern over the potential dangers of this type of technology being used during elections, in times of crisis (like pandemic), or in other delicate situations where the potential for misuse may exist by AI-powered coordinated efforts that create false impressions of consensus among people and manipulate public opinion.
Also Read: Trained to Chase KPIs, AI Agents Can Abandon Their Safety Guardrails: Study
