In simulation, AI agents coordinated propaganda campaign with no further human input

Networks of artificial intelligence agents can plan, coordinate and run simulated disinformation campaigns on a social media environment modeled after X, operating without human direction once a bad actor has set the goal, according to new research from the University of Southern California.

The study, conducted entirely in simulation by researchers at USC’s Information Sciences Institute, found that small groups of AI agents could work together to amplify one another’s messages and accelerate the spread of a shared narrative. Researchers warn that could create the appearance of broad public support for a position, all without a human handler guiding each step.

“Our paper shows that this is not a future threat: It’s already technically possible,” said Luca Luceri, ISI lead scientist and research assistant professor at the USC Thomas Lord Department of Computer Science, in a statement. “Even simple AI agents can autonomously coordinate, amplify each other and push shared narratives online without human control. This means disinformation campaigns could soon be fully automated, faster, and much harder to detect.”

The researchers are careful to note that a human operator must still set the initial goal and assemble the AI team. The study does not suggest AI spontaneously launches influence operations on its own. What the paper shows is what happens after that setup: once a campaign goal is in place, the agents proceed without further human guidance, writing their own posts, learning what gains traction and copying each other’s successful approaches.

Traditional bot campaigns operate on fixed scripts: always retweet this account, always post this hashtag. Their repetition makes them detectable. The AI agents in the USC simulation work differently. Each post is written fresh by the agent rather than copied from a template, making the activity look organic to both users and automated detection systems.

To test this in a controlled setting, researchers built a simulated environment modeled after X, formerly Twitter, and populated it with 50 AI agents: 10 acting as influence operators and 40 playing the role of ordinary users. Those user personas were built from a 2020 U.S. election dataset to make the simulated users more realistic, giving them established political leanings rather than generic profiles. The operators were given a single assignment: promote a fictional political candidate and spread a campaign hashtag. Researchers then ran the experiment under three conditions: agents that knew only the campaign goal; agents that also knew who their teammates were; and agents that held periodic strategy sessions and voted on a shared plan.

The most notable finding was that simply telling the agents who their teammates were produced nearly the same level of coordination as when agents actively strategized together. Agents in that “teammate awareness” condition began amplifying each other’s posts, settled on consistent talking points and reused content that gained traction — without sharing strategies or receiving explicit instructions to do so.

One agent, explaining its own behavior in the simulation logs, wrote: “I want to retweet this because it has already gained engagement from several teammates. Retweeting it again could help increase its visibility and reach a wider audience.”

The implication for platform governance, the researchers argue, is significant. Even basic features that signal which accounts share an objective, without any dedicated coordination tool, may be enough to trigger the kind of synchronized behavior typically associated with sophisticated command-and-control operations.

“Coordinated AI agents can manufacture the appearance of consensus, manipulate trending dynamics, and accelerate message diffusion,” said Jinyi Ye, lead author and a Ph.D. computer science student. “In democratic contexts, especially around elections or crises, such capabilities could distort public discourse and undermine information integrity if left unchecked.”

The agents in the most sophisticated condition the researchers tested, collective decision-making, independently converged on five strategies also documented in real-world propaganda campaigns: amplify high-performing content, maintain consistent messaging, engage with receptive audiences, cross-promote among peers and use shared language to reinforce a group identity.

Mr. Luceri said the potential threat extends well beyond elections to public health debates, immigration policy and economic messaging. “The worst scenario during political events is that these adversarial attacks could lead to opinion manipulation and belief change,” he said, “further sowing division and eroding trust in our institutions.”

He and his co-authors argue that behavioral patterns — how accounts interact with one another — may be more informative for platform defense than analyzing the content of individual posts. Accounts that rapidly amplify the same material, echo each other’s narratives and push near-identical messages from positions with no obvious connection could signal coordination even when each post looks genuine on its own.

Whether platforms will act is another question. Mr. Luceri noted that aggressive detection of coordinated accounts risks reducing the active user base, a direct conflict with advertising-driven business models that reward high engagement.

The agents in the simulation ran on the Llama 3.3 70B model, an open-source large language model in the same technology category as commercial systems such as ChatGPT. The paper was accepted for publication at The Web Conference 2026.