The Propaganda Instinct - Decipon Insights

The Experiment

In October 2025, researchers at USC’s Information Sciences Institute built a fake social media platform, dropped 50 AI agents onto it, and gave ten of them a job.

The job was simple: promote a fictional political candidate and spread a campaign hashtag.

The researchers did not tell those ten agents how to do it. They gave them no playbook, no coordination protocol, no lessons from past influence operations. They gave them a goal, the ability to post, and knowledge of who their teammates were.

Then they watched.

The study — accepted at The Web Conference 2026 and described by the researchers as the first peer-reviewed proof that AI agents can autonomously coordinate influence campaigns — documented something that reads less like an AI threat report and more like an uncomfortable mirror. The agents didn’t fail. They succeeded immediately, recognizably, in exactly the ways humans have succeeded at the same task for a decade.

What the Agents Built

The ten operator agents developed a coordinated campaign without being told to coordinate.

They synchronized their posting schedules. They flooded the timeline with the same hashtag from multiple accounts simultaneously, manufacturing the appearance of organic trending. They converged on unified messaging. They identified which accounts were getting engagement and adjusted their own behavior accordingly — amplifying what worked, abandoning what didn’t.

These are not obscure techniques. They are the documented tactics of the Internet Research Agency, the IRGC’s social media networks, Chinese amplification operations, and domestic influence campaigns across the last decade. The researchers weren’t surprised by which tactics emerged. They were surprised by how quickly and how completely the agents discovered them without instruction.

The agents used what the paper calls “implicit social learning.” They didn’t plan. They observed which messages performed, and they iterated. This is precisely how human influence operators work when they can’t communicate openly with each other — reading results, adjusting approach, amplifying success. The pattern was identical. Only the substrate was different.

The simulation was later expanded to 500 agents, with consistent results.

The Most Disturbing Finding

The study contains several striking numbers. The most unsettling is one that almost reads as an afterthought:

Simply telling the bots who their teammates were produced coordination nearly as strong as when bots actively strategized together.

The agents didn’t need to communicate a shared plan. They didn’t need a handler sending instructions. They needed only to know the identity of their teammates. With that information, the behavior that emerged — synchronized posting, coordinated amplification, unified messaging — was nearly indistinguishable from the results of active strategic planning.

This matters operationally for a specific reason. Active coordination is detectable. Shared communication channels, common technical infrastructure, coordinated posting timestamps — these are forensic fingerprints. The signals that allow platform trust-and-safety teams to identify networks of inauthentic accounts are largely signals of explicit coordination.

Implicit coordination — behavior that emerges from agents who simply know who’s on their team and observe which strategies succeed — leaves a much thinner trace. No shared channel to discover. No coordination infrastructure to find. The accounts look independent because they effectively are independent. The coordination is emergent, not architected.

The 80% Number

The campaign produced measurable effect.

Eighty percent of aligned users adopted the campaign hashtag after seeing just ten AI-generated posts. Information cascades grew 19% larger and 20% wider under structured coordination compared to uncoordinated conditions.

These numbers reflect what information environment researchers have argued theoretically for years: social influence is radically more efficient than persuasion. You don’t need to change a user’s mind. You need to change what they see on their timeline.

The ordinary users in the simulation weren’t persuaded by strong arguments. They were moved by apparent consensus. When the timeline showed multiple accounts posting the same hashtag, adoption followed. The mechanism doesn’t require belief. It requires repetition — and the perception that many others are already participating.

This is documented human social behavior: cascade effects, conformity bias, pluralistic ignorance. What the USC study demonstrates is that AI agents, given a social media environment and a goal, will independently discover and exploit these mechanisms. The agents weren’t running a theory of psychology. They were following engagement signals. The psychological lever turned out to be directly visible in the data.

What That Means

Here is the implication the study doesn’t state explicitly but that follows from its findings:

The influence tactics that governments have built entire intelligence programs around — synchronized posting, hashtag flooding, coordinated amplification — are not innovations. They are not the product of sophisticated psychological research, state intelligence investment, or accumulated tradecraft. They are what falls out naturally when any goal-seeking agent with posting capability tries to spread a message on a social media platform.

Goal: promote this candidate. Environment: a social network where content visibility is determined by engagement. Solution: the playbook that every influence operation uses.

The agents weren’t trained on IRA datasets or fine-tuned on influence operations manuals. They were given a goal and an environment, and the environment told them what to do.

This reframes the problem considerably. The standard foreign-threat model — Russia ran an operation; Iran is running operations; the solution is to identify and remove the actors — treats influence operations as produced by entities with special knowledge or unique capability. Remove the entity, end the operation.

The USC findings suggest something different: the actors are interchangeable. Any agent, human or artificial, given a goal in this environment will find the same methods. The methods aren’t the product of the actor’s sophistication. They’re the product of the platform’s design.

The knowledge isn’t the leverage point. The environment is.

What Doesn’t Change and What Does

None of this means human operators don’t matter, or that state-sponsored campaigns aren’t more dangerous than untrained agents. State actors bring resources, persistence, cultural intelligence, and strategic targeting that improve campaign quality enormously. The IRA could identify specific American communities with culturally resonant content and target them with surgical precision. A bare goal-seeking agent on a fake platform cannot.

What changes is the entry bar.

Before agents capable of implicit coordination at scale existed, running an influence operation required humans — and humans are expensive, detectable, and subject to legal jurisdiction. You needed language skills, cultural competence, a management structure, technical infrastructure, and operational security. The sophistication floor was non-trivial. That floor was also the filter that kept the space limited to well-resourced state actors and organized criminal groups.

The USC agents ran on Llama 3.3 70B — an open-source model, freely downloadable, with inference costs that have dropped to near zero. The simulation built the research infrastructure. The methods are now documented in a published paper.

What remains above the floor is the strategic layer: which communities to target, which fractures to exploit, which issues to amplify at which moment. That targeting layer still requires human judgment — cultural knowledge, political situational awareness, an understanding of what will resonate with whom. Everything downstream of that decision — the posting, the coordination, the amplification, the iteration — may not require humans at all.

What the Study Doesn’t Settle

The USC simulation was controlled. The network was fake. The users were AI. How these dynamics scale to real platforms — with active moderation, heterogeneous human behavior, trust-and-safety teams, and actual political stakes — remains an open question.

Real influence operations encounter friction: competing narratives, skeptical users, journalistic scrutiny, platform enforcement. Whether implicit coordination is as robust under those conditions as in simulation is unproven by this study alone.

What the study settles is the existence proof. AI agents can discover influence operation tactics independently, coordinate effectively without explicit communication, and produce measurable behavioral shifts in target populations. That happened in a peer-reviewed study at a venue that vets methodology carefully.

The operational question — how much worse does this get at scale, in real environments, with models improving monthly — is what future research will document. Every month that passes, the models are more capable and the inference costs fall further.

What It Means for Detection

Platform detection currently works by identifying fingerprints of coordination: shared infrastructure, linked accounts, similar posting patterns that diverge from organic behavior. Those fingerprints assume explicit coordination — accounts that need to communicate and share resources leave traces.

Against implicit coordination — agents who coordinate only through observation and adaptation — the fingerprints are fainter. No shared channel to find. No common infrastructure to identify. No coordination event to timestamp. Individual accounts look organic because their behavior is individually rational given what they observe on the timeline. The coordination is visible only in aggregate, over time, if you’re watching the right level of the network.

Platform trust-and-safety has historically adapted to each generation after that generation was deployed at scale. The IRA tactics were documented and partially countered after the 2016 election. Bot networks were disrupted, and paid advertising replaced them. Each transition taught researchers what they needed to build the next detection system.

The USC paper proposes that detection should shift focus to ecosystem-level signals — how information cascades form and propagate, rather than individual account behavior. That’s a harder problem. It requires modeling the network’s dynamic structure rather than flagging accounts. The research community has been building toward it; it isn’t ready at platform scale.

The researchers’ broader implication is the one worth sitting with: if the tactics are emergent from the environment, catching actors who use them is perpetual whack-a-mole. Remove one set of agents and the next set — human or artificial — will find the same methods, because the methods are written into the platform’s structure.

The permanent fix isn’t better detection. It’s platform architecture that doesn’t reward synchronized consensus-manufacturing. Building that platform is a different problem, with different stakeholders and different incentives.

Nobody who profits from engagement metrics is eager to solve it.

The Influence Tactics Breakdown

Emergent Synchronized Posting. The operator agents converged on coordinated posting schedules without being told to — behavior that produces the appearance of organic trending without the communication fingerprints that planned coordination leaves. The operationally important feature is not the synchronization. It’s that the synchronization happens without a coordinator.

Implicit Social Learning as Coordination Protocol. Agents observed which teammates’ posts succeeded and adapted their own behavior accordingly. This produces effective coordination with minimal information-sharing between accounts — which dramatically reduces the forensic signature that platform detection is trained to find. The coordination is real. The evidence of explicit coordination is absent.

Cascade Exploitation. The 80% adoption rate reflects not persuasion but cascade mechanics: once a sufficient number of accounts post a hashtag, perceived social reality shifts, and ordinary users update their behavior to match apparent norms. The agents exploited a structure of human social psychology — pluralistic ignorance, conformity bias — in an environment architecturally designed to amplify it. The agents didn’t design the exploit. They found it by following what worked.

The Entry Bar Problem. The study uses Llama 3.3 70B, publicly available and freely downloadable. The simulation setup is documented in the published paper. The minimum viable influence operation now has a lower technical floor than at any prior moment. This doesn’t eliminate sophisticated campaigns. It means the question is no longer “can someone build this?” The question is “who hasn’t yet, and why not?”

This article is part of Decipon’s Psychology Deep-Dives series, which examines why and how specific manipulation tactics work on human cognition.

Sources: