Log in
Every day, millions of pieces of fake content are produced. Videos, audio clips, posts, articles, generated by artificial intelligence, distributed at industrial scale, aimed at shifting public opinion across entire countries. The people producing them are often outside the country being targeted. The people receiving them almost never know they’re fake. And they have no idea how they’re made.
A few years ago, troll farms worked like this: entire buildings full of people, shifts, desks and workers paid to write posts, create fake profiles, comment and pick fights in online discussions. It was expensive, slow, and in the end, the real impact was marginal. Those buildings still exist today, mostly in India, split between teams specializing in scams and teams dedicated to disinformation. They work on commission and they’re mostly AI experts now. They no longer write the articles themselves and no longer do graphic design or photo editing. They have AI agents do everything: agents they create, configure, instruct, and supervise. Hundreds of thousands of autonomous agents that do in one hour what used to take weeks of human labor. Troll farms have become AI farms, producing synthetic content at industrial scale.
The report “From Trolls to Generative AI: Russia’s Disinformation Evolution,” published in February of 2026 by the Centre for International Governance Innovation (CIGI), tells one of these stories, specifically about disinformation campaigns originating from Russia. Networks like CopyCop, a disinformation operation linked to the GRU (Russian military intelligence), use uncensored open-source language models like modified versions of Llama 3, installed on their own servers, to transform press articles into political propaganda and distribute it across hundreds of fake websites without leaving a trace. Because the models run locally, there’s no watermark and no log. The model runs on their hardware, inside their borders, outside any Western jurisdiction.
The paper “How malicious AI swarms can threaten democracy,” published in Science in January 2026 describes well what is coming: coordinated swarms of AI agents with persistent identities, memory, and the ability to adapt in real time to people’s reactions. The authors call them “malicious AI swarms.” Fully autonomous agents, each producing original content, each one different, each one adapted to context.
They can simulate real communities that appear credible, and they build what we can call synthetic consensus: the illusion that an opinion is widely shared, that a position is held by the majority, when in reality it’s a single operator speaking through thousands of masks.
It works because we humans have bugs too, and the swarms exploit them at a scale that was never possible before or that would have required enormous human resources.
One bug is called the bandwagon effect. Combined with another bug, illusory truth: repetition plus apparent source independence equals perceived truth. So if we see the same position expressed by different sources, in different contexts, with different words, on different platforms, we register it as widespread. And if we perceive it as widespread, we consider it more credible. And if we consider it credible, we tend to align with it.
Swarms of autonomous agents exploit both mechanisms at the same time, at industrial scale.
What most people still haven’t grasped is the scale. We were used to automation: A system that sent a hundred thousand identical emails, at most changing the name and little else, or made just as many posts and similar comments with minor variations. It automated the publishing, but at its core it was recognizable spam. Our mental model is still that one: If it’s automated, it’s generic. If it’s generic, you can spot it. But that’s a perception error built on years of experience when AI agents didn’t exist. That model is over. These agents no longer fit the concept of automation, because they make decisions, they radically change the text based on the recipient. They aggregate data from heterogeneous sources in real time: social profiles, public records, leaked databases that you can now buy for a few dollars on any dark web marketplace. Billions of personal records are already out there, scattered across hundreds of breaches accumulated over the years, and AI can cross-reference them, reconcile them, and build a coherent profile of a single person in seconds. The computational cost is negligible: a few cents in tokens to generate a perfectly personalized message. Consider that a single agent with access to a language model and a couple of leak databases can produce thousands of unique pieces of content per day, each calibrated for a different person. Multiply that by a hundred thousand agents working in parallel, twenty-four hours a day, and you have the scale of what’s happening.
Another legacy from the past: “I’m just an ordinary person, why would anyone bother creating content specifically to convince me?” That may have been once true. Today, nobody is losing time because these agents don’t get tired, don’t sleep, and do nothing else: find connections, aggregate data, produce false content calibrated for each of us. The old demographic profiling is over. This is surgical media targeting at industrial scale.
But the capacity to respond and deny is not at industrial scale. If hundreds of thousands of coordinated agents spread a video of a politician saying something they never said, that politician can deny it all they want. The video is there. Millions of people have seen it. The denial arrives later, arrives slower, and will never reach the same scale. It arrives in a world where nobody knows what’s true anymore.
If the same swarms spread the news that a head of state has died, and the news is false, that head of state can make all the videos they want to prove they’re alive. Those videos will probably be dismissed as deepfakes. Because the swarm’s narrative got there first, took root, and at that point any evidence to the contrary looks fabricated.
Whoever controls the swarms today controls the version of the facts. Whoever tries to push back is already at a disadvantage because they have to prove that a real video is real in a world where everyone has learned that videos can be fake.
The attackers are often outside the country being hit. Groups aligned with governments that want to shift public opinion in another country, or that target specific demographics. Young people, for example, using platforms that are often owned by those very countries.
All of this is a massive threat to democracy because democracy operates on some premises, including that people form opinions based on real information, discuss with each other, and then decide. If the information is fabricated, if the debate is populated by entities that don’t exist, if the consensus we perceive is synthetic, that premise collapses. And with it, the entire mechanism. Elections become the result of who has the best swarms, not who has the best ideas. Public debate becomes a performance where most of the voices are generated, and public opinion stops being public and becomes the product of whoever has the resources to manufacture it.
We grew up thinking that threats to democracy came from coups, censorship, or regime propaganda broadcast on television or in national newspapers. Those were real threats, but they were at least visible. They were things you could identify and fight. Now the threat is bigger and, above all, invisible, personalized, and it operates inside the very channels we use to inform ourselves, to discuss, to participate. It contaminates information from within, to the point where nobody knows which voices are real and which are machines.
What can we do? Watermarking? Pattern detection? Unfortunately, they don’t work. The major AI platforms can embed markers in content generated by their models, true. But the people building autonomous swarms don’t use commercial platforms. They use open-source models with fine-tuning and capabilities that can’t be controlled from outside. And they often have no legal obligation to do anything because there are no global laws that can impose watermarking on every computer in the world. The result is paradoxical: The content produced by those who follow the rules stays marked, and the content produced by those who want to cause harm stays free.
Pattern detection systems have the same limits. They work for a while, then once the detection patterns are identified, the swarms adapt. They’re designed to do exactly that.
And the platforms where all of this circulates have a financial incentive to turn a blind eye. Internal Meta documents made public by Reuters in November 2025 estimated that roughly 10% of Meta’s global 2024 revenue, about $16 billion, came from advertising for scams and prohibited products. Fifteen billion high-risk ads served on average every day to users. The maximum revenue Meta was willing to sacrifice to act against suspicious advertisers was 0.15% of total revenue: $135 million out of $90 billion. When a platform’s business model depends on ad volume, removing the fraudulent ones has a cost that nobody wants to pay. I suspect Meta is not alone in this.
Regulation doesn’t solve this problem either. I’ve worked on the European AI framework, the GPAI task force, the Italian AI law, and I’ve brought my perspective to the UK Parliament. I’ve been in those rooms. Europe has the AI Act, the GPAI Code of Practice is currently being drafted, and has a regulatory apparatus that is more advanced than any other bloc in the world. The United States has no federal regulation, and twenty-eight states have tried to legislate with transparency requirements that amount to fine print. But even the most ambitious European framework has a structural limit: The attacks come from countries that answer to none of these rules. You can regulate your platforms, your developers, your companies. You can’t regulate a building in Saint Petersburg, Shenzhen, or New Delhi, where someone is instructing swarms of agents on open-source models running on local servers, outside any jurisdiction.
One way out is to return to the reputation of sources. Editors, news organizations, journalists with a name and a face. People and organizations that have a professional track record to defend and that risk something when they get it wrong. Sure, they can have political leanings and they can make mistakes. But they have a constraint that no AI agent will ever have: public accountability. A system that generates millions of pieces of false content answers to no one. An editor answers to their audience, to the law, to their reputation. That constraint is the only filter that still holds, and protecting it is the only thing we can do right now, while the laws try to catch up with a technology that moves faster than any legislative process in the world.
Are we completely at the mercy of AI swarms or can we fight back?
Machines should not get to overpower humans, especially when what’s at stake is how we govern ourselves. The antibodies exist. We need to activate them.
The more people understand how swarms work, the less effective they become. A swarm that manufactures fake consensus only works if the people receiving it don’t know synthetic consensus exists. A bit like deepfakes. We know about them now and we often spot them. Once you see how it works, it’s harder to fall for it.
Then we need investment in culture. In spreading digital literacy, which is not learning how to use a computer, but learning to understand the social and cultural effects of the digital world. It means teaching in schools how to verify a source and what the signs of manipulated content are. It means stopping the practice of treating media literacy as a school project and starting to treat it as democratic infrastructure, on the same level as bridges and hospitals. It means funding independent journalism instead of letting it die, strangled by the same mechanisms that reward false content because it generates more engagement. It means demanding that platforms give different visibility to those who have a verifiable reputation versus those who have none.
Because awareness is the only antibody that scales at the same speed as the threat. And unlike regulation or detection systems, awareness doesn’t need to be imposed. It can be built, taught, shared, and spread from person to person.
Before sharing a piece of content, check where it comes from. Before reacting to a video or a statement, stop. Ask yourself whether the source has a name, a history, something to lose. Treat every piece of content as potentially synthetic until a credible, accountable source confirms it. These are habits, not technologies. They cost nothing and they work immediately.
Finally, we need the help and collaboration of the tech community. Those who design platforms, write code, and make decisions about how feeds and ranking algorithms work are making choices that directly shape the information ecosystem. These are choices with democratic consequences. The people making them know it. Many have known it for years. This is the moment to stop treating it as someone else’s problem and to decide which side you’re on. Because the swarms are not waiting.
We can do this. The tools exist, the knowledge is there, and the threat is clear enough that pretending not to see it is already a choice. The question is whether we act now, while the window is still open, or later, when the damage will be harder to reverse.
