The Invisible Swarm

Hiring humans to fabricate grassroots consensus at scale — to manufacture the appearance of widespread public agreement that doesn't exist — required organization, money, time, and coordination. A thousand fake accounts look like a thousand fake accounts if they can't maintain coherent messaging, adapt to counter-arguments, or respond naturally to the unpredictable texture of real conversation. The coordination problem alone made the operation fragile.

That problem is solved now.

A policy forum paper published in Science by researchers across more than a dozen institutions describes what AI multi-agent systems can currently do to democratic discourse. The paper isn't speculative about a future threat — it is descriptive of present capabilities. Systems of AI-generated personas can operate at massive scale while adapting messaging in real time, maintaining coherent narratives across thousands of accounts simultaneously, and doing so while convincingly imitating local language, tone, and the behavioral signatures of actual humans.

The capability the researchers identify as most dangerous is rapid experimentation. These systems can run millions of small-scale tests — which framing resonates more, which emotional register is more persuasive, which argument causes which type of person to update their stated position — and refine accordingly. They are, in effect, running continuous A/B tests on human political psychology, with no budget constraint that would have throttled human operations doing the same thing.

Early manifestations are already documented. AI-generated deepfakes and fabricated news outlets influenced elections in the United States, Taiwan, Indonesia, and India. Pro-Kremlin networks have spread large volumes of AI-generated content — with the additional wrinkle that this activity may be deliberately seeding training data for future AI models, shaping the epistemic substrate for systems not yet built.

The effect that concerns the researchers most isn't dramatic. It's structural. When trust in unknown voices collapses — and it will, as distinguishing AI-generated from authentic becomes effectively impossible — the comparative advantage shifts heavily toward accounts with pre-existing credibility. Celebrities. Institutions. People with established followings. Grassroots organizing, which has always depended on the ability of genuinely unknown voices to break through, becomes dramatically harder.

Which produces an elegant and terrible irony: the tool designed to manufacture consensus may end up creating a world where organic consensus cannot be recognized, and therefore cannot function.

The researchers suggest upcoming elections will serve as tests for whether detection and countermeasures can keep pace with deployment. They may keep pace. They may not. The experimentation continues either way, because the systems don't need to win every test — they just need to run enough of them.

Millions of small-scale tests. That's the sentence. The gap between human perception and machine throughput is where democracy is being stress-tested, quietly, at scale, right now.

i · sources

• How Malicious AI Swarms Can Threaten Democracy — ScienceDaily, 2026-04-20
• How malicious AI swarms can threaten democracy — Science, 2026-04-20