Researchers at the University of Amsterdam created a social media platform with 500 AI chatbots. No humans, no ads, no algorithms curating feeds—just bots, powered by OpenAI’s GPT-4o mini, in a bare-bones digital space.
They wanted to see how these digital agents, each with a political leaning, would interact. The results, in a preprint on arXiv, are scary: even without algorithmic nudging, the bots created their own echo chambers, amplifying division and rewarding extreme voices. This is what an AI-driven social media future might look like.
Plaud Note AI Voice Recorder, Voice Recorder w/Case, App Control, Transcribe & Summarize with AI...
- Plaud Note AI TECHNOLOGY: Plaud APP integrates advanced transcription capabilities, enabling swift and accurate transcription in 112 languages. AI...
- YOUR PRIVACY COMES FIRST: Local data encrypted, cloud files exclusive to you, with data processing only upon your authorization. Plaud App and Web...
- DUAL-MODE PRECISION RECORDING: Plaud Note features a dual recording engine, capturing high-quality ambient sounds with the air conduction sensor for...
Each bot was given a persona, with a political affiliation—liberal, conservative or somewhere in between. The platform was stripped down, with no recommended posts or trending topics, just a space for the bots to post, follow and repost. Over five experiments, the researchers tracked 10,000 actions, from follows to posts to reposts. The bots quickly grouped themselves into clusters, gravitating towards others with similar views. Those posting the most polarizing content—sharp, partisan jabs—gained the most followers and saw their posts spread the widest. The pattern held across every trial, so it seems to be hardwired into the bots.
What’s interesting about this is how it mirrors human behavior, but removes the usual scapegoats. Polarization is blamed on social media algorithms, yet no algorithm was in control in this instance. The bots, which were trained on massive datasets of human online interactions, simply performed what they were told: they sought out like-minded views while amplifying the loudest, most controversial ones. So the question is: are we, or our digital reflections, doomed to fracture along ideological lines no matter what?
The researchers didn’t stop there, as they tested fixes to see if they could nudge the bots towards less polarized interactions. One trial introduced a chronological feed, hoping a neutral presentation of posts would break up partisan cliques. Another hid follower counts and repost numbers, to curb the race for viral fame. They even tried amplifying opposing viewpoints, a tactic that had worked in a previous study to boost engagement while keeping toxicity low. But none of these tweaks made a significant impact—engagement with partisan accounts shifted by 6% at most. In one example, obscuring user profiles backfired, exacerbating the split as bots leaned deeper into extreme posts seeking attention.
Even at its most basic level, social media appears to reinforce our worst instincts. The bots were not instructed to quarrel or polarize; rather, they just followed the trends in their training data, which is based on years of online human activity. If AI agents, free from corporate algorithms, still end up in digital trenches, it means the problem isn’t the tech—it’s us. Our habits, our biases and our need for validation shape platforms as much as any code.
[Source]
