The team of researchers has given AI bots their own social platform.
The chatbot split into creeks and boosted most partisan voices. A handful of “influencers” also quickly dominated the conversation, according to a survey released last Tuesday by researchers at the University of Amsterdam.
Researchers have built minimal social networks without advertising, recommended posts, and algorithms to determine user presentation. I then entered 500 chatbots powered by Openai's GPT-4O Mini, and assigned each a distinct persona with a specific political tendency.
Personas were drawn from the US National Election Research dataset and reflect “the real-world distribution of age, gender, income, education, partisanship, ideology, religion and personal interests,” the researchers said.
They added that the experiment modeled and reproduced users on Llama-3.2-8B and DeepSeek-R1, resulting in “the same qualitative pattern.”
The study was led by Dr. Petter Telberg, an assistant professor in computational social sciences at the University of Amsterdam, and Mike Roujge, a research engineer at the university.
Researchers Openai, Meta, and Deepseek did not respond to requests for comment from Business Insider.
The same toxicity patterns have appeared even without algorithms or humans
Over the course of five separate experiments, each performing over 10,000 actions, the bots were free to post, follow and repost. What happened looked like real-world social media.
This study found that chatbots shared political beliefs and were drawn to others who formed harsh echo chambers. Partisan voices gained a high-profile share, with the most extreme posts attracting the most followers and reposts. Over time, small groups of bots began to dominate the conversation, like the influencer-rich dynamics seen on platforms like X and Instagram.
The researchers also tested six interventions aimed at breaking polarization loops, including time series feed, downranking viral content, hiding follower counts, hidden user BIOS, and amplifying conflicting views.
The problem did not resolve. “Some of them had moderately positive effects, but none completely addressed the pathology of the core, and improvements in one dimension were often brought about by worsening another dimension,” the researchers said.
“Our findings challenge the general view that social media dysfunction is primarily the result of algorithmic curation,” the author writes.
“Instead, these issues may be rooted in the architecture of social media platforms: networks that grow through emotionally reactive sharing,” they added.
Related Stories
Business Insider tells the innovative stories you want to know
Business Insider tells the innovative stories you want to know
Researchers said their work was first to use AI to advance social science theory.
LLM-based agents can provide “a rich representation of human behavior” to study social dynamics, but researchers have warned that they remain “black boxes” and carry “a risk of embedded bias.”
Not the first AI social network experiment
This study is not the first time researchers have tested what happens when AI bots live in online spaces.
In 2023, Business Insider also reported on an experiment led by Törnberg. In this experiment, 500 chatbots read the news and discussed it on simulated social media platforms.
This project used ChatGPT-3.5 to build a bot for a very specific purpose. This is to explore ways to design a less polar and less toxic version of current social networks. Researchers created a social network model in the lab to test whether it could promote partisan interactions without promoting hostility.
“Is there a way to promote interaction beyond partisan disparities? Törnberg asked at the time.
In both studies, chatbots acted as stand-in for people, and researchers tracked interactions to better understand how users behave online.
Big Tech is also testing a similar approach.
In July 2020, Facebook introduced walled simulations where millions of AI bots live to study online toxicity.
