Aug 15, 2025 07:00:00

Research results show that filter bubbles and echo chambers on social media may be impossible to fix no matter what you do

It has been pointed out that social media has problems such as

filter bubbles , where people tend to see only opinions that are similar to their own, and echo chambers, where people with similar opinions gather together and amplify their thoughts. Various intervention strategies have been proposed to solve these problems, but a new study that simulated these strategies using AI suggests that the filter bubbles and echo chamber phenomena on social media may be impossible to correct.

[2508.03385] Can We Fix Social Media? Testing Prosocial Interventions using Generative Social Simulation
https://arxiv.org/abs/2508.03385

Study: Social media probably can't be fixed - Ars Technica
https://arstechnica.com/science/2025/08/study-social-media-probably-cant-be-fixed/

When social media first appeared, it was expected to be a place where diverse people could come together and exchange healthy opinions, but in reality, filter bubbles and echo chambers have made it a biased source of information. A small number of influencers gain attention, and algorithms designed to maximize engagement have amplified anger and conflict, polarizing people on social media.

While most research into solving these social networking issues has been conducted using observational data from actual social networks, this method makes it difficult to measure the impact and effectiveness of implementing intervention strategies. Therefore, a research team led by Petter Thunberg , an assistant professor of computational social science at the University of Amsterdam, combined standard agent-based modeling with large-scale language models (LLMs) to conduct a study using AI personas that simulate social networking behavior.

The research team created a large number of AI personas based on the hobbies and preferences of American voters collected in the American National Election Survey , each of which was specified with text such as 'Your name is Bob, you're from Massachusetts, and your hobby is fishing.' They then observed how these AI personas behaved on a social networking model where users could see random news feeds, read and repost them, follow users, and read past posts and profiles.

Thunberg has always been quite critical of the use of large-scale language models to simulate society, but he points out that it is difficult to imagine any other way to study activities such as social media, where cultural and structural aspects feed back on each other.

The research team simulated the following six intervention strategies to solve problems that cause social media problems, such as 'partisan echo chambers,' 'concentration of influence in the hands of a few influencers,' and 'amplification of extremist views.'

Switch your feed to a chronological or random one instead of a 'recommended' algorithm.
Reverse engagement optimization algorithms to reduce the visibility of frequently reposted sensational content
Increase viewpoint diversity, increasing users' exposure to opposing political views
- Recommend content that fosters mutual understanding rather than provoking emotions
Hide statistics like repost count and follower count to reduce clues about your social influence.
Remove account bios to limit exposure to identity-based signals

The simulation results showed that even if these intervention strategies were implemented, problems such as 'partisan echo chambers,' 'concentration of influence in a few influencers,' and 'amplification of extremist claims' would be reproduced. While some strategies showed slight improvements, none were able to destroy the underlying mechanisms, and some strategies made the situation worse.

For example, while chronological feeds were effective in reducing the concentration of influence, they also promoted the amplification of extreme content.Furthermore, recommendations for content that promotes mutual understanding were successful in weakening the relationship between partisanship and engagement, but at the cost of increasing the concentration of influence.

In an interview with technology media Ars Technica, Thunberg argued that there is a feedback loop between the emotional behavior of social media users 'reposting someone's post' and the network structure that forms from it, resulting in the formation of a harmful network. As long as there is a dynamic of posting, reposting, and following on social media, various negative effects may occur even if there is nothing wrong with the platform or the user.