Paper Review: Can We Fix Social Media?
Recently I came across the paper Can We Fix Social Media? Testing Prosocial Interventions using Generative Social Simulation. Besides being a great review of negative effects stemming from social media, authors propose and discuss focused interventions on core content-diffusion mechanisms of social media and evaluate the effect on a simulated environment (not real users, but simple LLM powered agents).Â
Before digging into the paper, it is worth sharing some context. The growth and information diffusion in social are governed by few principles:
The 1% rule:Â Â about 1% of Internet users create content, while 99% are just consumers of that content. This rule is slightly different on social media, where the barriers to content creation are lower. The percentage of users creating content is closer to the range 7-10%.Â
Richer get richer: most of the network activity (follows) is towards a fraction of popular users. The more popular a user gets, the more followers they will have the next day.
Homophily:Â humans have the tendency to form connections with similar-minded individualsÂ
Sensationalism drives eyeballs and engagement, best summarized by Zuckerberg as âwhen left unchecked, people will engage disproportionately with more sensationalist and provocative contentâ (see https://www.facebook.com/notes/751449002072082/)
The paper starts reviewing negative effects of social media to society, which stems from the principles above:
Fragmentation and echo chambers: users tend to fragment themself into homogeneous âecho chambersâ, with low content diversity and high ideologically homogeneity.Â
Social media prism and amplification of conflict. As algorithms are optimized to maximize user engagement, they amplify the effect and prominence of outrage, conflict, and sensationalism. This creates a distorted vision of reality.
Inequality of influence: a small number of highly visible users own the vast majority of attention and influence.Â
What would happen if we were to change these mechanisms ? Authors propose 6 interventions, which they test on a simulated environment made of LLM agents.Â
Each LLM agent is configured through age, gender, income,education, partisanship, ideology, religion, and personal interests (sampling profiles from the American National Election Studies dataset). At each step, a randomly selected user may write a new post in response to a news item, repost existing content, or follow another user. Timelines consist of ten posts: five from followed users and five drawn from high-engagement content posted by non-followed users.
At the end of the simulation, authors report on:
degree of homophily: profile similarity between users who are connected on the network
correlations between political extremity and engagement
inequality in followers and reposting activity.
Intervention and results:
Chronological ordering of content on feeds (removing engagement-based ranking). This broke the link between post visibility and popularity. The concentration of followers dropped sharply, and the concentration of reposts also declined. It did not reduce ideological homophily but intensified the correlation between political extremism and influence.Â
Downplaying dominant voices (prioritizing posts with fewer reposts). It lowered maximum follower and repost counts and reduced inequality , but had no measurable effect on partisan amplification or homophily.
Boosting out-partisan content:Â little impact across any outcome dimension.
Bridging attributes (promoting high-quality, constructive content). Reduced the link between partisanship and engagement and increased cross-partisan connections. However, it also increased inequality: visibility became concentrated among a narrow set of high-scoring posts.
Hiding social statistics and hiding biographies. It had minimal effect on the structural dynamics of the network, but lead to more followers, possibly reducing status-based filtering.