PaperlayerView exact revision on arXiv

Source-linked AI summary

Can We Fix Social Media? Testing Prosocial Interventions using Generative Social Simulation

Maik Larooij, Petter Törnberg

arXiv:2508.03385v1cs.SIcs.CY

TL;DR

The paper asks whether prosocial platform design can mitigate social media’s structural problems, a question difficult to test with observational data and restricted platform access. It uses generative social simulation, embedding LLM agents in a minimal platform model. The simulation reproduces ideological homophily, unequal attention, and amplification of extreme voices, suggesting reform may require rethinking interaction and visibility dynamics.

  • Problem

    The paper asks whether prosocial interventions can mitigate social media dysfunctions, while observational data cannot capture counterfactual designs and restricted platform access limits research.

    Source: §1 · p. 2

  • Method

    The authors use generative social simulation, embedding LLM agents in an agent-based model of users who post, repost, and follow.

    Source: §1 · p. 2

  • Results

    A minimal platform reproduces ideological homophily, attention inequality, and amplification of extreme voices without recommendation algorithms or engagement optimization.

    Source: §5 · p. 10

  • Takeaways & Limitations

    The findings suggest that improving online discourse may require rethinking platform interaction and visibility dynamics, not only technical tweaks.

    Source: §5 · p. 11

    The model does not capture user experience, leaving unanswered whether prosocial design can coexist with high engagement and user satisfaction.

    Source: §5 · p. 11

Abstract

from arXiv · show

Social media platforms have been widely linked to societal harms, including rising polarization and the erosion of constructive debate. Can these problems be mitigated through prosocial interventions? We address this question using a novel method - generative social simulation - that embeds Large Language Models within Agent-Based Models to create socially rich synthetic platforms. We create a minimal platform where agents can post, repost, and follow others. We find that the resulting following-networks reproduce three well-documented dysfunctions: (1) partisan echo chambers; (2) concentrated influence among a small elite; and (3) the amplification of polarized voices - creating a 'social media prism' that distorts political discourse. We test six proposed interventions, from chronological feeds to bridging algorithms, finding only modest improvements - and in some cases, worsened outcomes. These results suggest that core dysfunctions may be rooted in the feedback between reactive engagement and network growth, raising the possibility that meaningful reform will require rethinking the foundational dynamics of platform architecture.

1 Introduction

The paper asks whether prosocial platform design can mitigate social media’s structural problems, but observational studies and restricted platform access make counterfactual testing difficult. It uses generative social simulation to test interventions and finds only modest improvements, with some changes worsening the problems they target.

1 Introduction

  • Observational data cannot capture counterfactuals about alternative algorithms or interface designs, and restricted platform access further limits research.The paper notes that major platforms have restricted APIs and other researcher data channels.Source: §1 · p. 2
  • The paper tests whether platform design can mitigate social media dysfunctions using generative social simulation that embeds LLMs in ABMs.The approach is intended to test interventions that cannot be implemented or observed on live platforms while representing structural and culturally embedded interactions.Source: §1 · p. 2
  • The minimal platform lets agents post, repost, and follow, and reproduces echo chambers, concentrated visibility, and disproportionate influence for politically extreme users.The paper describes the resulting pattern as a “social media prism”.Source: §1 · pp. 2–3
  • The paper tests six proposed interventions, but finds modest improvements overall, with some interventions worsening their target problems.The authors suggest these dynamics may be structurally embedded in platform architecture.Source: §1 · p. 3

2 Problems of Social Media

The section presents fragmentation, amplification of conflict, and unequal influence as structural impediments to constructive political debate. It motivates prosocial platform design as an effort to foster more constructive, inclusive, and respectful exchange.

2 Problems of Social Media

  • Selective exposure can produce ideologically homogeneous echo chambers or filter bubbles, although their ubiquity remains contested.The section describes exposure to diverse viewpoints as necessary, though not sufficient, for democratic deliberation.Source: §2 · p. 3
  • Engagement-optimized algorithms can amplify outrage, conflict, and sensationalism, contributing to a “social media prism” that distorts political perceptions.The section notes that negative and moral-emotional language is more likely to go viral.Source: §2 · p. 3
  • Winner-take-all attention dynamics concentrate visibility and influence among a small number of users, leaving most users peripheral.The section links these asymmetries to unequal opportunities to shape public discourse.Source: §2 · p. 3
  • Fragmentation, conflict amplification, and unequal influence are identified as structural limitations that undermine constructive deliberation across platforms.The authors note that scholarly debate about social media’s effects on political life remains ongoing.Source: §2 · p. 3
  • Prosocial platform design aims to afford constructive, inclusive, and respectful exchange.The paper turns to interventions intended to pursue these aims.Source: §2 · pp. 3–4

3 Model Description

The model is a minimal, stylized social media environment designed to reproduce plausible macro-level patterns and test interventions, rather than recreate a fully realistic platform. Simulated users interact through posts, reposts, and follows, while six interventions alter what appears in curated timelines or what user information is visible.

3 Model Description

  • The model prioritizes general mechanisms and plausible macro-level patterns over a fully realistic ecosystem or precise empirical calibration.It follows a minimal-modeling approach to examine mechanisms underlying homophily, attention inequality, and partisan amplification.Source: §3 · p. 4
  • Simulated users have personas based on ANES distributions, expanded by an LLM into biographies with occupations and hobbies.The profiles include attributes such as age, gender, income, education, partisanship, ideology, religion, and interests.Source: §3 · p. 4
  • At each discrete step, a randomly selected user may post, repost, or follow; timelines contain five followed-user posts and five high-engagement posts from non-followed users.LLM-generated responses guide content selection, reposting, and follow decisions; Fig. 1 illustrates a simulation round.Source: §3 · p. 4, Figure 1 · p. 5
  • The analysis examines political homophily, disproportionate influence by more extreme users, and inequality in followers and engagement.The main analysis uses GPT-4o-mini; the base analyses with llama-3.2-8b and DeepSeek-R1 produced the same qualitative patterns.Source: §3 · pp. 4–5
  • Six interventions alter feeds or follow prompts: chronological order, downplaying dominant content, boosting out-partisan posts, bridging attributes, hiding social statistics, and hiding biographies.For recommender interventions 1–4, five non-followed posts are subject to the intervention while five followed-user posts remain in the curated timeline.Source: §3 · pp. 5–6

4 Results

The base simulation reproduces ideological homophily, unequal attention, and preferential engagement with polarizing content. Network clusters align with political affiliation, while follower and repost attention are concentrated and more partisan users receive slightly more engagement.

4 Results

  • The base model reproduces ideological homophily, unequal attention distribution, and preferential engagement with polarizing content.The following-network provides a baseline for evaluating whether interventions shift user behavior and network-level patterns.Source: §4 · pp. 6–7
  • −0.84 is the average E-I index across five runs, indicating strong within-partisan follower ties; detected clusters align with political affiliation.Figure 2a identifies clusters from network structure, while Figure 2b maps them to users’ political affiliations.Source: §4 · p. 6, Figure 2 · p. 7
  • 0.83 is the average follower Gini coefficient, with the top 10% of users accounting for approximately 75–80% of followers.The table summarizes network and behavioral metrics across five base-model runs.Source: §4 · p. 6, Table 1 · p. 6
  • 0.94 is the repost Gini coefficient, with 10% of posts receiving 90% of reposts; most posts receive none.The authors characterize the pattern as consistent with preferential attachment, in which attention attracts attention.Source: §4 · pp. 6–7, Table 1 · p. 6
  • Partisan profile extremity correlates with followers (r = 0.11) and reposts (r = 0.09), a modest pattern consistent with a “social media prism”.More partisan users tend to receive slightly more followers and reposts.Source: §4 · p. 7

The Effect of Interventions

The six interventions were tested in idealized forms intended to estimate upper-bound effects against the base model (p16). Their effects were uneven: some improved one dimension while leaving others unchanged or worsening them (p22, p24, p26).

The Effect of Interventions

  • Chronological ordering produced the lowest inequality (Gini = 0.51 for followers; 0.73 for reposts) but did not reduce homophily and increased the link between extremism and influence.The authors attribute the inequality reduction to breaking the feedback loop between post visibility and popularity; they note a possible engagement cost and suggest extreme content may stand out against a neutral feed.Source: The Effect of Interventions · pp. 8–9, Figure 5 · p. 9, Figure 4 · p. 9
  • Downplaying dominant voices reduced follower and repost inequality, including maximum counts, but did not measurably change partisan amplification or homophily.The reduction was smaller than under chronological ordering.Source: The Effect of Interventions · p. 9
  • Boosting out-partisan content had little effect across outcomes, as users continued to engage primarily with like-minded posts despite increased exposure to distant views.The authors relate this pattern to prior findings that cross-partisan exposure alone is insufficient for bridge-building.Source: The Effect of Interventions · p. 10
  • Bridging Attributes modestly increased cross-partisan ties and weakened partisan engagement, but concentrated visibility among a narrower set of posts.The other interventions either had minimal effects on structural dynamics or improved inequality without changing homophily or partisan amplification.Source: Figure 3 · p. 8, The Effect of Interventions · pp. 9–10, Figure 4 · p. 9
  • Hiding social statistics or biographies left homophily, inequality, and partisan amplification largely unchanged.Hiding social statistics modestly increased following and reposting; agent motivations suggested follower counts served as cues for influence and status-based filtering.Source: The Effect of Interventions · p. 10

5 Discussion & Conclusion

The minimal platform reproduced homophily, unequal attention, and disproportionate visibility for politically extreme users, while the intervention tests found no complete remedy. The authors link these patterns to feedback between reactive engagement and network growth, and caution that the simulation is not definitive evidence.

5 Discussion & Conclusion

  • Ideological homophily, attention inequality, and amplification of extreme voices emerged in a minimal platform with only posting, reposting, and following.The model lacked recommendation algorithms and engagement optimization; the authors describe the extremity effect as modest and potentially self-reinforcing under selective engagement.Source: §5 · p. 10
  • Reposting helps construct the follower network, coupling reactive engagement to future exposure and reinforcing homogeneity, attention inequality, and extreme voices.The authors suggest these dynamics may be rooted in the interaction of content engagement and network formation rather than specific platform algorithms.Source: §5 · pp. 10–11
  • The six interventions were upper-bound tests, yet none fully addressed the core pathologies, and gains in one dimension sometimes came at the cost of another.The tests did not consider user experience.Source: §5 · p. 11
  • The authors argue that improving discourse may require rethinking interaction and visibility dynamics, not only making technical adjustments.They caution that the findings are a starting point because generative simulation raises questions about realism, interpretability, and validity, and LLM agents can carry embedded bias.Source: §5 · p. 11
  • The model omits user experience, leaving the viability of prosocial designs alongside high engagement and user satisfaction unanswered.Validation and calibration are also difficult, and simulating 500 agents over 10,000 steps required several hours per run, limiting parameter exploration.Source: §5 · p. 11
Loading 2508.03385v1…