Bonsai: Intentional and Personalized Social Media Feeds

Paper Title

Bonsai: Intentional and Personalized Social Media Feeds

Publication Info

• Topic area: Personalized social media feed creation using natural language input and transparency mechanisms.
• Keywords: Intentional feeds, personalized feeds, social media, language models, transparency, user agency, feed customization, Bluesky, algorithmic control, natural language.

Background and Problem

• Problem / challenge: Engagement-optimized feeds dominate social media platforms, often misaligning user behavior with their deeper goals and promoting compulsive or addictive use. Existing tools for feed customization are either blunt (e.g., time management tools) or require substantial technical expertise.
• Significance: Misaligned feeds can harm user well-being, erode trust, and lead to platform abandonment. Creating feeds that align with explicit user intentions could improve mental health, productivity, and satisfaction.
• Motivation and related work: Prior work has explored friction tools, topic-based interfaces, and rule-based customization systems, but these approaches are limited in granularity and accessibility. Language models offer a promising avenue for enabling intent-driven personalization without requiring technical expertise.

Solution

• Proposed approach: Bonsai—a platform-agnostic system for creating intentional and personalized social media feeds using natural language input. It includes Planning, Sourcing, Curating, and Ranking modules to transparently generate feeds aligned with user-stated goals.
• Novelty:
  1. First end-user-facing system using language models to enable natural language-based feed creation.
  2. Platform-agnostic, modular architecture for feed customization.
  3. Empirical insights from a multi-week field study with Bluesky users.
• Procedure and key techniques:
  • Planning: Translate user’s natural language input into feed configurations using an LLM agent.
  • Sourcing: Retrieve posts from specified feeds, accounts, hashtags, and queries via APIs.
  • Curating: Score posts based on user preferences using a multimodal LLM.
  • Ranking: Order posts using a Weighted Borda Count algorithm based on relevance, recency, and popularity.

Results

• Concrete findings:
  • 70% of posts in Bonsai-generated feeds were rated as relevant or better (26% "perfectly relevant," 44% "relevant").
  • 16% of posts were rated "not relevant," and 7% "should have been excluded."
  • Substantial inter-annotator agreement (κ = 0.77).
• Advantage over baselines: Compared to existing feed builders (e.g., Skyfeed), Bonsai reduced setup effort and allowed nuanced preference articulation via natural language.
• Experiments / evaluation:
  • Field study with 15 Bluesky users over a median of 11 days.
  • Participants created 36 feeds, manually regenerated feeds 103 times, and consumed Bonsai feeds 1,413 times.
  • Interviews and interaction logs analyzed for themes like user agency, effort-reward tradeoffs, and transparency.
• Limitations and future work:
  • Limited generalizability due to Bluesky’s smaller ecosystem and participant skew toward academics.
  • Short study duration (median 11 days) may not capture long-term adoption patterns.
  • System constraints (e.g., lack of network-based or location-based suggestions) caused friction.
  • Risks of reinforcing echo chambers and filter bubbles need further exploration.

Summary

Bonsai introduces a novel system for creating intentional and personalized social media feeds using natural language input, enabling users to articulate their goals and transparently shape feed algorithms. A field study demonstrated that Bonsai helps users discover new content, filter out irrelevant or harmful posts, and decouple engagement from intent, though feedbuilding required more effort than users were accustomed to. Participants valued the sense of agency and transparency but highlighted the need for tighter feedback loops and hybrid approaches combining intentional and emergent feeds. Bonsai offers a promising path toward social media systems that better align with user goals and values.