Bluesky: Decentralized Social Network

• Bluesky Social Network is a decentralized microblogging platform built on the AT Protocol, featuring modular architecture, open data access, and user-driven curation.
• Its innovative data model uses append-only Merkle search trees and a centralized relay to efficiently aggregate and curate diverse content feeds.
• Rapid user migration, fine-grained moderation tools, and academic engagement make Bluesky a key testbed for decentralized network and social dynamics research.

Bluesky is a decentralized, microblogging social network founded on the AT Protocol. Initiated in 2023, it has emerged as a key actor in the current generation of federated social platforms, featuring open data access, modular architecture, and user-driven content curation. Adoption has been accelerated both by technical innovations—such as user-customizable feeds and fine-grained moderation tools—and by mass migration events from legacy platforms, notably Twitter/X. Bluesky serves as a valuable case study for decentralized network engineering, platform-driven community formation, and the shifting modalities of online discourse and polarization.

1. Protocol Architecture, Decentralization, and Data Model

Bluesky's backbone is the AT Protocol, which disaggregates core social networking functions into interoperable, replaceable modules. Every user has a Decentralized Identifier (DID)—most commonly issued by a central directory (did:plc) but supporting self-hosted (did:web) roots. Each DID resolves to a document containing the user’s handle, PDS (Personal Data Server) endpoint, and public keys for signature validation (Kleppmann et al., 2024, Balduf et al., 2024).

User repositories are append-only Merkle Search Trees, where each action (post, follow, like, block) is a signed record addressable by globally unique URIs (at://<did>/<collection>/<rkey>), supporting verifiable proofs of inclusion and deletion. A PDS hosts at least one repository and exposes APIs for data reads, writes, and subscription to updates.

Aggregation across the distributed corpus is accomplished via a centralized Relay, which re-broadcasts events from all known PDSes as a “Firehose” stream (with a typical three-day retention). Higher-level “AppView” instances ingest the Firehose and corresponding label streams, constructing materialized indices for efficient timeline and search queries. Moderation and recommendation are handled by “Labelers” and “Feed Generators” respectively, each of which can be run by any party and selected by users according to their preferences (Kleppmann et al., 2024, Balduf et al., 2024).

Open APIs and explicit data access policies underpin Bluesky’s ecosystem, enabling fine-grained, temporally precise (millisecond) social-graph and content datasets such as BlueTempNet (Jeong et al., 2024) and the pairwise/higher-order hypergraph “A Blue Start” (Smith et al., 16 May 2025).

2. Network Growth, Activity Patterns, and User Migration

Bluesky’s user base expanded through distinct migration waves: initial invitation-only seeding, public access opening in February 2024, exogenous shocks (X/Twitter policy changes, geopolitical events), and academic migration in late 2024 (Seckin et al., 17 Apr 2025). As of January 2025, there were approximately 30 million total registrations, with a stable active core of 26 million (≥1 follow-edge), and daily activity stabilizing at 4–5 million DAU (≈15% of accounts) (Seckin et al., 17 Apr 2025). Daily mean activity per active user exceeds ten actions (posts, likes, reposts).

The undirected follower network exhibits high clustering (orders of magnitude above random), a giant connected component (70–80% of users), and heavy-tailed degree distributions. Heterogeneity metrics such as Gini and Kappa indices show a pronounced rise after major migration surges, indicating the rapid emergence of high-degree “hub” accounts (Seckin et al., 17 Apr 2025, Failla et al., 2024, Smith et al., 16 May 2025).

Waves of growth exhibit language and regional signatures: the public opening saw an influx of Japanese users, a subsequent wave reflected Brazil’s X ban (influx of >500,000 Portuguese speakers), and late 2024 shifts reflect U.S. election dynamics and English-dominant user growth (Seckin et al., 17 Apr 2025, Balduf et al., 2024).

3. Structural Features and Higher-Order Network Properties

The interaction landscape is multilayered: follows, blocks, feed-creation, feed-membership (“liking” a feed), replies, reposts, and quotes. All core layers are available with fine-grained, time-resolved datasets (Jeong et al., 2024, Smith et al., 16 May 2025, Failla et al., 2024).

The follower network (n≈26–27M nodes, m>1.6B edges) and starter-pack hypergraph (≈300,000 hyperedges) capture classical and group-mediated dynamics. Degree and hyperdegree distributions are heavy-tailed, with hubs in both users and starter-pack participation (Smith et al., 16 May 2025). Starter packs—bundles of suggested accounts/feeds, which facilitate “one-click” social bootstrapping—introduce higher-order structure that shapes group formation, information diffusion, and network percolation (Balduf et al., 20 Jan 2025, Smith et al., 16 May 2025). Most starter packs cross community boundaries, as quantified by high normalized entropy in 2-section projections.

The clustering coefficient, path lengths, and small-world properties of both pairwise and higher-order projections closely mirror those found in established platforms, supporting rapid diffusion (Seckin et al., 17 Apr 2025, Quelle et al., 2024). Connected components and k-core analysis reveal a resilient core and the presence of high-coreness users in both pairwise and higher-order views.

4. Content Dynamics: Originality, Curation, and Custom Feeds

Content generation is dominated by original posts: in the post-public-opening window, original posts account for ≈62% of content, replies ≈28%, and reposts ≈10%—inverting the proportions typical of Twitter/X, where retweets often exceed originals (Nogara et al., 5 May 2025, Sahneh et al., 2024). Activity distributions for both posts and follows are heavy-tailed (power-law with scaling exponent α≈2.5 for posts) (Sahneh et al., 2024, Failla et al., 2024).

Algorithmic curation operates via “Feed Generators”—API-exposed services producing ranked streams of eligible post URIs. Over 40,000 feed generators are registered, but a handful of platforms (e.g. Skyfeed) account for the vast majority of user engagement (Balduf et al., 2024). The default home timeline is inverse-chronological; a small minority of users employ advanced custom-feed logic, despite the ready availability of domain- or topic-specific generators (e.g. Paper Skygest for research content, Astrosky for astronomy feeds, BlackSky for minority-centered content) (Greenwood et al., 6 Jan 2026, Hunt et al., 23 Jan 2026, Failla et al., 2024).

Starter packs are a major driver of network densification: over 20% of follow edges have been seeded via pack “follow-all” actions, resulting in rapid personal network expansion and substantial, persistent lifts in both visibility (up to +85% followers for members, +117% for pack creators in four weeks) and activity (+60% post rate) (Balduf et al., 20 Jan 2025). However, these mechanisms can reinforce echo chambers, over-curation, and rich-get-richer effects, as measured by Gini coefficients and community-concentration metrics (Balduf et al., 20 Jan 2025).

5. Polarization, Political Discourse, and News Reliability

Bluesky’s political content is marked by high topical engagement (≈13% of posts), with issue prevalence reflecting real-world event salience. Political discussions concentrate on U.S. politics (Trump, Musk), international conflicts (Russia–Ukraine, Israel–Palestine), and socio-technological debates (LGBTQ+ rights, AI) (Salloum et al., 3 Jun 2025).

Polarization is characterized by high “structural polarization” (Adaptive EI index) on salient topics (AEI up to 0.95 in Israel–Palestine), but with extreme group-size imbalance: minority (“B-side”) stances rarely constitute more than 1–2% of participants in the most polarized topics. Content-based polarization (measured via Kullback-Leibler divergence over key vocabulary) is highest on international conflicts and culture-war topics. Simpson diversity indices are very low (0.01–0.02) in these topics, confirming dominance by a single stance (Salloum et al., 3 Jun 2025). Bluesky thus demonstrates “homogeneous cluster polarization”—a reconfiguration where cross-platform sorting leads to ideologically uniform communities, rather than internal “echo chambers” on a single site (Salloum et al., 3 Jun 2025, Quelle et al., 2024).

News sharing exhibits a pronounced left-of-center skew: over 90% of news-link posts reference high-credibility domains (NewsGuard ≥60), with >80% from left or lean-left outlets. Unreliable-source news links are rare (~2% of all news-linking posts) and skew more extremely (also left), with a flatter frequency-rank profile indicating dispersion rather than dominance by a single outlet (Reddy et al., 17 Jan 2025). Misinformation and toxicity indicators remain consistently low in both content and user distributions across pre- and post-public-opening phases (<1% toxic posts, <0.13% of users share low-credibility links) (Nogara et al., 5 May 2025, Sahneh et al., 2024).

6. Scientific, Societal, and Research Applications

Bluesky is increasingly used for altmetric and science communication purposes. Recent analyses covering over 87,000 research-linked posts (Feb 2024–Apr 2025) show a sharp inflection in scholarly engagement from November 2024, coinciding with prominent academic migration (Zheng et al., 24 Jul 2025). Disciplinary coverage centers on social, environmental, medical, and physical sciences. Posts exhibit a high degree of textual originality (median cosine similarity with article title <0.3), contrasting with the more surface-level sharing common on Twitter/X. Engagement rates (likes/reposts) per paper-linked post on Bluesky are at least equivalent to, and often higher than, benchmarks from X, especially for replies and quotes (Zheng et al., 24 Jul 2025).

Custom academic feeds (e.g. Paper Skygest) leverage Bluesky’s open-feed infrastructure to provide personalized, research-focused timelines; these drive measurable increases in exposure and engagement with academic content (+3.5pp in paper-like share) (Greenwood et al., 6 Jan 2026). Domain-specific networks, such as Astrosky for astronomy, use the AT Protocol’s modularity, offering independent, resilient scientific community infrastructure (Hunt et al., 23 Jan 2026).

Bluesky’s open APIs, public Firehose, and modular data model have enabled the construction of advanced, real-time sentiment/narrative pipelines (e.g., CognitiveSky), “ground-truth” experimental datasets for algorithmic curation and exposure, and high-resolution studies of network evolution, polarization, and higher-order social contagion (Chhetri et al., 14 Sep 2025, Jeong et al., 2024, Smith et al., 16 May 2025, Failla et al., 2024).

7. Limitations, Open Questions, and Future Prospects

Despite the modular protocol, key system choke points remain: handle assignment and PDS hosting are almost entirely centralized (bsky.social >98% of handles, >99% repo hosting), the Relay and AppView layers form de facto single points of aggregation, and main feed and label-generators are dominated by very few providers (e.g., Skyfeed controls >85% of all custom feeds) (Balduf et al., 2024).

Current engagement, while robust, is dominated by early adopter and left-leaning user bases, with nontrivial overrepresentation of art, technology, and academic communities; generalizability to the broader social media landscape will depend on future migration and adoption patterns (Jeong et al., 2024, Smith et al., 16 May 2025). Customization, transparency, and user agency are advanced relative to legacy platforms, but usability and economic sustainability of third-party modules remain unresolved challenges (Balduf et al., 2024).

A plausible implication is that Bluesky’s experiment in modular decentralization will reveal new dynamics of polarization, group formation, and content diffusion under federated, user-controllable protocols. As research continues, the platform functions as a unique testbed for computational social science, recommender-system experimentation, and federated moderation algorithm development (Kleppmann et al., 2024, Greenwood et al., 6 Jan 2026, Chhetri et al., 14 Sep 2025).