Characterizing Bluesky Content Moderation Service: From Automation of Service to Landscape of Harms

Motivation

• Auditing matters because under-moderation risks systemic harms, while over-moderation risks censorship of legitimate expression.The introduction identifies this as a delicate balance for deployed moderation systems.Source: §1 · p. 1

Audit opportunity

• Public moderation logs on decentralized platforms such as Bluesky make foundational inquiry into moderation mechanisms, efficacy, and purpose possible.The audit examines Bluesky’s publicly accessible moderation actions.Source: §1 · p. 1

RQ2: Effectiveness

• Precision was 0.837 and recall was 0.222; annotators flagged 4.5× more posts as harmful in a random sample than the moderation system detected.The study compared annotations of 1,000 BMS-labeled posts and 1,000 random posts from Bluesky’s firehose.Source: §1 · p. 2

RQ1: Automation and Human Oversight

• Median labeling delay was 7.1 seconds for porn but 8.4 days for rude, consistent with faster automation for clear-cut content and more human oversight for nuanced labels.The study uses labeling delay as a proxy for automation versus human oversight.Source: §1 · p. 2

RQ3: Landscape of Harms

• Clustering found harms from hostility toward protected groups to sexually explicit and graphic content, while the most taken-down posts contained violent death wishes to political leaders.The authors also report overlap among categories across labels, which may challenge moderators and confuse end-users.Source: §1 · p. 2

2.1 Content Moderation

• Earlier studies examined moderation mechanisms and transparency, but centralized-platform opacity constrained comprehensive mapping of systemic harms and moderation actions.The related work positions this audit across content moderation, regulatory intervention, and decentralized platforms.Source: §2.1 · p. 2, §2 · p. 2

2.2 Regulatory Intervention

• The DSA Transparency Database was created to improve transparency, yet studies report limited platform reporting and insufficient information about moderated content.These inadequacies limit researchers’ ability to understand systemic harms and detection effectiveness.Source: §2.2 · p. 2

2.3 Decentralized Platforms

• Open-protocol platforms such as Mastodon and Bluesky have enabled large-scale studies, including work on Bluesky activity, misinformation, polarization, and toxicity.Failla and Rossetti (2024) released a dataset covering over 4M accounts and 235M posts.Source: §2.3 · pp. 2–3
• Balduf et al. (2024) analyzed Bluesky’s architecture and labeling ecosystem, but not moderation decisions or the specific harms addressed; this study provides a large-scale empirical characterization of those areas.The authors describe their work as the first such characterization of a transparent moderation system.Source: §2.3 · p. 3

3 Bluesky and its Content Moderation

• Bluesky users receive a unique Decentralized Identifier (DID), and the PLC directory contains 66.9M DID documents.DID documents contain information including a user handle and the location of the user’s data repository.Source: §3 · p. 3

3.1 Key Architectural Components in Bluesky

• Bluesky’s Relay aggregates records from active personal data servers into the publicly accessible Firehose, which labelers consume and use to publish public label streams.The App View indexes architecture-wide data and enforces moderation labels in the user interface.Source: §3.1 · p. 3

3.2 Content Moderation on Bluesky

• BMS is applied to user accounts by default; labels can denote a harm category or an action, and default actions include hide, warn, or ignore.Third-party labelers require users to subscribe, and their relationship to BMS is left for future work.Source: §3.2 · pp. 3–4

3.3 Data Collection for Moderated Content

• The 2025 dataset contains 10,681,824 post labels across 13 label values, with images dominating many labels and text-only posts more common for rude and !takedown.The broader label stream contained 18,343,953 labels across posts, accounts, profiles, and other record types.Source: §3.3 · p. 4
• 88% of labeled posts were root posts, English represented 49% to 97% across label categories, and 69 country-specific BMS-labeled post records were collected.Root-post proportions varied by label, from 91–92% for porn, sexual, and nudity to 96% for rude, 99% for spam, and 79% for threat.Source: §3.3 · p. 4

4 RQ1: Automation and Human Oversight in the Bluesky Moderation Service

• Some post records could not be retrieved because repositories were missing, records deleted, or PDS endpoints unavailable.The passage does not quantify the resulting coverage loss.Source: §4 · p. 4
• The audit defines labeling delay as the interval between post creation and label application, treating seconds-scale delays at scale as likely automated and hours or days as indicating human oversight.Posts with backdated creation times or negative labeling delays were filtered out.Source: §4 · p. 5
• Seconds-scale delays for porn, sexual, nudity, self-harm, and graphic-media labels suggest predominantly automated application, while rude, intolerant, threat, and sexual-figurative labels take hours to days, indicating human oversight.The reported delay patterns are similar across content types with different multimedia attachments.Source: §4 · pp. 5–6, Figure 1 · p. 5, Table 3 · p. 6
• Porn labels have a median delay of 5.5 s and a Q3 delay of 9.8 s across more than five million labels; thumbnail embeds often take longer because external URLs add fetch and crawl latency.For nudity and self-harm, despite predominantly automated medians, 95th-percentile delays reach a few days.Source: §4 · pp. 5–6
• Longer interquartile ranges for labels such as intolerant, rude, threat, and sexual-figurative indicate variability in human moderation times; some borderline content also requires contextual judgment.The paper notes that wartime atrocities or breastfeeding may be labeled as graphic media or nudity, although whether to label them can be subjective.Source: §4 · p. 6
• Bluesky’s 2025 transparency report corroborates manual application of rude, intolerant, and threat labels, while appeal-based corrections explain delayed sexual-figurative labels despite a reported 0% manual application rate.The report lists manual application rates of 8% for self-harm and 18% for nudity, consistent with observed delays extending to hours or days.Source: §4 · p. 6

5.1 Precision and Recall of the BMS

• The study estimated precision from 1,000 BMS-labeled posts and recall from 1,000 randomly sampled firehose posts, using two annotators and a third to resolve disagreements.The primary annotators achieved κ = 0.802 on the binary safe/unsafe task.Source: §5.1 · pp. 6–7
• BMS precision was 0.837 on the labeled set, while recall was 0.222 on the random firehose sample.In the random set, annotators identified 27 unsafe posts and BMS had labeled six.Source: Figure 2 · p. 7, §5.1 · p. 7
• Precision varied by label: it was 0.903 for automated labels and 0.755 for manual labels, with rude and intolerant especially low at 0.509 and 0.667.Porn precision was 1.000; graphic-media precision was 0.802.Source: §5.1 · p. 7

5.2 Blindspots of Automated Moderation

• To investigate automated misses, the authors compared 336 labeled single-image posts with 336 similar unlabeled posts; Automod flagged 311 labeled posts, a 92.6% match with BMS.The unlabeled posts were selected for visual and semantic similarity to labeled posts.Source: §5.2 · p. 8
• Automod flagged 47 of 336 similar unlabeled posts; annotators judged 36 of the remaining 289 unsafe, a 10.7% lower bound on recall loss for automated labels.The 336 unlabeled posts were selected for similarity to known harmful content.Source: §5.2 · p. 8
• Of 36 missed posts, 22 (61.1%) had Hive scores on Automod-used heads below their hard-coded thresholds.These near-misses fell in the [0.55, threshold) range.Source: §5.2 · p. 8
• The other 14 misses (38.9%) scored at least 0.80 on Hive heads outside Automod’s rules, which ignored those categories.The failures arise in translating continuous scores into binary decisions using fixed thresholds and a fixed set of heads.Source: §5.2 · pp. 8–9
• The blindspot diagnosis used a smaller Hive sample because of its developer API limit, and the authors could not confirm that the 2024 open-sourced Automod rules match BMS’s current rules.The authors treat the open-sourced rules as the best available approximation.Source: §5.2 · p. 8

6.1 Clustering of Labeled Posts

• The analysis clusters multimodal labeled posts and uses an automated pipeline to name clusters, with human validation rating names 4.58 on average and 92.5% appropriate or very appropriate.The pipeline embeds text and media, projects embeddings with UMAP, and clusters them with HDBSCAN; posts assigned to noise are excluded.Source: §6.1 · pp. 9–10

6.2 Detected Harmful Contents by Bluesky

• Intolerant, rude, and threat clusters include identity-based discrimination, political and personal attacks, and calls for violence against people, institutions, and public figures.Examples include transphobic and anti-religious content, slurs and insults, and political death wishes and violent fantasies.Source: Figure 4 · p. 10, §6.2 · p. 10
• Graphic-media clusters include violent and gory imagery, while self-harm clusters include depictions or discussion of injury and some historical imagery used as political commentary.The authors note that a self-harm label on an image of an enslaved man’s scars suggests visual appearance may be classified without its broader context or intent.Source: §6.2 · p. 10
• Porn clusters center on explicit commercial adult material, while sexual, nudity, and sexual-figurative labels cover less explicit, nonexplicit, and stylized content.The paper also reports that porn is the top applied label and that action-label clusters often center on extreme political hostility, conspiracies, and extreme violence.Source: §6.2 · pp. 10–11

6.3 Implications of the Observations

• The observed intolerant, rude, and threat clusters may correspond to DSA risks involving illegal content, fundamental rights, and civic discourse.Graphic-media and self-harm labels suggest possible risks to public security and physical or mental well-being.Source: §6.3 · p. 11
• Sexual-content labels may indicate risks to minors and gender-based violence, but granular post-level mapping to DSA risk categories is outside this study’s scope.The authors characterize the findings as broadly suggestive of potential systemic harms.Source: §6.3 · p. 11

• The analysis uses publicly available posts containing text and images, excluding non-public account data and direct user interactions.The authors state that the work does not constitute human-subjects research.Source: §7 · p. 11
• Data storage and processing took place on secure institutional infrastructure, using in-house, open-source tools.Source: §7 · p. 11
• The researchers did not interact with users, deploy automated agents, or manipulate platform behavior, and present results as aggregated statistics.Source: §7 · p. 11

8 Concluding Discussion

• The audit finds that BMS labels operate more broadly in practice than their written definitions suggest, and that the service combines fast action on clear-cut violations with slower judgment for nuanced harms.The authors identify labeling delay as a proxy-dependent measure, note that removed posts may be missed, and leave downstream effects on user behavior and content prevalence for future work.Source: §8 · p. 11
• The authors propose using cluster analyses to guide increased label granularity and developing joint moderation analytics through open standards.They identify extending the audit to third-party labelers as a next step.Source: §8 · pp. 11–12

A Bluesky Moderation Service: Statistics

• Account labels most often use !takedown, while porn is the most frequent profile label; account labels show warnings on all content, whereas profile labels usually blur avatars.Source: A Bluesky Moderation Service: Statistics · p. 13

• Collection from March to December 2025 yielded 1.14B post records, of which 11.9M were labeled by BMS.Source: B Additional Experimental Details for… · p. 13
• For recall evaluation, the researchers randomly sampled 1,000 posts; a separate 40M-post subset was drawn from March–June 2025 and restricted to text and single-image posts.Source: B Additional Experimental Details for… · p. 13
• The 40M-post subset was embedded and indexed; for each of 336 labeled posts, the procedure selected the highest-ranked eligible unlabeled, single-image neighbor from 1,000 candidates.The resulting set contained 336 semantically similar but unlabeled posts.Source: B Additional Experimental Details for… · p. 13

C Automod and Hive AI Moderation

• Hive AI returns scores for 128 classes across 54 model heads, and Automod converts selected scores into actionable BMS labels.The heads span sexual content, violence and gore, drugs and vices, hate imagery, and miscellaneous image attributes.Source: C Automod and Hive AI Moderation · p. 13, Figure 5 · p. 14
• Hive’s classes are mutually exclusive within each head, with scores summing to 1; some heads distinguish categories such as levels of bloodiness.Automod reads only a subset of Hive’s classes, leaving many outputs outside its rule set.Source: C Automod and Hive AI Moderation · pp. 13–14
• The accompanying tables catalog available post records and their media, reply, and language breakdowns by label, alongside top account and profile labels in 2025.The tables also list Hive classes used by Automod and classes outside its rule set.Source: Table 5 · p. 14, Table 6 · p. 14, Table 7 · p. 14
• Automod applies hard-coded thresholds to 16 classes across 10 heads; the other 112 classes across 44 heads are unused by its rules.For example, a self-harm score of at least 0.96 triggers self-harm, and a sexual-activity score of at least 0.90 triggers porn; sexual labels follow a priority cascade.Source: C Automod and Hive AI Moderation · p. 14, Figure 5 · p. 14

D.1 Country-specific Moderation Services:

• Country-specific !hide labels range from days to weeks in Turkey and Brazil to nearly a year for a few German posts.The authors characterize these delays as suggesting country-specific services act as reactive compliance rather than proactive safety systems.Source: §D.1 · p. 14, Figure 6 · p. 15

D.2 Negation labels & Redressal mechanism:

• 206,724 negated post labels across 13 categories show varied removal delays, including 44–47 days for spam and !warn, versus 21 hours for self-harm.Rude and intolerant labels have median delays of 12–15 days; graphic-media and sexual labels are negated in 1.5 and 3.4 days, respectively. Many negated content labels may have been automated and erroneously applied before redressal.Source: §D.2 · p. 15, Figure 6 · p. 15

E.1 UMAP dimensions and HDBSCAN clustering:

• The clustering analysis selects five UMAP components as a trade-off and uses HDBSCAN to form clusters with soft membership probabilities.When a label corpus exceeds 300,000 posts, hyperparameter trials use a uniform random subsample of that size; Bayesian optimization selects the final configuration by DBCV score.Source: §E.1 · pp. 15, 17
• Cluster coherence is assessed quantitatively with DBCV, while a separate LLM-based pipeline supplies qualitative descriptions and names.The selected hyperparameters and DBCV scores are reported for each label.Source: Table 9 · p. 15, §E.1 · p. 17

E.2 Interpreting clusters

• Each cluster is represented by a statistically justified sample, with posts weighted by HDBSCAN membership probability to cover its thematic range.The sampling formula uses a 95% confidence level and a 5% margin of error; the approach avoids processing every post in clusters that may contain tens of thousands.Source: §E.2 · p. 17
• Qwen3-VL summarizes sampled posts in batches of 32, then merges chunk summaries into a final description using a map-reduce strategy.Model inputs include post text and attached images or video, and greedy decoding makes outputs deterministic.Source: §E.2 · p. 17
• Cluster names are generated jointly within each label to distinguish sibling clusters and reduce generic, near-identical names.The authors report that cluster assignments are generally coherent and meaningful, despite some overlap.Source: §E.2 · p. 17
• Human reviewers rated names 4.58 out of 5 on average, with 92.5% of ratings at 4 or 5 and ordinal Krippendorff’s α = 0.686.Three annotators independently rated 31 clusters; images and video in sensitive categories such as graphic-media and self-harm were avoided in the validation materials.Source: §E.2 · pp. 17–18

F Action Labeled Contents and Spam

• Taken-down posts predominantly involve celebrations of real-world violence and explicit calls for harm against named individuals.Clusters include violent rhetoric against right-wing figures, celebration of Charlie Kirk’s shooting, and Luigi Mangione as a populist avenger.Source: F Action Labeled Contents and Spam · p. 18
• Hidden and warned posts include threats, harassment, self-harm promotion, political hostility, and doxxing targeting public figures or DOGE employees.The hide clusters include threats against Donald Trump and Elon Musk and a doxxing campaign targeting young tech professionals; warn clusters include leaked email lists targeting DOGE employees.Source: F Action Labeled Contents and Spam · p. 18
• Most spam-labeled post records were unavailable, limiting analysis; the few clusters include political and activist content such as humanitarian appeals and protest.The spam label is described as covering unwanted activity from bot-like behavior and solicitation to coordinated inauthentic posting.Source: F Action Labeled Contents and Spam · p. 18