Characterizing Bluesky Content Moderation Service: From Automation of Service to Landscape of Harms
Pushpdeep Singh, Sayeh Jarollahi, Ayan Majumdar, Vabuk Pahari, Abhijnan Chakraborty, Krishna P. Gummadi, Ingmar Weber, Abhisek Dash
cs.CY · cs.AI
Quick verdict
A convincing first large-scale audit of Bluesky's moderation service; novel transparency findings, but recall rests on 27 posts and coauthor ground truth.
Summary
This paper presents the first large-scale audit of Bluesky's default content moderation system (BMS), exploiting its public label stream to analyze 10.6M post labels applied during 2025. Using labeling delay as a proxy for automation, the authors show that visual labels are applied within seconds (porn: median 5.5s), while rude (median 8.4 days), intolerant, and threat labels show delays consistent with human oversight — a pattern corroborated by Bluesky's own transparency report. Coauthor annotation of 1,000 stratified labeled posts and 1,000 random firehose posts yields an estimated precision of 0.837 (1.000 for porn, 0.509 for rude) and a recall of 0.222: annotators flagged 27 unsafe posts in the random sample, of which BMS had caught 6. Replaying the automated pipeline (Hive AI classifier plus open-source Automod rules), the paper attributes misses to near-threshold scores and to Hive output classes that no rule reads. Clustering (UMAP/HDBSCAN with LLM-generated names) maps each label's harm landscape, from protected-group hostility to violent death wishes against political leaders. Main limitations: ground truth rests on three coauthors annotating replies without thread context; the recall estimate has a 27-post denominator with no confidence intervals; clustering discards up to 53% of posts as noise; and only BMS, not third-party labelers, is audited.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- BMS operates as a human-AI collaborative system: sexual and graphic-content labels are applied automatically within seconds (porn median 5.5s per Table 3), while nuanced labels like rude (median 8.4 days) show delays consistent with human oversight, corroborated by Bluesky's 2025 transparency report.
- Human annotators agree with 83.7% of BMS's labeled posts (precision 0.837; 1.000 for porn, 0.509 for rude) but flag 4.5× more harmful content in a random firehose sample, implying recall of 0.222.
- Low automated-label recall traces to two pipeline failure modes in Hive AI + Automod: scores just below hard thresholds (22 of 36 missed posts) and Hive classes that no Automod rule reads (14 of 36).
- Unsupervised clustering shows BMS's operational label semantics are broader than written policies, spanning protected-group hostility, political death wishes, sexual content, and graphic media, with substantial cross-label overlap.
Strengths
First large-scale audit of a deployed moderation system via public logs
Uses BMS's public label stream for all of 2025 (18.34M labels; records retrieved for 10,681,824 post labels across 13 label values, Sec. 3.3/Table 2), establishing a reusable collection methodology (PLC directory → PDS resolution → getRecord/getBlob).
Proxy inference externally validated, discrepancies honestly resolved
Labeling-delay-based automation classification (Table 3: porn median 5.5s vs rude median 8.4d) is corroborated by Bluesky's 2025 transparency report; the sexual-figurative conflict (0% manual in 2025 report) is resolved via the 2024 report's appeal-driven application.
Mechanistic diagnosis of low recall is novel and actionable
Candid error analysis
Clustering pipeline human-validated
Weaknesses
Ground truth produced by the authors themselves
Fragile recall estimate without uncertainty
Aggregate precision is not prevalence-weighted
Delay proxy conflates mechanisms; internal number inconsistency
Harm landscape skewed by noise exclusion
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper conducts the first large-scale audit of a live, deployed content moderation system — Bluesky's default Bluesky Moderation Service (BMS) — exploiting the AT Protocol's public label stream. From 18,343,953 labels applied in 2025 (15.7M on posts), the authors retrieved records for 10,681,824 post labels across 13 label values (Sec. 3.3, Table 2). Three questions structure the audit: degree of automation vs. human oversight (RQ1), detection effectiveness as precision/recall (RQ2), and the landscape of harms the system actually catches (RQ3). Central claims: (1) BMS is a human-AI collaboration with a sharp dichotomy — automated visual labels in seconds, human-mediated labels in hours-to-days; (2) precision is high (0.837) but recall low (0.222, with annotators flagging 4.5× more harm); (3) clustering reveals de facto label semantics broader than written policy. A footnote states acceptance at ICWSM 2027. Claims are generally well-scoped; the authors repeatedly hedge the automation inference (likely/potentially automated).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What precision and recall does Bluesky's content moderation system have?
According to this audit of about 10.6 million Bluesky moderation labels from 2025, the default Bluesky Moderation Service achieved an estimated precision of 0.837 (83.7% of flagged posts were judged genuinely harmful by annotators) but a recall of only 0.222 in a random 1,000-post sample: annotators flagged 27 posts there as unsafe, and the service had caught 6 of them. Per-label precision ranged from 1.000 (porn) to 0.509 (rude).
Is Bluesky's moderation automated or done by humans?
Both. The study uses labeling delay as a proxy: image-based labels like porn (median 5.5 seconds), sexual, and graphic-media are applied automatically within seconds, while labels like rude (median 8.4 days), intolerant, and threat show delays consistent with substantial human review. The split matches what Bluesky's own transparency report discloses about manual application rates.
Why does Bluesky's automated moderation miss harmful content?
The authors replayed Bluesky's automated pipeline — the Hive AI vision classifier plus the open-source Automod rule engine — on unlabeled posts similar to known harmful ones. Misses came from two failure modes: classifier scores falling just below hard-coded thresholds (e.g., a 0.98 cutoff on underwear-related heads), and Hive output classes that no Automod rule reads at all (such as general_suggestive and yes_cleavage) even when scored near 1.0, because Automod acts on only 16 of Hive's 128 output classes.
What content gets moderated most on Bluesky?
Porn is by far the most applied post label — 7.36 million of the 10.68 million collected post labels in 2025 — followed by sexual (2.24 million) and nudity (271 thousand). Among manually reviewed categories, the most taken-down content contained violent death wishes directed at political leaders, and account-level takedowns (2.04 million in 2025) were the most common account label.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure panels (Figs. 1-4, 6-7) are extracted only as axis labels, titles, and captions; figure-internal content such as CDF shapes and cluster maps could not be assessed directly, though the key quantitative values are available in Tables 1-9 and the body text.
- Individual cluster names and sizes referenced in Sec. 6 (shown in Figs. 4 and 7) are not enumerated in the extracted text, so cluster-level claims beyond those named in the prose could not be independently checked.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.