Original title: Characterizing Bluesky Content Moderation Service: From Automation of Service to Landscape of Harms
Pushpdeep Singh, Sayeh Jarollahi, Ayan Majumdar, Vabuk Pahari +4 · September 2026
Most social media apps keep how they police content secret, making it hard to study. Researchers took advantage of Bluesky, a decentralized platform that shares public moderation logsOpen records of every moderation action a platform takes, which let outsiders study how the system behaves., to audit its default tool, the Bluesky Moderation Service. The researchers studied how the system works, how accurate it is, and what harms it catches. They found that humans and artificial intelligence work together. Automated tools flag sexual and graphic content in seconds. Meanwhile, nuanced and high stakes issues rely on human reviewers, taking hours or days. The system rarely makes mistakes when it flags something, achieving high precisionOf all the posts the system flagged as harmful, the fraction that really are harmful.. However, it suffers from low recallOf all the harmful posts that exist, the fraction the system actually catches., meaning it misses most bad content. Human annotators found 4.5x more harmful posts than the tool did in a random sample. Finally, the system catches harms ranging from hostility directed at protected groups to graphic and sexually explicit posts.
Imagine a lifeguard watching a crowded beach. A drowning swimmer is easy to spot — bright clothes, thrashing — and the lifeguard reacts instantly. But a swimmer being pulled out by a rip current can look calm, and those cases take long, careful watching to notice. In this paper, Bluesky's moderation system works the same way: obvious sexual or graphic content is flagged automatically in seconds, while subtle harms, like hostility toward a protected group, need human reviewers and take hours or days. And just as a lifeguard misses many quiet drownings, the system misses most harmful posts.
Content moderation shapes what people see on social media, yet major platforms keep its inner workings secret. Bluesky's public logs let outsiders measure, for the first time at scale, how a real moderation system performs — and the audit shows it leans heavily on people for tricky calls while missing most harmful posts. That gives users, researchers, and policymakers a concrete baseline for demanding better.
The work is an audit of one platform's default system in 2025. It does not establish whether the precision/recall figures or the harm categories generalize to other platforms, time periods, or moderation configurations; the abstract also does not report the size or selection details of the manual annotation sample.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Characterizing Bluesky Content Moderation Service: From Automation of Service to Landscape of Harms — Pushpdeep Singh, Sayeh Jarollahi, Ayan Majumdar et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.