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

TEXT START: Empirical research on content moderation is fundamentally constrained by the opaque deployment of moderation systems on major social media platforms.

The Dissection

The paper converts Bluesky’s moderation system from a black box into an auditable labor-and-classification pipeline. Its central finding is structural: automation handles fast, legible categories, while humans remain assigned to ambiguous and high-stakes judgments. The 10.6 million labels expose both operational scale and systematic failure—high precision, but catastrophically low recall.

This is not evidence that AI has failed. It is evidence of an intermediate hybrid system: machines process volume; humans ration judgment at the edges.

The Core Fallacy

The paper treats moderation primarily as a service-quality problem solvable through better accuracy, transparency, and system design. Under the Discontinuity Thesis, that framing is too narrow. Improving recall does not preserve broad human productive participation; it improves the machinery that replaces it.

The human oversight layer is therefore not a stable institutional solution. It is a temporary bottleneck and a future automation target. Once models become capable of handling nuanced cases, the remaining human role contracts toward appeals, liability allocation, policy exceptions, and reputational damage control.

Hidden Assumptions

• Harmful content can be reliably defined, labeled, and measured.
• Human annotators provide valid ground truth rather than culturally situated judgments.
• Higher recall is unambiguously beneficial; false positives and over-removal are treated as secondary costs.
• Human oversight can scale economically as moderation volume grows.
• Transparency automatically produces accountability or effective coordination.
• “Human-AI collaboration” represents durable complementarity rather than a transition phase.
• Better moderation can resolve the social harms generated by platform incentives and mass communication systems.

Social Function

Primary classification: partial truth.

Secondary classification: transition management and prestige signaling. The paper is not empty copium; its empirical results are damaging and useful. But its proposed endpoint—more effective, transparent moderation—channels a structural crisis into technical optimization. It makes the machinery more inspectable without questioning the economic order that continually expands the machinery’s required scale.

The Verdict

This is a strong autopsy of one moderation service and a weak account of the system’s historical direction. Its most important result is not that BMS needs better recall. It is that moderation is already organized as automated volume processing with a shrinking human judgment layer.

The paper documents P1 locally but does not establish the full Discontinuity Thesis. It shows the labor substitution mechanism in miniature: automation absorbs routine cognition, humans are retained for ambiguity, and the retained work is slower, scarcer, and more expensive. As those edge cases become automatable, the humans do not reclaim the system. They become its compliance furniture.

The corpse is still being annotated. That does not mean it is alive.