Content moderation
The work of deciding what may remain on a platform. It consists of rules, automated classification and human reviewers — in that order, and with shrinking margins at every step.
Automation handles the obvious: hash matching against known databases for child abuse material, classifiers for spam, nudity and violence. Everything borderline — satire, news reporting about violence, threats in local context, sarcasm — lands with humans, often at subcontractors, with documented psychological harm; Facebook settled with US moderators for $52 million in 2020. Scale makes perfection impossible: at millions of decisions a day even 99 percent accuracy produces tens of thousands of errors daily, and both error types generate headlines. The EU's DSA has since 2024 imposed requirements on appeals, transparency reports and risk assessments.