Skip to content
anythingai.space
Case Studiesadvanced14 min · 8 steps

Content Moderation at Scale

Moderate 10M posts a day with a four-tier cascade — filters, classifiers, an LLM and humans — so each tier only sees what the cheaper one couldn't decide.

harm exposure2%
false positives15%
critical review240min
every flagged postACTION DEADLINE BY CATEGORYCSAMimmediateViolence< 1 minHate speech< 5 minHarassment< 15 minSpam / misinfo< 1 hourAdult content< 1 hourIngestion10M posts/dayHuman moderators500 moderators4h backlog
requestheavy workerrorresponsedatasignal
metrics are illustrative
1 / 8
ProblemStep 1 of 8

50M users, 10M posts a day, 500 moderators

A social platform with 50M daily active users gets 10M posts per day. Today 500 human moderators take an average of 4 hours to review a post, 15% of removals are false positives, and 2% of harmful content reaches users.

Goals: harmful exposure < 0.1%, priority content reviewed in < 15 minutes, false positives < 5% — and scale without hiring moderators linearly. Severity sets the deadline:

CategoryActionLatency
CSAMBlock + reportImmediate
Violence / goreBlock + review< 1 min
Hate speechBlock + review< 5 min
HarassmentReview + warn< 15 min
Spam · misinfo · adultDeprioritise · label · age-gate< 1 hour

Lesson 5 of 20 in Case Studies