Hacker News Reader

I was an early-ish Facebook employee, and started before they contracted out moderation. I once passed a content moderator in the lunch room, flipping through flagged images, about twenty per second. Presumably she was just reviewing user & automatic flags for accuracy.

What they see is: Penis. Pages and pages of penis. Penis in all shapes and sizes and orientations and conditions. But it was all penis.

View on HN

Porn detection with Deep Learning is 99% solved problem, I bet those moderators only get to view content where the model reports lower confidence, the rest is automatically removed. The same for gore etc.

Think how many image posts and messages there are on Facebook and Messenger. Assume your Deep Learning model has 99.9% recall. It will still miss millions per day.

That's conceivably manageable if you're Facebook's size with tons and tons of moderators, while improving the data set.

Which is exactly what they do! For better or worse.

20 images per second x 8 hours a day is half million images per day per employee (quoted by gp, though its seem to be unsustainable), and Facebook has an army of content moderators so they probably manage somehow.

“Oh, look, a hot dog!”

Reply on news.ycombinator.com