I’ve spent and continue spending a large portion of my time on that problem. We have at least 3 automated mechanisms to identify offensive content, and the last layer is human-driven. It is not easy or cheap, but alternative is magnitudes more expensive. Hope that answers your question