The contractors’ feedback is certainly considered a high-signal dataset. I am not directly familiar with that aspect of the company (I worked on things somehow related but not this) but I suspect there are several ways of training and assessing estimations that certain content breaks certain rules (gore, direct threat, child endangerment, etc.). There could be challenges (legal, quality issues) of using automated decision to block or not block different types of flagged content.

One thing that, as a data scientist, I noticed: none of those reports complains about seeing the same image repeatedly. I understand that posters are often tempted to spread new content, so I would expect there to be many repeats. It could be because they see so many similar and shocking things they don’t recognise. Or companies could use techniques to pool together content that is identical or similar enough.