One advice from a legal counsel that I received at that time still stands in my memory; she said: "Never look at the images, nevermind how curious you are. It's not just illegal and can land you in prison, but it's permanent. You cannot un-see something you've seen."
They returned to non Facebook employed humans doing the moderation
You'd still have to investigate/review partial matches, but something like that could cut down a lot on duplicate effort and could auto-identify a lot of things that would need manual review.
For some things liken nudity ML based solution work fine.
But other things such as hate speech are more nuanced and require human oversight.
That said ML is used everywhere possible to flag content to moderators.
For some categories the work of human moderators is used to train the ML models.
Facebook and other companies like YouTube are well insentivized to do the right thing here and automate as much as possible for all the reasons outlined in OP
I can't even imagine doing this job. This should be 100% automated with only 0.01% error rate for the first-pass filter (e.g. before it reaches humans).
They are willing to spare their users the horrors but not their own employees/contractors...