I have to imagine they contract that work out. Unless doing that was my business, I wouldn't want to touch that training set with a thousand foot pole. Well I wouldn't regardless..., but you see what I'm saying about encapsulating and isolating that.

I've always wondered about that. Surely possesing the 'training data' (basically child porn) is illegal. Only police can posses it for investigations etc. So how can you make something that can recongnise it?

There are companies who have special licenses, etc to do that type of work.

I was supposed to meet with a Scandinavian vendor a few years ago that provides a service where you can match SHA hashes of files against known illegal content. I recall the pitch deck said that they had millions of records to match against, which I found incredibly depressing.

I ended up bailing on the meeting... it's a subject I don't want to learn more about.

It's humans. There was a story about the 'dark side of Google' within the last few months about how Google contracts this out, and they burn out, but I'm not searching for that story directly myself.