I think the bigger issue is that a lot of the demand for labelling training data is now in highly-specialised fields (i.e. things like medical imaging), and mechanical turk's focus was on the generalist problems