Thoughts about the utility of schools to serve this function? We are already teaching humans very simple to very advanced tasks and we do so millions of times a day. Is the fact that tasks/knowledge are being structured for paced human learning irrelevant for this purpose?

It's not irrelevant, since there is a minor area of ML called 'curriculum learning' which asks how to order examples to teach a ML algorithm most efficiently, and it comes up in some other contexts (for example, a boosting algorithm which focuses on optimizing performance on hard/misclassified cases can be seen as somewhat like curriculum learning, and there are variants of gradient descent which focus on hard examples rather than wasting time on cases where the NN can already get the right answer; and for 'active learning', you want to pick the example which will teach the algorithm the most), but there's not much you can take from known pedagogy at the moment and apply straight to NNs. Even the non-bullshit parts of education like spaced repetition have no clear analogues for tasks like 'train an RNN to write news headlines based an article text using this corpus of human-written headlines'.

In fact, "expert systems" were built around this principle, though generally more in the industrial context.

I don't think that the way knowledge transfer is structured for humans is by default distinct from how we would do it for AI in the long term, but that's a pretty wide range of pedagogy so some tasks make more sense than others for the state of the art today.

In fact the way we train ANN/CNN is similar to how infants & toddlers do early classification. Frank Guerin at the university of Aberdeen has done some study in this area that is in it's early stages.

As far as I know there have not been efforts to set up schools to train AI's from the general population.

The closest thing to it are the efforts by Google and Facebook to train their classifiers (facial recognition, text detection etc...). We are working on an internal process that uses user input to classify objects in the home (furnishings mostly) for auto segmentation and detection.