One new job I can think of, machine learning psychiatrist. A big category of AI, machine learning using neural networks are really interesting because we have essentially thrown together a network of back propagating weighted gates and then thrown training at them. Its brilliant, it works much of the time, but we can't troubleshoot it easily (or much at all?) by examining its internal state, as its essentially millions upon millions of values that are produced not by entering them by a keyboard, but by training the network with "input" coupled with what we want the "output" to be.
A particularily illustrative example is training a network to categorize animals present in pictures. When shown many pictures of animals and what they are (ie., training the network), the network internally assigns weights in a particularily non transparent manner that with each iteration allows the network to become better at recognizing animals. How good does this neural network deep learning ai get?
So this is a new thing, we can't seem to troubleshoot these things by reading the internal state (because essentially they're just millions of "weights" attached to digital constructs that emulate neuron like behavior (I guess, feel free to correct me there if you actually know about these things), but rather we become behavioral psychiatrists to these silicon constructs.