The early successes with AI on chess programs led to a false confidence that we would have machine intelligence shortly thereafter. It was decades after the first reasonable chess playing program before we had any kind of reasonable natural language recognition, which is something that three year olds master. GO is a much more complex game than chess and it was quite an accomplishment to create programs that could win at the game, but even so it is not a very good analogy for real life. Games are significantly easier to create learning networks for than real life because with a game you have a perfect knowledge of your current state (no failed AOA sensors) and a very deterministic outcome -- you either win the game, or you lose it.
There are many challenges to the problem. Pattern recognition is one: it is easy to recognize a pattern in a game, a little harder to recognize a pattern in a picture, and I have no idea how you recognize the pattern of what your jet feels like when it is hit by wake turbulence on takeoff -- but I am sure all of the real pilots here 'instinctively' recognize that pattern and can distinguish it from the feeling of taking off from a wet runway, or in a crosswind, or what it feels like if a tire blows out on takeoff (if that is something that you practice in the sim.)
Weighing the outcome is another major difference between game play and real life. The player with the most enclosed spaces wins the game of GO, so it is pretty simple to score. When the ending condition has been met, count up the number of enclosed squares and the winner is the one with the most of them. In an excellent post earlier, a poster brought up what the "Sully" question -- with no engines, is it better to try to return to the airport with a nonzero probability of accidentally recreating 9/11, or is it better to try a water landing? How do you score the neural network's decision? If 3 times out of 10 the "return to base" scenario kills 1,000 people on the ground is that a failure? What about 3 times out of 100? What are the weather conditions? What if you are an American plane in this situation over Moscow at a time when the US and Russia are on the brink of nuclear war? Does it make a difference if the plane is full of Mexicans, or if Mitch McConnell is onboard? The neural network will faithfully reproduce whatever value decisions that you make (which is one of the real dangers of using AI for police and military work.)
There is also the question of transparency. Do the people onboard have the right to know about these value judgments? I'd certainly like to know if the automatic plane is programmed to self destruct if a failure occurs over a populated area. With a human pilot, I can be fairly sure that the decisions that are made will closely match the decisions that I would make in the same situation, since the pilot shares my fate. A computer pilot doesn't care about survival at all, and a remote pilot knows he is going home at the end of the day no matter what the outcome for the passengers.
It is a tricky issue, and I think it will be a long time before the public cottons onto self driving airplanes, and by that time we may not be able to fly anymore anyway. The problem with the AI approach is that it only works with perfect humans, of which there have been very few in history. The self driving Uber car sounds great, but who wants to get into a car in the morning that got puked all over last night?