A lot of machine “learning” is based on neural networks, but the results of this learning are - at least currently - impossible to audit or unravel. There are interesting experiments that show that minor alterations to e.g. traffic signs can render them unreadable to or even cause current technology traffic sign awareness systems to “see” a different sign. Rather innocuous looking stickers can foil camera-based perception systems. Humans and neural networks rely on entirely different means to perceive the world and we certainly don’t want to end up at a point where a strategically placed dot pattern on a billboard under a final approach will cause an automated airliner to go-around or worse. The “if it acts like a dog” approach to test automated piloting systems by submitting them to the same tests human pilots have to pass to get a license is a fallacy. I have no doubt that autopilots are able to fly better than pilots as they lack the slow biological processing between perception and action. But pilots are required to demonstrate but a small percentage of their skills during a check ride because of the assumption that they will be able to use their systems knowledge to extrapolate in case something requiring extra skills happens to them. There is no base in assuming that a machine learning system will be able to deal with a situation it has never encountered before and much of our current certification system is not based on multiple failures happening (although they do, as demonstrated by the Quantas A380 incident. (As impressive as the Garmin Autoland is, it is designed as a measure of last resource to avoid an even worse outcome - crashing without a pilot - and AFAIK is based on the assumption that there is no other failure beyond the pilot not being able to perform his duties. Should the pilot pass out because e.g. an engine failure excites him to the point of a heart attack, the Garmin system won’t be as helpful as another pilot. The much cited dogfight scenario was based on the computer opponent having total situational awareness, something not usually afforded to human pilots who have to fly airplanes the actual status of many parts of which they don’t know for lack of sensors, external cameras, etc.)