Originally Posted by
ferry pilot
The big problem with AI is that it is not coded and then verification tested at unit level, it is a program built by the Machine Learning. While the engineers working with the ML know
how the ML is generating its code they have no idea what the code is and cannot verify the code - check that the code is correct. They can only validate the code - check that the system does what it is meant to do in every test that is given to it. This does not meet the certification requirements of current systems. So an AI driven system built based on ML cannot be certified under the current approach. There is also a fundamental issue with AI/ML it cannot use analogous reasoning so something that it has not seen before happens and it cannot as a human would take action that worked in a similar 'analogous' event. Or as someone once put it if the AI/ML learns only on left hand circuits it will not be able to carry out a right hand circuit. There is a LOT of research in this area including research into the CRM aspects of a human pilot supported by an AI/ML pilot; it could be orders of magnitude more than the 'what's it doing now?' How do you brief an AI/ML first officer on what you will do in particular events so it does not jump in to stop you?