At least bugs in regular control systems can be identified and fixed, with high confidence that the problem has been resolved.

If the system is heavily based on machine learning, there may be no such code bugs or algorithmic mistakes to find, no easy way to show exactly what went wrong and why a fix will prevent it from happening again.

Tesla computer vision is done using neural networks. The tech has migrated to make driving decisions using neural nets as well. The code is basically just an artifact used to create a digital mind. However, it's possible that digital mind is operating vehicles at the level of a chimpanzee.