Except AI code inside self driving cars isn't crafted by hand, but is a matter of training data as much as code. How can you determine that a training data is "bug free" ? It barely makes any sense.