Interesting article. I think that, without deeper insights into machine learning it will be very difficult to make it better. At best, like the article says, we can guess why it made its decision. But we really don't know for sure. It seems like we will end up tweaking algorithms until we get the 'best' model, but will it ever be truely correct. Or just happen to be mostly correct. Does it really know what its looking for, or does it just guess correctly most of the time.
Calling machine learning 'it' is probably already a step too far.