I know most are a bit tired of articles on AI (and many other technology trends for that matter), but thought I would provide a link to an insightful post from MIT Technology Review on deep learning, its complexity, and considerations on what is a reasonable level of transparency. Of course, the level of transparency should be different based on the application. Personally I don't mind fraud detection being more opaque, but applications in medicine and military should certainly have a greater level of transparency. Anyway, here is the link to the article. Enjoy.