I think your point about the frequently gradual nature of the concept is a very good one, and I'll buy that there may be appropriate analogies to a black box in

I think your point about the frequently gradual nature of the concept is a very good one, and I'll buy that there may be appropriate analogies to a black box in this context, especially if some sort of poetic or metaphoric description is what one is aiming for. I see no indication of that in the article. What the present journalists, and many others, are doing in this respect is not analogy but rather destructive appropriation of terminology. I get that they mean something like 'is difficult to understand' too, which is in many ways completely uncontroversial. (I'm certainly not going to claim it's a solved problem to quickly and effectively come understand intellectually how an arbitrary ANN does what it does. I doubt there are many people who would.) If that's what they mean, then they can say that, or make up their own picture language that isn't already busy meaning the polar opposite of the situation they're alluding to. A black box is characterized by observability only at the edges and unknown inner workings. It is by definition an inappropriate term for a convolutional network where every single weight and operation and intermediate result is trivially inspectable and you can do things like follow the effects of an experimental perturbation along every step of every path through the network. 'But it's really hard to get an intuitive understanding of what that means in the big picture' is a perfectly legitimately concern for something to improve on, but it isn't remotely good enough as an excuse for effectively claiming we have no observability or control where both are clearly abundant. Abusing the term like this detracts from its established role in engineering, systems theory, etcetera and in my opinion also from communicating the actual problems of understanding how ANN:s do their thing. I really wish they'd stop making that claim. Now I'm going to leave my computer before I get started on the ######s talking about steep learning curves as if they were an obstacle.