The deep learning buzz-word applies to models which learn feature interactions in a fairly complicated manner. While it is possible to explain what a model is doing, the model is a black-box in the sense that without a computer you could not develop the model. The model is making a lot of decisions, and explaining why all those choices are made is not really feasible. In contrast, I could create a linear model by hand easily, it just won't be as good at complicated predictions. However I think the stigma against black-boxes is undeserved. The human brain is a black-box and no one argues against using it.