When we model a complex system using theory (explanation), we must ignore some details about its nature: the parameters we pick, although good for developing our human intuition, often misses some interaction which is unexplainable or unknowable to us. With a large enough dataset which is trained to produce many parameters purely in a predictive fitting way, there is no room for our biases on what the model should look like. Best we can hope for, all cases in the data are accounted for, but how we got there is now hard to explain.
