In the case of machine learning, it HAS to be useful that we know in some detail how the machines were built, at least in the sense of knowing the overall structure of the machine, what algorithmic process was used to optimize its parameters, what environment it evolved in (which is to say, what the training set was), and the objective function.