The program, is the machine learning conditional framework. The training is data that influences program execution. The program is what the machine follows, the behavior exhibited by the program is determined by the training data.
The problem, as you correctly put, is that we dont really have good feedback on what elements of the training data it is weighing on reliably.
The actual program is what defines the conditional framework at the lowest level. It does this faithfully. The emergent properties? That's another story. That is related to the data the AI has collected.