In D-to-A conversion since you are going from periodic, discrete data to a continuous signal, it is inherent that you have to fill in the gap with fabricated information. WIthout upsampling you are leaving it to the conversion process to decide how to join the dots. With upsampling, you guide the process by inserting samples which constrain the conversion to what you consider more realistic. In either case you are fabricating missing information but if you do it intelligently you can expect the end result to be more constrained towards realism.
