An aside on why it would make sense for an increase of load on one phase could lead to a decrease of demand load on the other phases:
The purpose of the demand factor for receptacles is to try to capture the fact that not all the receptacles are going to be used at once. But whether a receptacle is used is not dependent upon what phase it's connected to. They're just taken as a collective whole. The more receptacles you have, the less likely it is that many of them will be used at once. So the demand factor for all of them decreases. That makes sense in my head, at least.
An aside to the aside: the statement that it doesn't matter what phase receptacles are connected to is not necessarily universally true. I worked on a project once that had several dozen sets of loads with identical kVAs that happened to be grouped into three circuits each. Although on paper the loads were identical, the way I knew the loads would be used would very likely result in one of the three circuits being used more than the other. Had I simply placed these loads in order of appearance on the panel (ABC, ABC, ABC...), one of the phases was likely to be a problem. Therefore, I placed each instance of the groups so that they were rotated on the phases (ABC, BCA, CAB). Just looking at the identical kVAs of each individual load, it didn't make sense to do this, but I knew it was necessary. This is just an engineering judgement situation that code and software can't really account for.