In these animations it's easy to see it, as the shape recognition is not very stable. For example, in the last video when the prompt changes from "people" to "bears" you can see a backpack turning into a bear head, which then turns into an open-mouth bear head. And then, the arm of the person carring the backpack is also turned into more bear heads.
The next steps in the evolution of this technique should be in exploring the relation between noise and subjects, so that you can create variations of the same image maintaining the recognizable parts stable.