From a technical perspective, they're doing exactly what they're designed to do:

Self learn (i.e. unsupervised) patterns (using multi-head attention mechanism) from a large volume of input and then generate new patterns based on the. prompt.