Essentially, if we have criteria of ``fitness,'' anything which can be turned into a bit string can be evolved by genetic algorithms. As should be evident to anyone reading this, literature, pictures, sound and movies can all be turned into strings of bits. Once we have a measure of fitness, there is no a priori reason we could not turn standard techniques loose on an initial population of pictures, or sonatas, or sonnets. There are even techniques, outlined at least by Holland, which will allow our system to modify the means it uses to evalute fitness. In particular, rather than mutating and recombining essentially random lengths of bits, the system could come to recognize that certain blocks of bits are meaningfully connected.
It also does not seem impossible, or even terribly difficult, to modify the standard techniques of genetic programming so as to work directly on two-dimensional or three-dimensional objects. Whether it would be worthwhile, I cannot say.