This isn't terribly hard to achieve. E.g with a pretrained LSTM generator downloaded from some place that does neural network stuff + a homebrew text classifier (I'm fond of character-level online linear models with long n-gram ranges) that's just underpowered enough to overspecialize to a certain style, but unpredictably so. Make this poor man's model generative by choosing random combinations of characters that hit a high enough score; and then feed it to the LSTM (along with normal text samples).

The underpowered generator creates weird stuff that the good generator absorbs into its language model. Boom, you are witnessing the technological singularity. Or something.

Yeah!! "You keep using that acronym..."