Unsurprisingly much like all ML things this (presumably) only works well if you're close to the training set. Even writing out pangrams the generated text only looks a little like my handwriting.

I specifically wrote double-decker As and Gs yet the the generated data has single story ones.