The shakespeare generator isn't just reproducing the syntactic structures, it occasionally seems to capture meter. The samples you've reproduced here aren't iambic, but they are around ten or eleven syllables per line, which is impressive enough in itself. In the longer passages, it manages some proper iambic pentameter:
My power to give thee but so much as hell:
Some service in the noble bondman here
It doesn't seem to have managed to pick up on rhyming couplets, though.
A quick search of Shakespeare's corpus also shows that Shakespeare never called a bondman 'noble'; there must be some conception of parts of speech being captured by the RNN, to enable it to decide that 'bondman' is a reasonable word to follow 'noble'.
I'd imagine the lack of rhyme is likely due to the fact that English pronunciation is ambiguous. Given only the text, it would have no way of picking up the fact that, say, "here" and "beer" rhyme, while "there" does not.
(Put another way, English text is a lossy representation of English speech.)
Perhaps if you were to feed the IPA representation of each word in alongside the text, the RNN would do a bit better, though admittedly I'm not sure how you would do so.
If this is the case, I'd imagine training it against Lojban text would see similar results.
A quick search of Shakespeare's corpus also shows that Shakespeare never called a bondman 'noble'; there must be some conception of parts of speech being captured by the RNN, to enable it to decide that 'bondman' is a reasonable word to follow 'noble'.
So yes, "unreasonable" seems about right.