Markov chains help a lot in making the output look like the real stuff, that's kinda their purpose in this case. Heck, there have been automatically-generated conference papers that have been accepted and went through peer review (although this might tell a lot about that process, too ...).

git manpages are apparently full of very git-specific terminology and quite dense. I would guess the same model applied to the J documentation would be equally unreadable for the uninitiated. If the source material was more readable and easier to grasp the automatically-generated versions would reflect that (to some degree, until it's obvious that's nonsense).