Generating Almost Gibberish

An interesting thing I saw today:

Kind of silly, but after some further thought, it would be interesting if you created a semantic and syntactic model of the world that was rich enough such that you could run a "parser" backwards and generate sentences that were both semantically and grammatically sound, and yet random... and then to use a probabilistic model to generate a second sentence based on the first, and so on and so forth, until you had written an entire page.

If your system was strong enough, then it might be the case that 1 out of 1,000,000 generate pages would be a bit creepy in terms of its sensibility, right?

As it stands, this 100% random approach comes off as 1% interesting and 99% "sigh".