Just for giggles, I fired up my Markov chain implementation, in Java, and fed it El Reg's test data set. Here's the first 250 words:
"I want to kiss your lips I love you.
Oh, if only you I am in love with you is the greatest thing that has ever happened to me, and I've never been happier in my day, I always know it will be a good one because at the end of it matters because I'll soon be in your arms.
Thank you for being my vacation from the rest of the only ways to get out of a tight box is to have succeeded."
From this and a few other runs that produced similar results, I reckon Ye Olde Markov Chayne beats the pants off AI, not only for readability, but especially in terms of cost per Kloc of code written (50 Java statements, 0.05 Kloc) not to mention run time and CPU used ( 84mS run time, 108 mS cpu for 250 words of output on a Lenovo T440).