Oh yeah, obviously tenuki would confuse the heck out of it, but I expect the human player might be surprised by some unexpected "tenuki" too, even in the late opening (why did he just attach to the top of my 10-3 stone? Answer: because it was a normal extension in the most common sequence after the last joseki was played)
Really, I was driving home a few nights ago listening to the System Shock 2 soundtrack and realized that the evil computer voice "speech" during the credits track, while probably an actual human speaking into a microphone and then distorted beyond intelligibility, could also be done by taking samples of phonemes and analyzing a few audiobooks to provide some good patterns of sequences that are possible to pronounce, then using a low-order Markov chain to construct your almost understandable gibberish.
From there it was a short jump (in my tired state) to say "hey, I wonder how a Markov-based Go bot would play?" and decide it might be amusing enough to pursue.
Like all Markov-based things, it's more amusing if you try to assume the naive mindset that there's actually some sort of intelligence making these bizarre decisions (which is easy to do - after all, they all look logical in isolation, right?)
Anyone who doesn't know what a Markov chain is, it's an algorithm that constructs the next step in a given sequence based on statistical analysis of past data. Garkov is a good example, where it replaces the speech bubbles in Garfield strips with words based on analyzing a database with the text of all the strips (and generally is way more amusing than the original comic.) One thing I always wonder about things like Garkov though - how does it choose the first word? Based on what the most likely starting words are? Or does it pick a word from the database at random?