Making the right mistakes.

Soon six computers will be throwing it down Turing Test style. Maximum PC has an article about it here.
The Turing Test is a test where a researcher chats with the computer, and if the researcher can't tell if he is talking with a computer or a person then the machine passes the test. The article has an example of the test; here is the part of the computer conversation that I want to highlight:

KW: What mysterious subjects?
Subject: Think deeply about mysterious dubjects.
KW: What's a dubject?

KW is administering the test. "Subject" is the computer.

Why would a computer misspell the word subjects? I am guessing to be tricky. You see, computers can't make a mistake, so having a computer make a mistake will throw people off the trail. However, when I was reading the two conversations in the article this exchange made me immediately think that this was computer making an intentional mistake.
Why was it so obvious? I think the first thing is that a real person who is trying to prove that he is a human will first make sure that he doesn't make obvious mistakes. Why? Because everyone is looking at what you type and you don't want to look like a fool. So I doubt that any mistakes made would be done in obvious places like the first letter of a word. A better choice would have been to misspell subjects as subects, or subjectc.
Also I think that a human would pay more attention to focus of the sentence, perhaps the computer should have misspelled the word "deeply" or 'about" instead.

So how would I instruct a computer is purposely misspell words? The first thing that I would do is to make different levels of chatting. Writing text for a casual conversation will have more mistakes than when writing for a formal matter. Secondly, I would weight the chances for each word to be misspelled based on their importance in the sentence. I would also place the misspellings in the center of the words since it is harder to catch those when checking over the sentence for errors. And finally I would reference a large bank of words that are typed, maybe gathered from the internet, that has common ways to misspell a word so that the misspelling seems more natural (because you can't just pick a word to misspell, you also need to know how to misspell it.)
I would think that having a computer make logical mistakes would be one of the simpler things for a programmer to figure out.

Those are my ideas on the dubject.