Update: I took a corpus of personal chat data (this way it wouldn't be seen in training), and tried asking it some paraphrased questions. It performed quite poorly.

Oh, interesting!