What I find entertaining is that while answers may be wrong, you can see from the question statement and the returned sample text that it is working perfectly in a mechanical sense. The parsing of the question is correct, the searching works, and so does the ranking. But at the end of the day, it is just applied NLP at a large scale, not AGI, so it doesn't know that it is returning a wrong answer. It doesn't know anything at all!