The question of should we is already being discussed, in depth, in another thread.
A friend of mine came up with an alternate to the Turing test (he has this weird mind that wants to reverse everything... I think that's why I like him). He said that the real benchmark of AI is not whether it is indistinguishable from a human in the turing test, but whether it can predict AI vs human better than humans can. That is to say can the AI tell that it is talking to a computer?