The problem currently is that a machine could not translate as good as an average interpreter. Languages are not one-to-one mapping, and machines are not good at analyzing and choosing which words to use and in what order. And this is why interpret is still a career.
Whenever a computer is not good at something, we could try to train it in hope that it will eventually learns and masters it. The source of training material would be from thousands of millions of people. As in image recognition, tags from Internet users helps classify and recognize pictures a lot. If, in one day, computers could seamlessly translate one language into another, we would be happy that we no longer need to learn a second language, and we would be equally sad that we will never drive to learn other than native language.
Currently, machine translation in instance message system should be more than experimental, but at least it could provide some information which might be helpful to international talkers. And, have fun to laugh at silly computers.