Your question makes some broadly misleading assumptions about both translation and machine translation.
The problem here is not "grammatical" correctness. It's not a problem for Google to generate superficially grammatical sentences. But sentences only appear as part of text/utterance which is the really carrier of communicative meaning. Any one word or phrase only makes sense in that context. Also, texts have their own idiomatic conventions which depend on many factors.
So you have to evaluate MT (machine translation) at the level of the whole text, utterance or conversation. And here you find that often the coherence breaks down. You see translations of texts where there are multiple contradictory interpretations of an event/situation in the target text that are not present in the source text.
The reason is that unlike with a human translation, machine translation does not start with a thorough understanding of the source text. It combines some basic parsing with stochastic methods - essentially dealing with probabilities based on huge corpora.
But because language is basically infinitely variable, even the vast corpora at Google's disposal are not enough to guess at the right equivalents for every sentence. This becomes exponentially more difficult because most sentences only make sense in the context of other sentences. Current MT has almost no way of evaluating the role of context in disambiguation and has to pretty much stay at the level of single sentences with some basic anaphoric parsing.
It is hard to predict the future of MT but replacing human translators and interpreters is not even on the horizon. I suspect that NLP is very near the peak of the gains that can be made through stochastic processing in the same way it reached the end of what can be done with pure parsing twenty years ago. It will keep improving incrementally but the field is waiting for a new breakthrough.