Of course one should never make bold claims about the wonders of mechanical translation, or indeed of any claim within the broad area of artificial intelligence (if there is such a discipline) without providing proof that these claims could be reasonable.

As a simple test I used the latest web-demo of Systrans translation software on the log entry about machine translation in practice. Here are the results.
I took the paragraph:

The problem of correctly converting speech to meaning is hard enough, but if you can limit the responder to use a limited functional grammar - that could even be made an unambiguous closed subset of the open ambiguous grammar of the local language, you could conceivable have a valuable interchange with the speakers of this other language.

In Systran-French this is :

Le probl?me de convertir correctement la parole en signification est assez dur, mais si vous pouvez limiter le r?pondeur pour utiliser une grammaire fonctionnelle limit?e que pourrait m?me ?tre fait un sous-ensemble ferm? non ambigu de la grammaire ambigu? ouverte du langage local, vous pourriez imaginable avoir un ?change valable avec les parleurs de ce autre langage.

Which again (the true test) back-translates to this:

The problem to correctly convert the word into significance is rather hard, but if you can limit the responder to use a limited functional grammar which a nonambiguous closed subset of the open ambiguous grammar of the local language could be even made, you could conceivable have a valid exchange with the loudspeakers of this other language.

Clearly, this is not English spoken by a speaker of English - but for situations where conversational context can supply a lot of the meaning it isn't really all that bad. And certainly with the ability to further interrogate, speakers at both ends of the system could identify the typical style of error and adjust both their own utterances and their interpretation of the counterpart, to make some sense of the text.

As a further test, I iterated once more English->French->English and arrived at:

Problem to convert correctly word into significance is rather hard, but if you can limit the responder to use a limited functional grammar that a subset closed nonambiguous of the open ambiguous grammar of the local language could be even made, you could conceivable have a valid exchange with the loudspeakers of this other language.

This is really quite OK for degradation after four translations. Knowing the context of the conversation, I think you would get at least some of the meaning of the sentence.